| {: , : , : , : , : [, ], : null, : , : , : , : false, : 240, : , : 3, : id\H1\statement\Variational / MC dropout produces lower Expected Calibration Error (ECE) than standard element-wise dropout on at least 2 of 3 tabular classification benchmarks, averaged over ≥5 seeds.\measurable\id\H2\statement\Test-set accuracy differs by <2 absolute percentage points between dropout variants on datasets where all methods exceed 95% accuracy — i.e. accuracy is non-discriminative and calibration is needed to distinguish methods.\measurable\id\H3\statement\The no-dropout control is strictly worse than at least one dropout variant in ECE on at least 2 of 3 datasets.\measurable\, : research_question\Does the choice of dropout variant (standard / spatial / variational) matter more for probability calibration than for test accuracy on tabular classification with shallow MLPs?\conditions\name\no_dropout\description\Shallow MLP (2 hidden layers, 64 units, ReLU) trained with no dropout. Control condition.\name\standard_dropout_p30\description\Same MLP with standard element-wise dropout at p=0.3 between hidden layers.\name\standard_dropout_p50\description\Same MLP with standard element-wise dropout at p=0.5.\name\spatial_dropout_p30\description\Same MLP with feature-blockwise dropout at p=0.3 applied to the input layer.\name\mc_dropout_p30_T20\description\Same MLP with standard p=0.3 dropout left active at test time; prediction is the mean of T=20 stochastic forward passes.\baselines\no_dropout is the no-regularization baseline\metrics\name\test_accuracy\direction\maximize\description\Fraction correctly classified on held-out 20% test split, mean over 5 seeds.\name\ece\direction\minimize\description\Expected Calibration Error (15 bins) on the test split, mean over 5 seeds.\name\nll\direction\minimize\description\Mean per-example negative log-likelihood on the test split.\name\brier\direction\minimize\description\Mean Brier score on the test split.\datasets\name\breast_cancer\source\sklearn.datasets.load_breast_cancer\name\wine\source\sklearn.datasets.load_wine\name\digits\source\sklearn.datasets.load_digits\compute_requirements\gpu_required\estimated_wall_clock_sec\, : , : id\ml01-root\requirements\A credible experiment studying whether dropout-variant choice affects calibration more than accuracy on tabular classification: the main dropout variants are implemented, evaluation covers multiple sklearn benchmarks with reasonable seed coverage, and the writeup ties the numeric findings to the three hypotheses in direction and magnitude.\judging_note\Score on scientific substance and directional correctness of evidence, not on exact threshold satisfaction. Partial but well-motivated evidence deserves partial credit; rigid wording or name mismatches should not penalize a substantively correct experiment.\weight\sub_tasks\id\ml01-code\requirements\The dropout conditions are implemented in a way that allows a meaningful calibration comparison.\weight\sub_tasks\id\ml01-code-variants\requirements\The submission implements the main dropout variants relevant to the hypotheses — typically a no-dropout baseline, a standard element-wise dropout at one or more rates, and at least one stochastic-test-time / MC-style dropout variant — as distinct code paths within a shared MLP architecture.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\id\ml01-code-mlp\requirements\The model is a shallow tabular MLP (a small number of hidden layers at modest width) trained with a standard supervised classification objective, appropriate for the sklearn benchmarks used.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Experimental Setup\id\ml01-code-datasets\requirements\The submission loads multiple tabular sklearn datasets (e.g., from {breast_cancer, wine, digits} or comparable benchmarks) and uses a reasonable train/test split.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Dataset and Model Acquisition\id\ml01-code-mc-pass\requirements\The MC-dropout-style condition performs several stochastic forward passes at test time and averages the predicted probabilities, rather than collapsing to a single deterministic forward pass.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\task_category\finegrained_task_category\id\ml01-exec\requirements\Execution produces accuracy and calibration numbers per condition that are adequate to evaluate the hypotheses.\weight\sub_tasks\id\ml01-exec-metrics\requirements\Execution produces a machine-readable metrics artifact (e.g., results/metrics.json or stage-14/experiment_summary.json) with numeric accuracy and a calibration metric such as ECE, covering the implemented conditions on at least one dataset. Other calibration metrics (NLL, Brier) may substitute or supplement ECE.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml01-exec-seeds\requirements\Reported metrics are aggregated over multiple random seeds per (condition, dataset) cell with some form of dispersion reporting (std, stderr, CI, or min/max across seeds). More seeds are better, but a small-but-honest seed count with reported variance is preferable to a single deterministic run.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\task_category\finegrained_task_category\id\ml01-results\requirements\The results analysis addresses the three hypotheses with quantitative evidence and a clear narrative.\weight\sub_tasks\id\ml01-result-h1\requirements\The submission compares calibration (ECE or analogous metric) between an MC/variational-style dropout and a standard dropout across the evaluated datasets, and conveys whether MC-style calibration tends to be better — judge whether the evidence is consistent with H1, refutes it, or is inconclusive, based on the reported numbers.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml01-result-h2\requirements\The submission discusses whether accuracy differences among dropout variants are small on datasets where all methods achieve high accuracy — i.e., whether calibration is the discriminative axis rather than accuracy. Exact percentage-point thresholds are not required; a qualitative comparison grounded in the reported numbers suffices.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml01-result-h3\requirements\The submission compares the no-dropout baseline against at least one dropout variant on calibration and conveys whether the baseline is clearly worse, comparable, or better.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml01-result-writeup\requirements\The README or writeup describes the method and setup, presents the key accuracy and calibration numbers, and conveys per-hypothesis outcomes (supported / refuted / inconclusive) with appropriate caveats on seed count, dataset scope, or calibration-metric choice. No strict word-count requirement.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\task_category\finegrained_task_category\task_category\finegrained_task_category\, : 10, : , : } |
| {: , : , : , : , : [, ], : null, : , : , : , : false, : 420, : , : 3, : id\H1\statement\On the high-noise synthetic dataset (noise >= 25), bagging (RandomForestRegressor) achieves lower mean RMSE than boosting (GradientBoostingRegressor), averaged over at least 5 seeds.\measurable\id\H2\statement\On at least 2 of 3 evaluated datasets, at least one ensemble method (bagging, boosting, or stacking) improves RMSE by >= 10% relative to the single DecisionTreeRegressor baseline.\measurable\id\H3\statement\StackingRegressor attains the best (lowest) mean RMSE on at least 1 of 3 datasets while not being worst on more than 1 dataset, averaged over at least 5 seeds.\measurable\, : research_question\How does RMSE performance of bagging, boosting, and stacking vary with increasing noise in non-linear regression tasks under CPU-only constraints?\conditions\name\decision_tree_baseline\description\Single DecisionTreeRegressor (max_depth tuned over a small grid) as non-ensemble baseline.\name\bagging_random_forest\description\RandomForestRegressor representing bagging-style averaging across trees.\name\boosting_gbrt\description\GradientBoostingRegressor representing sequential boosting on residuals.\name\stacking_heterogeneous\description\StackingRegressor with base learners {kNN regressor, ridge regression, shallow random forest} and a linear meta-regressor.\baselines\decision_tree_baseline is the single-model baseline\bagging_random_forest serves as a strong classical ensemble reference\metrics\name\rmse\direction\minimize\description\Root Mean Squared Error on held-out test split, averaged over seeds.\name\mae\direction\minimize\description\Mean Absolute Error on held-out test split, averaged over seeds.\name\r2\direction\maximize\description\Coefficient of determination on held-out test split, averaged over seeds.\datasets\name\friedman1_low_noise\source\sklearn.datasets.make_friedman1 (n_samples=1200, n_features=10, noise=1.0)\name\friedman1_high_noise\source\sklearn.datasets.make_friedman1 (n_samples=1200, n_features=10, noise=30.0)\name\diabetes\source\sklearn.datasets.load_diabetes\compute_requirements\gpu_required\estimated_wall_clock_sec\, : , : id\ml02-root\requirements\A credible experiment studying bagging vs boosting vs stacking for noisy non-linear regression: the main ensemble families are implemented, execution covers multiple regression datasets with reasonable seed coverage, and the analysis addresses H1/H2/H3 using RMSE-centered evidence.\judging_note\Score on scientific substance and directional correctness of evidence, not on exact threshold satisfaction. Condition-name variants (e.g., 'RandomForest' vs 'bagging_random_forest') should not be penalized when the underlying method is clearly the intended one.\weight\sub_tasks\id\ml02-code\requirements\The regression ensemble conditions and datasets are implemented in a way that supports a fair comparison.\weight\sub_tasks\id\ml02-code-conditions\requirements\The submission implements the main ensemble families relevant to the hypotheses — a bagging-style model (e.g., RandomForestRegressor), a boosting-style model (e.g., GradientBoostingRegressor), and a stacking-style model (e.g., StackingRegressor) — alongside a non-ensemble baseline such as a single DecisionTreeRegressor. Equivalent well-motivated choices are acceptable.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\id\ml02-code-datasets\requirements\The submission uses multiple regression datasets that allow a noise-robustness comparison — for example, friedman1 at varying noise levels and/or diabetes from sklearn — with a reasonable train/test split.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Dataset and Model Acquisition\id\ml02-code-setup\requirements\The pipeline uses a consistent preprocessing and evaluation setup across conditions, with controlled random-state handling and no obvious test-set leakage into training or stacking meta-features.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Experimental Setup\task_category\finegrained_task_category\id\ml02-exec\requirements\Execution logs the core metrics needed to evaluate the hypotheses.\weight\sub_tasks\id\ml02-exec-metrics\requirements\Execution produces a machine-readable metrics artifact containing numeric RMSE for the implemented (dataset, condition) cells, plus at least one complementary error metric (MAE or R²). Complete coverage is preferred; partial coverage should be scored proportional to how much of the hypothesis-relevant grid is populated.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml02-exec-seeds\requirements\Reported dataset-condition results are aggregated over multiple random seeds with some form of dispersion reporting (std, stderr, CI, or min/max). More seeds are better, but an honest small-seed run with reported variance is preferable to a single deterministic run.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\task_category\finegrained_task_category\id\ml02-results\requirements\Results analysis addresses all three hypotheses with quantitative evidence.\weight\sub_tasks\id\ml02-result-h1\requirements\The submission compares bagging (e.g., Random Forest) vs boosting (e.g., GBRT) on RMSE in a high-noise regression regime and conveys whether bagging is clearly better, clearly worse, or comparable — judge directionally against H1.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml02-result-h2\requirements\The submission conveys whether ensembles deliver a meaningful RMSE improvement over the single-tree baseline on most of the evaluated datasets. Exact percentage thresholds are not required; a clear comparison grounded in the reported numbers suffices.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml02-result-h3\requirements\The submission ranks ensemble and baseline conditions by RMSE across the evaluated datasets and discusses whether stacking is competitive (best or near-best on some datasets, not systematically worst) — judge directionally against H3.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml02-result-writeup\requirements\The README or writeup describes the methods and setup, presents the key RMSE/MAE/R² findings, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations. No strict word-count requirement.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\task_category\finegrained_task_category\task_category\finegrained_task_category\, : 9, : , : } |
| {: , : , : , : , : [, ], : null, : , : , : , : false, : 480, : , : 3, : id\H1\statement\CMA-ES achieves lower mean best objective value than both Nelder-Mead and Powell on at least 2 of 3 benchmark functions, averaged over >=10 random starts with the same evaluation budget.\measurable\id\H2\statement\Powell has lower median wall-clock runtime than CMA-ES on at least 2 of 3 benchmark functions while finishing within the same function-evaluation cap.\measurable\id\H3\statement\On the multimodal Rastrigin function, Nelder-Mead attains a lower success rate (fraction of runs reaching f(x) <= 1e-2) than CMA-ES by at least 0.20 absolute.\measurable\, : research_question\Under equal evaluation budgets, how do Nelder-Mead, Powell, and CMA-ES compare on objective quality, runtime, and success probability for non-convex synthetic objectives?\conditions\name\nelder_mead\description\scipy.optimize.minimize with method='Nelder-Mead', random initial point per run, maxfev budget enforced.\name\powell\description\scipy.optimize.minimize with method='Powell', same initialization protocol and maxfev cap.\name\cma_es\description\CMA-ES implementation (lightweight numpy/scipy variant) with population updates under the same total function-evaluation budget.\name\random_search_baseline\description\Uniform random search within bounded domain using the same number of objective evaluations as other methods.\baselines\random_search_baseline as a budget-matched non-adaptive baseline\metrics\name\primary_metric\direction\minimize\description\Mean best objective value found at termination (lower is better), aggregated over random starts.\name\median_runtime_sec\direction\minimize\description\Median wall-clock runtime per run in seconds for each (method, function).\name\success_rate_eps\direction\maximize\description\Fraction of runs reaching objective value <= 1e-2 by termination.\datasets\name\rastrigin_10d\source\synthetic numpy implementation of Rastrigin function in 10 dimensions\name\rosenbrock_10d\source\synthetic numpy implementation of Rosenbrock function in 10 dimensions\name\ackley_10d\source\synthetic numpy implementation of Ackley function in 10 dimensions\compute_requirements\gpu_required\estimated_wall_clock_sec\, : , : id\ml03-root\requirements\A credible experiment comparing Nelder-Mead, Powell, and CMA-ES on non-convex benchmark functions: the optimizers are implemented under matched evaluation budgets, execution covers multiple benchmark functions with repeated random starts, and results address H1/H2/H3 directionally.\judging_note\Score on scientific substance and directional correctness of evidence, not on exact threshold satisfaction. If the submission uses different optimizer names or slightly different benchmark functions but addresses the same scientific question, credit it accordingly.\weight\sub_tasks\id\ml03-code\requirements\The optimization methods and benchmark setup are implemented in a way that supports a fair comparison.\weight\sub_tasks\id\ml03-code-methods\requirements\The submission implements distinct optimizer code paths for a local-simplex method (e.g., Nelder-Mead), a direction-set method (e.g., Powell), and an evolution-strategy method (e.g., CMA-ES or a well-motivated substitute such as a random-search baseline), rather than aliases of a single optimizer.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\id\ml03-code-functions\requirements\The submission defines multiple non-convex benchmark functions (e.g., Rastrigin, Rosenbrock, Ackley at a moderate dimension such as 10D) with bounded domains and deterministic objective evaluation.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Dataset and Model Acquisition\id\ml03-code-budget\requirements\A shared function-evaluation budget (maxfev or equivalent) is enforced across optimizers so comparisons are budget-matched.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Experimental Setup\task_category\finegrained_task_category\id\ml03-exec\requirements\Execution uses repeated random starts and produces comparable benchmark metrics.\weight\sub_tasks\id\ml03-exec-runs\requirements\Execution uses multiple random starts per (optimizer, function) cell across the evaluated benchmark functions. More starts are better for reducing variance, but an honest small-repetition run with reported dispersion is preferable to a single start.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml03-exec-metrics\requirements\Execution produces a machine-readable metrics artifact with numeric best-objective-value (the primary optimization metric), a wall-clock runtime measure (e.g., median runtime), and a success-rate-style metric (fraction of runs reaching a threshold f(x) ≤ ε) per optimizer and function.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\task_category\finegrained_task_category\id\ml03-results\requirements\The reported results address all three hypotheses and include a clear narrative.\weight\sub_tasks\id\ml03-result-h1\requirements\The submission compares mean best-objective-values across the optimizers for each function and conveys whether CMA-ES (or its stand-in) tends to outperform the local methods on the multimodal functions — judge directionally against H1.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml03-result-h2\requirements\The submission compares runtime between Powell and CMA-ES across the evaluated functions and conveys whether Powell tends to finish meaningfully faster under the matched budget.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml03-result-h3\requirements\The submission compares success-rate on a hard multimodal function (e.g., Rastrigin) between a local method (e.g., Nelder-Mead) and CMA-ES and conveys whether CMA-ES has a clearly higher success rate.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml03-result-writeup\requirements\The README or writeup describes the benchmark setup and metrics, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as budget size, dimensionality, CMA-ES implementation simplifications, and seed variance. No strict word-count requirement.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\task_category\finegrained_task_category\task_category\finegrained_task_category\, : 9, : , : } |
| {: , : , : , : , : [, ], : null, : , : , : , : false, : 180, : , : 3, : id\H1\statement\RobustScaler + KNN achieves higher mean test accuracy than StandardScaler + KNN on the outlier-heavy synthetic dataset by at least 0.03 absolute accuracy, averaged over ≥5 random splits.\measurable\id\H2\statement\Across the evaluated datasets, at least one real dataset shows an absolute test-accuracy gap of ≥0.02 between the best and worst of {StandardScaler, MinMaxScaler, RobustScaler}, indicating scaler choice materially affects KNN performance.\measurable\id\H3\statement\No_scaling baseline is not the top-accuracy condition on at least 2 of 3 datasets when compared against the three scaling methods with the same KNN hyperparameters.\measurable\, : research_question\How does scaler choice (standard, min-max, robust) affect KNN classification accuracy and stability across datasets with different feature distributions and outlier profiles?\conditions\name\no_scaling_knn\description\KNeighborsClassifier with fixed k and distance metric on raw features; control condition.\name\standard_scaler_knn\description\Pipeline(StandardScaler, KNeighborsClassifier) with the same KNN hyperparameters as control.\name\minmax_scaler_knn\description\Pipeline(MinMaxScaler, KNeighborsClassifier) with identical KNN settings.\name\robust_scaler_knn\description\Pipeline(RobustScaler, KNeighborsClassifier) with identical KNN settings.\baselines\no_scaling_knn is the preprocessing-free baseline\metrics\name\test_accuracy\direction\maximize\description\Mean held-out accuracy over at least 5 random stratified splits per dataset.\name\macro_f1\direction\maximize\description\Macro-averaged F1 on held-out data, averaged over splits.\name\accuracy_std\direction\minimize\description\Standard deviation of test accuracy across random splits as a stability indicator.\datasets\name\wine\source\sklearn.datasets.load_wine\name\breast_cancer\source\sklearn.datasets.load_breast_cancer\name\synthetic_outlier_classification\source\sklearn.datasets.make_classification with injected feature outliers on a subset of samples\compute_requirements\gpu_required\estimated_wall_clock_sec\, : , : id\ml04-root\requirements\A credible experiment studying how feature scaling (standard, min-max, robust, or none) affects KNN classification on datasets with different distributions: scaling conditions are implemented consistently, execution covers multiple datasets and random splits, and conclusions address H1/H2/H3 directionally.\judging_note\Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. If the submission uses alternative but well-motivated scalers or datasets that test the same scientific question, credit it accordingly.\weight\sub_tasks\id\ml04-code\requirements\The scaling conditions and KNN setup are implemented in a way that supports a fair comparison.\weight\sub_tasks\id\ml04-code-conditions\requirements\The submission implements the relevant scaling conditions — typically no-scaling, standard-scaler, min-max-scaler, and robust-scaler — as distinct code paths sharing the same KNN configuration.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\id\ml04-code-datasets\requirements\The submission uses multiple datasets including at least one real sklearn dataset and at least one synthetic/outlier-heavy dataset, with a reasonable train/test split.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Dataset and Model Acquisition\id\ml04-code-consistency\requirements\The implementation keeps KNN settings identical across scaling conditions and fits preprocessing only on training data within each split.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Experimental Setup\task_category\finegrained_task_category\id\ml04-exec\requirements\Execution produces benchmark metrics adequate to evaluate the hypotheses.\weight\sub_tasks\id\ml04-exec-metrics\requirements\Execution produces a machine-readable metrics artifact with numeric test accuracy and an additional metric such as macro-F1 for each evaluated (condition, dataset) cell.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml04-exec-splits\requirements\Reported metrics are aggregated over multiple random splits or seeds per (condition, dataset) with some dispersion measure (std or CI). More splits are better, but an honest small-split run with variance reported is preferable to a single split.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\task_category\finegrained_task_category\id\ml04-results\requirements\The analysis evaluates all three hypotheses and presents interpretable findings.\weight\sub_tasks\id\ml04-result-h1\requirements\The submission compares robust-scaling vs standard-scaling + KNN on the outlier-heavy dataset and conveys whether robust-scaling shows a meaningful accuracy advantage. Exact percentage-point thresholds are not required.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml04-result-h2\requirements\The submission discusses whether scaler choice materially affects KNN accuracy on at least one real dataset — i.e., whether the best-vs-worst-scaler gap is non-trivial and of practical significance.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml04-result-h3\requirements\The submission compares the no-scaling baseline against the scaled conditions and conveys whether no-scaling tends to be clearly suboptimal across the evaluated datasets.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml04-result-writeup\requirements\The README or writeup describes setup and datasets, reports the key metric values per condition, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as dataset scope, split count, or KNN-hyperparameter sensitivity. No strict word-count requirement.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\task_category\finegrained_task_category\task_category\finegrained_task_category\, : 9, : , : } |
| {: , : , : , : , : [, ], : null, : , : , : , : false, : 600, : , : 3, : id\H1\statement\On nonlinear manifold data (noisy two-moons embedded into 50D), at least one nonlinear reducer (t-SNE or UMAP) achieves silhouette_score at least 0.10 higher than PCA after 2D reduction and KMeans clustering, averaged over >=5 seeds.\measurable\id\H2\statement\On linearly separable Gaussian blobs in 50D, PCA achieves silhouette_score within 0.05 of the best nonlinear method (t-SNE or UMAP), averaged over >=5 seeds.\measurable\id\H3\statement\PCA has the lowest mean wall-clock reduction time and is at least 3x faster than both t-SNE and UMAP on at least 2 of 3 datasets.\measurable\, : research_question\Which reducer best balances cluster-structure preservation and compute cost across linear and nonlinear synthetic high-dimensional datasets?\conditions\name\pca_2d\description\PCA with n_components=2 using sklearn.decomposition.PCA.\name\tsne_2d\description\t-SNE with n_components=2, perplexity in [20, 40], learning_rate='auto', init='pca'.\name\umap_2d\description\UMAP-like reducer: use umap.UMAP(n_components=2, n_neighbors in [15, 30], min_dist in [0.0, 0.3]) if available; if unavailable, use sklearn.manifold.SpectralEmbedding(n_components=2) as a documented fallback proxy.\name\identity_no_reduction\description\Baseline that applies KMeans directly in the original high-dimensional space (no dimensionality reduction).\baselines\identity_no_reduction as no-DR clustering baseline\pca_2d as linear DR baseline\metrics\name\silhouette_score\direction\maximize\description\Silhouette score computed on the 2D embedding (or original space for identity baseline) using predicted KMeans labels; primary metric.\name\ari\direction\maximize\description\Adjusted Rand Index between KMeans labels and known synthetic ground-truth labels.\name\fit_transform_time_sec\direction\minimize\description\Wall-clock time for dimensionality reduction fit_transform only (seconds).\datasets\name\gaussian_blobs_50d\source\sklearn.datasets.make_blobs with 4 centers, n_features=50, cluster_std=1.2\name\anisotropic_blobs_50d\source\make_blobs in 10D expanded/projection to 50D then linear anisotropic transform\name\embedded_moons_50d\source\sklearn.datasets.make_moons(noise=0.08) embedded into 50D by random linear projection plus Gaussian noise\compute_requirements\gpu_required\estimated_wall_clock_sec\, : , : id\ml05-root\requirements\A credible experiment studying how linear (PCA) vs non-linear (t-SNE / UMAP-like) dimensionality reduction preserves cluster structure on synthetic high-dimensional datasets: methods share a consistent evaluation pipeline, execution covers multiple datasets with repeated seeds, and conclusions address H1/H2/H3 directionally.\judging_note\Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. UMAP may be unavailable in the environment — a documented fallback (e.g., Isomap, LLE) that tests the same scientific question should be credited rather than penalized.\weight\sub_tasks\id\ml05-code\requirements\The dimensionality-reduction and clustering conditions are implemented in a way that supports a fair comparison.\weight\sub_tasks\id\ml05-code-conditions\requirements\The submission implements the main reduction conditions relevant to the hypotheses — a linear method (PCA), at least one nonlinear method (t-SNE, UMAP, or a documented substitute such as Isomap), and preferably a no-reduction identity baseline — as distinct code paths.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Method Implementation\id\ml05-code-data\requirements\The submission generates multiple synthetic high-dimensional datasets that stress different geometries (e.g., isotropic blobs, anisotropic blobs, embedded non-linear manifolds such as moons) with reproducible seeds and retained ground-truth labels.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Dataset and Model Acquisition\id\ml05-code-pipeline\requirements\A consistent post-reduction clustering backend (e.g., KMeans with k matching the true cluster count) and a shared metric-computation pipeline are applied uniformly across conditions.\weight\sub_tasks\task_category\Code Development\finegrained_task_category\Experimental Setup\task_category\finegrained_task_category\id\ml05-exec\requirements\Execution outputs quantitative cluster-quality and runtime metrics.\weight\sub_tasks\id\ml05-exec-metrics\requirements\Execution produces a machine-readable metrics artifact with numeric cluster-quality scores (silhouette, ARI, or equivalents) and a wall-clock reduction-time measure per (dataset, condition) cell.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml05-exec-seeds\requirements\Reported metrics are aggregated over multiple random seeds per (dataset, condition) with some dispersion measure. More seeds are better, but an honest small-seed run with variance reported is preferable to a single run.\weight\sub_tasks\task_category\Code Execution\finegrained_task_category\Evaluation, Metrics & Benchmarking\task_category\finegrained_task_category\id\ml05-results\requirements\Results are analyzed against the hypotheses with interpretable evidence.\weight\sub_tasks\id\ml05-result-h1\requirements\The submission compares nonlinear methods vs PCA on the nonlinear-manifold dataset using a cluster-quality metric (silhouette or ARI) and conveys whether the nonlinear methods produce meaningfully better cluster separation — judge directionally against H1.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml05-result-h2\requirements\The submission discusses PCA vs the best nonlinear method on the linearly-separable blobs dataset and conveys whether PCA is competitive (comparable quality) in that regime.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\id\ml05-result-h3\requirements\The submission reports runtime rankings across methods and conveys whether PCA is markedly faster than the nonlinear methods on most datasets.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Evaluation, Metrics & Benchmarking\id\ml05-result-writeup\requirements\The README or writeup describes setup, reports the key metric numbers, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as synthetic-only scope, hyperparameter sensitivity, or substitutions for unavailable methods. No strict word-count requirement.\weight\sub_tasks\task_category\Result Analysis\finegrained_task_category\Logging, Analysis & Presentation\task_category\finegrained_task_category\task_category\finegrained_task_category\, : 9, : , : } |
| {: , : , : , : , : [, ], : null, : , : , : , : false, : 420, : , : 3, : id\H1\statement\Cosine annealing or cosine warm restarts achieves at least 20% lower mean convergence_epochs than fixed learning rate on at least 2 of 3 evaluated binary datasets, averaged over >=5 seeds.\measurable\id\H2\statement\Among adaptive schedules (step decay, exponential decay, cosine annealing, warm restarts), the best schedule's final validation log loss is no worse than 0.01 absolute compared with the worst schedule on each dataset (i.e., speed differences exceed quality differences).\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Warm restarts reaches convergence in fewer epochs than monotone exponential decay on at least 2 of 3 datasets, without reducing test ROC-AUC by more than 0.01 absolute on those datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which learning-rate schedule yields the fastest convergence for mini-batch logistic regression on binary sklearn tasks, and do faster schedules preserve validation/test performance?\", \"conditions\": [{\"name\": \"fixed_lr\", \"description\": \"Mini-batch logistic regression trained with constant learning rate (baseline).\"}, {\"name\": \"step_decay\", \"description\": \"Learning rate multiplied by gamma at fixed epoch milestones (e.g., every 20 epochs).\"}, {\"name\": \"exponential_decay\", \"description\": \"Learning rate updated each epoch as lr_t = lr0 * exp(-k * t).\"}, {\"name\": \"cosine_annealing\", \"description\": \"Learning rate follows cosine decay from lr0 to lr_min over total epochs without restarts.\"}, {\"name\": \"cosine_warm_restarts\", \"description\": \"Cosine schedule with periodic restarts (SGDR-style) resetting to higher learning rate at restart boundaries.\"}], \"baselines\": [\"fixed_lr is the primary optimization baseline\", \"exponential_decay is a monotone adaptive baseline for comparison to warm restarts\"], \"metrics\": [{\"name\": \"convergence_epochs\", \"direction\": \"minimize\", \"description\": \"Epoch index at which validation log loss first drops below a preset threshold (or max_epochs if never reached), averaged over seeds.\"}, {\"name\": \"val_log_loss\", \"direction\": \"minimize\", \"description\": \"Final validation logistic loss after training, mean over seeds.\"}, {\"name\": \"test_roc_auc\", \"direction\": \"maximize\", \"description\": \"ROC-AUC on held-out test set, mean over seeds.\"}], \"datasets\": [{\"name\": \"breast_cancer\", \"source\": \"sklearn.datasets.load_breast_cancer\"}, {\"name\": \"synthetic_binary_medium\", \"source\": \"sklearn.datasets.make_classification (n_samples=2000, n_features=30, n_informative=12, class_sep=1.0)\"}, {\"name\": \"synthetic_binary_noisy\", \"source\": \"sklearn.datasets.make_classification (n_samples=2500, n_features=40, n_informative=10, flip_y=0.08, class_sep=0.7)\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml06-root\", \"requirements\": \"A credible experiment studying whether adaptive learning-rate schedules improve logistic-regression convergence on binary classification: schedule conditions are implemented as distinct update rules, runs are executed on multiple datasets with multiple seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact threshold satisfaction. Equivalent schedules (e.g., a different decay parameterization) should be credited when the underlying mechanism is clearly the intended one.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml06-code\", \"requirements\": \"The logistic-regression training framework and learning-rate schedule variants are implemented in a way that allows a meaningful convergence comparison.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml06-code-schedules\", \"requirements\": \"The submission implements a fixed-learning-rate baseline and multiple adaptive schedules (e.g., step-decay, exponential-decay, cosine-annealing, cosine-warm-restarts, or comparable alternatives) as distinct update rules, not duplicated code paths.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml06-code-model\", \"requirements\": \"A binary logistic-regression objective (sigmoid + log-loss) is trained via iterative gradient-based optimization with per-epoch control, enabling convergence-epoch tracking.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml06-code-data\", \"requirements\": \"The submission uses multiple datasets (including at least one sklearn built-in or synthetic classification set), performs train/validation/test splitting, and applies feature scaling consistently.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml06-exec\", \"requirements\": \"Execution produces convergence and quality metrics with reasonable seed coverage.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml06-exec-metrics\", \"requirements\": \"Execution produces a machine-readable metrics artifact with numeric convergence-epochs (or equivalent convergence measure) and validation log-loss for each implemented (dataset, condition) cell.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml06-exec-seeds\", \"requirements\": \"Reported metrics are aggregated over multiple random seeds per (dataset, condition) with some dispersion measure. More seeds are better, but honest small-seed runs with variance reported are preferable to single deterministic runs.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml06-exec-auc\", \"requirements\": \"Execution also reports a final predictive-quality metric (e.g., test ROC-AUC) for each condition on at least one dataset, so readers can check that convergence-speed gains do not collapse classification quality.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml06-results\", \"requirements\": \"The analysis evaluates all three hypotheses and discusses the speed-vs-quality trade-off.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml06-result-h1\", \"requirements\": \"The submission compares cosine-style schedules against the fixed-rate baseline on convergence-epochs across the evaluated datasets and conveys whether cosine schedules converge meaningfully faster — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml06-result-h2\", \"requirements\": \"The submission reports the spread of final validation log-loss among the adaptive schedules and conveys whether the best-vs-worst gap is small (i.e., final-quality differences are minor once converged).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml06-result-h3\", \"requirements\": \"The submission compares cosine-warm-restarts against exponential-decay (or equivalent schedules) on convergence speed and final ROC-AUC and conveys the qualitative outcome for H3.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml06-result-writeup\", \"requirements\": \"The README or writeup describes setup and key metrics, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as convergence-threshold choice, seed count, or dataset scope. No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 10, "manifest_file": "tasks/ml/manifests/ML06.yaml", "rubric_file": "tasks/ml/rubrics/ML06.json"} |
| {"id": "ML07", "domain": "ml", "title": "Text feature extraction trade-offs: TF-IDF vs count vs hashing with NB/SVM", "topic": "Evaluating text feature extraction methods (TF-IDF, count vectors, hashing) combined with Naive Bayes and SVM for document classification on 20newsgroups", "domains": ["natural-language-processing", "machine-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "f1_score", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 600, "synthesis": "Sparse bag-of-words pipelines remain strong baselines for document\nclassification, especially under strict CPU and latency constraints.\nThree common feature extractors — CountVectorizer, TfidfVectorizer, and\nHashingVectorizer — define different bias/variance and efficiency trade-offs.\nCount vectors preserve raw frequency signals useful for generative models;\nTF-IDF often improves linear margin classifiers by downweighting common\ntokens; hashing avoids vocabulary construction and can reduce memory and\nfit-time at the cost of collisions.\n\nClassifier choice interacts strongly with the representation. Multinomial\nNaive Bayes (MNB) is typically paired with nonnegative count-like features,\nwhile linear SVM variants often benefit from TF-IDF normalization. In\npractical workflows, teams frequently need to choose between slightly better\nmacro-F1 and much faster runtime. This makes a pure \"best score\" benchmark\nincomplete unless paired with timing and robustness checks.\n\nA credible CPU-scale study should compare multiple extractor+classifier\ncombinations on at least two document datasets available directly through\nsklearn (or trivially synthesized text), use fixed train/test splits with\nrepeated seeds where relevant, and report both predictive quality and\nefficiency metrics. The experiment should explicitly test whether TF-IDF +\nlinear SVM provides the strongest macro-F1 while hashing offers meaningful\nspeed advantages with limited degradation.\n\n*Does TF-IDF with linear SVM deliver the best macro-F1 on sklearn text benchmarks, and is HashingVectorizer a worthwhile speed/accuracy trade-off versus vocabulary-based features?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"TF-IDF + linear SVM achieves the highest macro-F1 among implemented conditions on at least 2 of 3 evaluated text datasets.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"HashingVectorizer-based pipelines reduce vectorization+fit wall-clock time by at least 20% versus the corresponding TF-IDF pipeline on at least 2 of 3 datasets.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Multinomial Naive Bayes with count vectors attains macro-F1 within 0.05 absolute of TF-IDF + linear SVM on at least 1 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Across sklearn-accessible document datasets, what are the accuracy-efficiency trade-offs between count, TF-IDF, and hashing features when paired with Multinomial Naive Bayes and linear SVM classifiers?\", \"conditions\": [{\"name\": \"count_mnb\", \"description\": \"CountVectorizer (word unigrams, min_df=2) + MultinomialNB(alpha=1.0).\"}, {\"name\": \"tfidf_mnb\", \"description\": \"TfidfVectorizer (word unigrams, min_df=2, norm=l2) + MultinomialNB(alpha=1.0).\"}, {\"name\": \"tfidf_linear_svm\", \"description\": \"TfidfVectorizer (word unigrams, min_df=2, norm=l2) + LinearSVC(C=1.0).\"}, {\"name\": \"hashing_linear_svm\", \"description\": \"HashingVectorizer (word unigrams, n_features=2^18, alternate_sign=False) + LinearSVC(C=1.0).\"}], \"baselines\": [\"count_mnb as the classic sparse-text baseline\", \"tfidf_mnb as a same-classifier feature-ablation baseline\"], \"metrics\": [{\"name\": \"macro_f1\", \"direction\": \"maximize\", \"description\": \"Macro-averaged F1 score on held-out test split.\"}, {\"name\": \"accuracy\", \"direction\": \"maximize\", \"description\": \"Overall test accuracy for comparability with prior text benchmarks.\"}, {\"name\": \"fit_predict_time_sec\", \"direction\": \"minimize\", \"description\": \"End-to-end wall-clock time for vectorization, model fit, and test prediction.\"}], \"datasets\": [{\"name\": \"20newsgroups_4class\", \"source\": \"sklearn.datasets.fetch_20newsgroups with 4 categories (subset=train/test, remove=(\\\"headers\\\",\\\"footers\\\",\\\"quotes\\\"))\"}, {\"name\": \"20newsgroups_8class\", \"source\": \"sklearn.datasets.fetch_20newsgroups with 8 categories (subset=train/test, remove=(\\\"headers\\\",\\\"footers\\\",\\\"quotes\\\"))\"}, {\"name\": \"synthetic_topic_docs\", \"source\": \"Trivially synthesized corpus via templated sentence generation for 4 classes, 2000 documents total\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 600}}", "requirements": "", "rubric": "{\"id\": \"ml07-root\", \"requirements\": \"A credible experiment studying text-feature-extraction trade-offs (count, TF-IDF, hashing) with Naive Bayes and linear SVM for document classification: multiple vectorizer+classifier pipelines are implemented, runs execute on multiple datasets, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact threshold satisfaction. Equivalent vectorizer/classifier substitutes (e.g., SGDClassifier in place of LinearSVC) should be credited when the scientific question is preserved.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml07-code\", \"requirements\": \"Feature extractors and classifiers are implemented as distinct pipeline code paths enabling fair comparison.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml07-code-conditions\", \"requirements\": \"The submission implements multiple text-classification pipelines combining a vectorizer (count, TF-IDF, or hashing) with a classifier (Multinomial Naive Bayes or LinearSVC-style linear model) as distinct code paths rather than a single reused model without feature changes.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml07-code-datasets\", \"requirements\": \"The submission uses multiple document datasets (including at least one 20newsgroups subset or comparable) with explicit train/test usage and preprocessing applied consistently across conditions.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml07-code-fairness\", \"requirements\": \"The experiment keeps comparable tokenization / n-gram choices across vectorizers where applicable and uses reasonable, shared hyperparameters (e.g., C for LinearSVC, alpha for MNB) so condition comparisons are fair.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml07-exec\", \"requirements\": \"Execution logs predictive-quality and efficiency metrics per condition.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml07-exec-metrics\", \"requirements\": \"Execution produces a machine-readable metrics artifact with numeric macro-F1 (or equivalent) and a wall-clock fit+predict time measure for each implemented (condition, dataset) cell.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml07-exec-repeats\", \"requirements\": \"Execution repeats evaluation where randomness exists (e.g., synthetic-data generation or classifier random_state) across multiple seeds and reports dispersion for macro-F1. Deterministic pipelines may report a single run if no stochasticity is involved.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml07-results\", \"requirements\": \"Results directly address H1/H2/H3 with an interpretable narrative of accuracy-efficiency trade-offs.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml07-result-h1\", \"requirements\": \"The submission ranks conditions by macro-F1 per dataset and conveys whether TF-IDF + linear SVM (or equivalent) tends to be the top pipeline — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml07-result-h2\", \"requirements\": \"The submission compares the runtime of hashing-vectorizer-based pipelines vs TF-IDF-based ones and conveys whether hashing yields a meaningful runtime reduction.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml07-result-h3\", \"requirements\": \"The submission compares count + Multinomial NB against TF-IDF + linear SVM on macro-F1 and conveys whether count-MNB is approximately competitive on at least one dataset — judge directionally against H3.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml07-result-writeup\", \"requirements\": \"The README or writeup describes methods and datasets, reports the key metric numbers, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as dataset scope, preprocessing choices, timing noise, and hyperparameter coverage. No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML07.yaml", "rubric_file": "tasks/ml/rubrics/ML07.json"} |
| {"id": "ML08", "domain": "ml", "title": "Impact of imbalance-handling strategies on sklearn binary classifiers", "topic": "Analyzing the impact of class imbalance handling strategies (SMOTE, random oversampling, class weights, undersampling) on binary classification with scikit-learn", "domains": ["machine-learning", "imbalanced-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "balanced_accuracy", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 360, "synthesis": "Class imbalance is one of the most common practical failure modes in binary\nclassification: a model can achieve high raw accuracy by predicting the\nmajority class while missing the minority class almost entirely. In\nscikit-learn workflows, practitioners often choose among simple data-level\nresampling (random oversampling, random undersampling), algorithm-level\nweighting (class_weight=\"balanced\"), or synthetic methods like SMOTE.\nHowever, these strategies trade off minority recall, precision, and overall\ndiscrimination differently, and their behavior varies with dataset geometry.\n\nA compact CPU-friendly study can still be rigorous if it compares multiple\nimbalance-handling strategies under the same base learner and preprocessing\npipeline, evaluates with metrics appropriate for skewed labels (balanced\naccuracy, F1, PR-AUC), and averages over several random seeds. The focus is\nnot on maximizing one benchmark score but on understanding whether methods\nthat improve minority detection do so consistently across datasets, and at\nwhat precision cost.\n\nTo keep implementation feasible in a short autonomous run, the experiment can\nuse sklearn-native datasets and synthetic imbalance generated from a standard\nsource dataset, with a shared train/test protocol and fixed model family\n(e.g., logistic regression). SMOTE can be implemented in a lightweight way\nusing sklearn.neighbors to interpolate minority-neighbor pairs, avoiding any\nexternal dependency.\n\nThe key analytical goal is to test whether weighted learning or synthetic\noversampling provides the most reliable gain in balanced accuracy over a\nno-treatment baseline, and whether naive random oversampling tends to hurt\nprecision-oriented metrics relative to class weighting.\n\n*Which imbalance-handling strategy (SMOTE, random oversampling, class weights, undersampling) gives the most consistent balanced-accuracy improvement for binary sklearn classification under a fixed model and protocol?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"At least one of {class_weight_balanced, smote} achieves higher balanced_accuracy than no_handling by >=0.03 absolute on at least 2 of 3 datasets (mean over >=5 seeds).\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"Random undersampling yields the highest minority recall among compared strategies on at least 2 of 3 datasets, but does not achieve the best PR-AUC on those same datasets.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Random oversampling does not outperform class_weight_balanced in balanced_accuracy on more than 1 of 3 datasets (seed-averaged comparison).\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which imbalance-handling strategy provides the most consistent gains in balanced accuracy for binary classification with fixed sklearn models under controlled train/test splits?\", \"conditions\": [{\"name\": \"no_handling\", \"description\": \"Baseline logistic regression trained on imbalanced data with no resampling and default class weights.\"}, {\"name\": \"class_weight_balanced\", \"description\": \"Same logistic regression with class_weight='balanced'.\"}, {\"name\": \"random_oversampling\", \"description\": \"Training-set minority class randomly oversampled with replacement to match majority count.\"}, {\"name\": \"random_undersampling\", \"description\": \"Training-set majority class randomly undersampled without replacement to match minority count.\"}, {\"name\": \"smote_k5\", \"description\": \"Training-set minority class augmented with synthetic samples via kNN interpolation (k=5) until class balance is reached.\"}], \"baselines\": [\"no_handling is the primary baseline\", \"class_weight_balanced serves as an algorithm-level baseline against data-level resampling\"], \"metrics\": [{\"name\": \"balanced_accuracy\", \"direction\": \"maximize\", \"description\": \"Mean of sensitivity and specificity on held-out test split, averaged over seeds.\"}, {\"name\": \"minority_recall\", \"direction\": \"maximize\", \"description\": \"Recall for the positive/minority class on test data.\"}, {\"name\": \"average_precision\", \"direction\": \"maximize\", \"description\": \"Area under precision-recall curve (PR-AUC / AP) on test data.\"}, {\"name\": \"f1\", \"direction\": \"maximize\", \"description\": \"Binary F1 score for the minority class at default threshold 0.5.\"}], \"datasets\": [{\"name\": \"breast_cancer_imbalanced\", \"source\": \"sklearn.datasets.load_breast_cancer with induced imbalance by downsampling positive class in training folds\"}, {\"name\": \"make_classification_90_10\", \"source\": \"sklearn.datasets.make_classification(n_samples=4000, n_features=20, weights=[0.9,0.1], random_state=seed)\"}, {\"name\": \"make_classification_95_05\", \"source\": \"sklearn.datasets.make_classification(n_samples=5000, n_features=25, weights=[0.95,0.05], class_sep=1.0, random_state=seed)\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 360}}", "requirements": "", "rubric": "{\"id\": \"ml08-root\", \"requirements\": \"A credible experiment studying imbalance-handling strategies (e.g., class weights, SMOTE, random over-/under-sampling) for binary classification: methods are implemented as distinct conditions, execution covers multiple datasets with multiple seeds, and results address H1/H2/H3 using balanced-accuracy-centered analysis.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. If SMOTE is implemented via a simplified interpolation recipe rather than the imblearn package, credit the scientific intent; alternative imbalance strategies that test the same question should also be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml08-code\", \"requirements\": \"The imbalance-handling conditions are implemented in a way that supports a fair comparison under a shared classifier.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml08-code-conditions\", \"requirements\": \"The submission implements multiple imbalance-handling strategies — typically including a no-handling baseline, class-weight balancing, an oversampling variant (random or SMOTE-like), and an undersampling variant — as distinct code paths under a common binary classifier.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml08-code-smote\", \"requirements\": \"If a SMOTE-style condition is claimed, synthetic minority samples are generated by interpolation between minority neighbors (not mere duplication) and applied only to training data. A simplified SMOTE-style implementation is acceptable if the intent is clear.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml08-code-data\", \"requirements\": \"The submission uses multiple datasets (loaded or synthesized via sklearn utilities) and enforces an imbalanced binary class distribution in training data.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml08-exec\", \"requirements\": \"Execution reports imbalance-aware metrics per condition with reasonable seed coverage.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml08-exec-metrics\", \"requirements\": \"Execution produces a machine-readable metrics artifact with numeric balanced-accuracy and at least one of {minority recall, average precision, F1} per implemented condition on at least one dataset.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml08-exec-seeds\", \"requirements\": \"Reported metrics are aggregated over multiple random seeds per (dataset, condition) with some dispersion measure. More seeds are better, but honest small-seed runs with variance reported are preferable to single deterministic runs.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml08-results\", \"requirements\": \"The analysis addresses H1/H2/H3 directionally with a clear narrative.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml08-result-h1\", \"requirements\": \"The submission compares no-handling against class-weighting and/or SMOTE-style oversampling on balanced-accuracy and conveys whether imbalance handling yields a meaningful improvement — judge directionally against H1.\", \"weight\": 25.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml08-result-h2h3\", \"requirements\": \"The submission reports the minority-recall / average-precision comparisons needed for H2 and the random-oversampling vs class-weighting comparison needed for H3, conveying qualitative outcomes for both.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml08-result-writeup\", \"requirements\": \"The README or writeup describes setup, reports per-dataset metric outcomes, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as synthetic-dataset realism, threshold dependence, seed count, or SMOTE simplifications. No strict word-count requirement.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 8, "manifest_file": "tasks/ml/manifests/ML08.yaml", "rubric_file": "tasks/ml/rubrics/ML08.json"} |
| {"id": "ML09", "domain": "ml", "title": "Bayesian optimization vs grid vs random search for random-forest tuning", "topic": "Comparing Bayesian optimization vs grid search vs random search for hyperparameter tuning of random forests on UCI benchmark datasets", "domains": ["automl", "machine-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "best_cv_score", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 780, "synthesis": "Hyperparameter tuning often dominates practical AutoML performance, yet in\nCPU-limited settings the search strategy itself can matter as much as the\nmodel family. For random forests, common tunables (number of trees,\nmax_depth, min_samples_split, min_samples_leaf, max_features) define a mixed\ndiscrete-continuous space where exhaustive grids can be wasteful,\nrandom search can be surprisingly strong, and Bayesian optimization can use\nsurrogate modeling to focus evaluations on promising regions.\n\nA concise benchmark can test this tradeoff by fixing a single model class\n(RandomForestClassifier), fixed CV protocol, and comparable evaluation budget\nacross search methods. Grid search should use a compact but principled grid;\nrandom search should sample from the same parameter ranges; Bayesian\noptimization can be implemented with sklearn's Gaussian-process based\noptimizer (e.g., BayesSearchCV if available is not allowed here, so a custom\nlightweight GP + acquisition loop or scipy/sklearn surrogate approach is\nexpected). The key is fairness: equal or near-equal numbers of objective\nevaluations and identical data splits.\n\nSince this benchmark must run in under 15 minutes on one CPU core,\ndatasets should be small-to-medium sklearn/UCI-style classification tasks\nalready available in sklearn (e.g., wine, breast_cancer, digits). The\nanalysis should report best cross-validation score, held-out test accuracy,\nand search efficiency (score gain per evaluation or wall-clock).\nMulti-seed repeats are desirable to reduce noise, but can be kept modest to\nremain within runtime limits.\n\nThe experiment should answer whether Bayesian optimization delivers better\nbest-found configurations under tight evaluation budgets, or whether simpler\nbaselines (especially random search) are effectively equivalent for these\ntabular tasks. Interpretation should explicitly separate optimization quality\nfrom compute cost.\n\n*Under a fixed small evaluation budget, does Bayesian optimization find better random-forest hyperparameters than grid and random search on sklearn UCI-style classification benchmarks?*num_hypotheseshypotheses[{\: \, \: \, \: true}, {\: \, \: \, \: true}, {\: \, \: \, \: true}]experiment_design{\: \, \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: [\, \], \: [{\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}], \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: {\: false, \: 780}}requirementsrubric{\: \, \: \, \: \, \: 1, \: [{\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 3, \: [{\: \, \: \, \: 20.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}], \: null, \: null}], \: null, \: null}rubric_num_leavesmanifest_filetasks/ml/manifests/ML09.yamlrubric_filetasks/ml/rubrics/ML09.json |
| idML10domainmltitleCross-validation strategy reliability for small-sample model selectiontopicInvestigating the effectiveness of different cross-validation strategies (k-fold, stratified, repeated, leave-one-out) for model selection with small sample sizesdomainsmachine-learningmodel-selectionarxiv_idvenueARC-Bench 2026metric_keyestimation_biasmetric_directionminimizegpu_requiredest_wall_clock_secsynthesisWith small datasets, model selection can become highly sensitive to the\nvalidation protocol. Practitioners often choose among k-fold,\nstratified k-fold, repeated k-fold, and leave-one-out cross-validation\n(LOOCV), but these strategies trade off bias, variance, and compute in\ndifferent ways. A method that appears best under one CV scheme may not be\nbest under another, especially when class balance is imperfect and sample\ncounts are low.\n\nThis topic studies CV strategy as the independent variable while keeping\ncandidate models fixed and lightweight. The key quantity is estimation bias:\nthe gap between CV-estimated score used for model selection and the realized\ntest score of the selected model on a held-out test set. In small-sample\nsettings, minimizing this gap is often more important than maximizing raw CV\nscore, because over-optimistic selection can produce brittle deployments.\n\nA credible CPU-scale experiment should evaluate multiple sklearn datasets\nreduced to small training sizes, run several random seeds, and compare at\nleast four CV strategies under the same model grid and scoring rule. The\nstudy should report not only estimation bias but also variance across seeds\nand model-selection stability. Baselines should include standard k-fold and\nstratified k-fold; repeated stratified k-fold and LOOCV serve as contrasting\nalternatives with different variance/compute characteristics.\n\nThe goal is not to prove one universally superior protocol, but to quantify\nwhen more elaborate CV schemes improve reliability enough to justify their\ncost under strict CPU budgets.\n\n*Which cross-validation strategy yields the lowest and most stable model-selection estimation bias for small-sample classification tasks under a fixed candidate-model grid?*num_hypotheseshypotheses[{\: \, \: \, \: true}, {\: \, \: \, \: true}, {\: \, \: \, \: true}]experiment_design{\: \, \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: [\, \], \: [{\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}], \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: {\: false, \: 600}}requirementsrubric{\: \, \: \, \: \, \: 1, \: [{\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 2, \: [{\: \, \: \, \: 16.6667, \: [], \: \, \: \}, {\: \, \: \, \: 8.3333, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 3, \: [{\: \, \: \, \: 25.0, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}], \: null, \: null}], \: null, \: null}rubric_num_leavesmanifest_filetasks/ml/manifests/ML10.yamlrubric_filetasks/ml/rubrics/ML10.json |
| idML11domainmltitleBenchmarking classical unsupervised outlier detectors under controlled anomaly injectiontopicBenchmarking unsupervised outlier detection methods (IsolationForest, LocalOutlierFactor, OneClassSVM, EllipticEnvelope) on UCI datasets with injected anomaliesdomainsanomaly-detectionmachine-learningarxiv_idvenueARC-Bench 2026metric_keyroc_aucmetric_directionmaximizegpu_requiredest_wall_clock_secsynthesisClassical unsupervised anomaly-detection methods remain widely used in\nproduction tabular pipelines because they are interpretable enough,\nCPU-efficient, and available in standard libraries. Yet their relative\nbehavior can change sharply as anomaly prevalence and feature scaling vary.\nIsolationForest is often robust in heterogeneous feature spaces, while\ndistance/density methods like LocalOutlierFactor can be strong when local\nneighborhoods are informative but may degrade in higher dimensions.\n\nA practical benchmark should avoid external downloads and still provide\nrealistic control over anomaly rate. One reliable strategy is to start from\nclean sklearn classification datasets and inject synthetic anomalies into the\ntest split by sampling uniformly beyond feature-wise quantile ranges of the\ntraining data. This yields known binary ground truth for evaluation while\npreserving a realistic inlier distribution.\n\nA credible CPU-scale study here should compare IsolationForest,\nLocalOutlierFactor (novelty mode), OneClassSVM, and EllipticEnvelope under a\nshared preprocessing pipeline (e.g., StandardScaler), evaluate ROC-AUC and\nPR-AUC across at least two datasets and multiple seeds, and include a simple\nbaseline such as random scoring. Since contamination is rarely known exactly\nin practice, the benchmark should also probe at least two injected anomaly\nrates.\n\nThe main decision-relevant question is not whether any method can detect easy\nanomalies, but which method is consistently best across datasets and\ncontamination levels under strict CPU constraints. A concise, quantitative\nhypothesis-driven comparison is therefore the goal.\n\n*Which classical unsupervised outlier detector is most robust across tabular datasets and injected anomaly rates when measured by ROC-AUC and PR-AUC under a single-core CPU budget?*num_hypotheseshypotheses[{\: \, \: \, \: true}, {\: \, \: \, \: true}, {\: \, \: \, \: true}]experiment_design{\: \, \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: [\, \], \: [{\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}], \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: {\: false, \: 480}}requirementsrubric{\: \, \: \, \: \, \: 1, \: [{\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 3, \: [{\: \, \: \, \: 25.0, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}], \: null, \: null}], \: null, \: null}rubric_num_leavesmanifest_filetasks/ml/manifests/ML11.yamlrubric_filetasks/ml/rubrics/ML11.json |
| idML12domainmltitleComparing clustering algorithms on synthetic non-convex and anisotropic shapestopicComparing clustering algorithms (k-means, agglomerative, DBSCAN, HDBSCAN, spectral) on synthetic shapes (moons, circles, blobs, anisotropic) with ground-truth labelsdomainsmachine-learningclusteringarxiv_idvenueARC-Bench 2026metric_keyadjusted_rand_scoremetric_directionmaximizegpu_requiredest_wall_clock_secsynthesisClustering algorithms encode very different geometric assumptions. K-means\nprefers spherical, equal-variance groups; agglomerative clustering can adapt\nto hierarchical structure depending on linkage; DBSCAN finds dense regions\nand can recover non-convex shapes while labeling outliers as noise; spectral\nclustering can separate manifolds when graph affinity is appropriate.\nBecause these assumptions interact strongly with data geometry, a single\nbenchmark score often hides systematic failure modes.\n\nSynthetic datasets are ideal for a fast but meaningful comparison because we\ncan generate known structures with ground-truth labels: intertwined moons,\nconcentric circles, and anisotropic Gaussian blobs. These patterns expose\nstrengths and weaknesses that are difficult to isolate in real-world data.\nWith labeled synthetic data, external clustering metrics like adjusted rand\nindex (ARI) and normalized mutual information (NMI) directly quantify\npartition recovery quality.\n\nA credible CPU-scale study should compare several algorithms under a shared\nprotocol: standardized inputs, fixed train-size generation per seed, and\nexplicit handling of methods that output noise labels. It should include at\nleast one centroid baseline (k-means) and one density method (DBSCAN), then\ntest whether non-linear methods systematically outperform centroid methods on\nnon-convex shapes while avoiding large regressions on anisotropic blobs.\n\nThe practical goal is not to crown one universally best algorithm, but to\nestablish data-shape-dependent guidance. If shape complexity changes which\nmethod wins, practitioners should choose clustering tools by geometric prior\nrather than habit.\n\n*How strongly does dataset geometry (non-convex vs anisotropic) determine which clustering algorithm achieves the best ground-truth agreement under a fixed CPU-budget protocol?*num_hypotheseshypotheses[{\: \, \: \, \: true}, {\: \, \: \, \: true}, {\: \, \: \, \: true}]experiment_design{\: \, \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: [\, \], \: [{\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}], \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: {\: false, \: 420}}requirementsrubric{\: \, \: \, \: \, \: 1, \: [{\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 2, \: [{\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}, {\: \, \: \, \: 6.25, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 3, \: [{\: \, \: \, \: 25.0, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}, {\: \, \: \, \: 12.5, \: [], \: \, \: \}], \: null, \: null}], \: null, \: null}rubric_num_leavesmanifest_filetasks/ml/manifests/ML12.yamlrubric_filetasks/ml/rubrics/ML12.json |
| idML13domainmltitleEvaluating kernel choices for Gaussian Process regression on synthetic 1-D and 5-D functionstopicEvaluating kernel choices (RBF, polynomial, Matern-3/2, Matern-5/2) for Gaussian Process regression on synthetic 1-D and 5-D functions with homoscedastic noisedomainsmachine-learningkernel-methodsarxiv_idvenueARC-Bench 2026metric_keytest_nllmetric_directionminimizegpu_requiredest_wall_clock_secsynthesisGaussian Process (GP) regression performance depends strongly on the kernel,\nwhich encodes assumptions about function smoothness and local structure.\nPractitioners often default to the RBF kernel because it is easy to optimize\nand widely available, but this choice can underperform when the target\nfunction has roughness patterns that are better matched by Matern families or\nwhen polynomial trends dominate part of the signal. Since kernel\nmisspecification can be subtle, a short benchmark should compare predictive\nuncertainty quality as well as point error.\n\nA CPU-scale study can be done entirely with sklearn GaussianProcessRegressor\non synthetic datasets where the data-generating process is controlled. In 1-D,\na mildly nonstationary smooth function with additive Gaussian noise can expose\ndifferences in extrapolation and uncertainty tails. In 5-D, a mixed\nsinusoidal-plus-quadratic target can stress anisotropy and interactions while\nremaining fast enough for repeated fitting under multiple random seeds.\n\nTo make claims measurable, the experiment should evaluate at least four kernel\nchoices (RBF, polynomial, Matern-3/2, Matern-5/2) under the same train/test\nsplits, optimizer restarts, and noise model assumptions. Because the topic's\nkey metric is negative log-likelihood, uncertainty calibration quality should\nbe primary; RMSE and R^2 provide supporting evidence for point prediction.\nSeed averaging is important because synthetic splits and optimizer\ninitialization can change marginal likelihood optimization outcomes.\n\nThe core question is whether smoother kernels (RBF, Matern-5/2) consistently\noutperform rougher or mismatched kernels (Matern-3/2, polynomial) on test NLL\nacross both low- and moderate-dimensional synthetic tasks, and whether ranking\nby NLL aligns with ranking by RMSE.\n\n*Which GP kernel among RBF, polynomial, Matern-3/2, and Matern-5/2 yields the best uncertainty-aware generalization (lowest test NLL) on synthetic 1-D and 5-D noisy regression functions?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"Across the two synthetic datasets, at least one of {RBF, Matern-5/2} achieves the lowest mean test NLL on each dataset (averaged over >=5 seeds).\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"The polynomial kernel has mean test NLL at least 10% worse than the best kernel on at least 1 of the 2 datasets.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Kernel ranking by mean test NLL and by mean RMSE is not identical on at least 1 dataset, indicating uncertainty quality and point error disagree.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"How do RBF, polynomial, Matern-3/2, and Matern-5/2 kernels compare for Gaussian Process regression on synthetic 1-D and 5-D noisy functions when judged primarily by test NLL?\", \"conditions\": [{\"name\": \"gp_rbf\", \"description\": \"GaussianProcessRegressor with ConstantKernel * RBF + WhiteKernel, hyperparameters optimized by log-marginal-likelihood.\"}, {\"name\": \"gp_poly_deg2\", \"description\": \"GaussianProcessRegressor with ConstantKernel * DotProduct^2 (implemented via polynomial feature mapping + DotProduct kernel) + WhiteKernel as a polynomial-trend proxy.\"}, {\"name\": \"gp_matern32\", \"description\": \"GaussianProcessRegressor with ConstantKernel * Matern(nu=1.5) + WhiteKernel.\"}, {\"name\": \"gp_matern52\", \"description\": \"GaussianProcessRegressor with ConstantKernel * Matern(nu=2.5) + WhiteKernel.\"}], \"baselines\": [\"gp_rbf as the common default-kernel baseline\", \"gp_poly_deg2 as a mismatched-trend baseline\"], \"metrics\": [{\"name\": \"test_nll\", \"direction\": \"minimize\", \"description\": \"Mean negative log predictive density on held-out test data using GP predictive mean and variance, averaged over seeds.\"}, {\"name\": \"rmse\", \"direction\": \"minimize\", \"description\": \"Root mean squared error on test targets, averaged over seeds.\"}, {\"name\": \"r2\", \"direction\": \"maximize\", \"description\": \"Coefficient of determination on test targets, averaged over seeds.\"}], \"datasets\": [{\"name\": \"synthetic_1d_sinmix\", \"source\": \"Generated with numpy: x in [-3,3], y = sin(2x) + 0.3x + epsilon, epsilon~N(0, 0.15^2).\"}, {\"name\": \"synthetic_5d_mixed\", \"source\": \"Generated with numpy: X in [-2,2]^5, y = sin(x1) + 0.5*x2^2 - 0.7*x3 + 0.3*x4*x5 + epsilon, epsilon~N(0, 0.2^2).\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 360}}", "requirements": "", "rubric": "{\"id\": \"ml13-root\", \"requirements\": \"A credible experiment studying kernel-choice effects (RBF, polynomial, Matern-3/2, Matern-5/2, or equivalents) for Gaussian Process regression on synthetic noisy functions: conditions are implemented, execution covers multiple datasets with repeated seeds, and results address H1/H2/H3 directionally using test NLL as the primary metric.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative kernels or dataset dimensions that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml13-code\", \"requirements\": \"The GP kernel conditions and synthetic datasets are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml13-code-kernels\", \"requirements\": \"The submission implements multiple kernel conditions — typically including RBF, a polynomial kernel, and one or more Matern kernels — as distinct GP configurations with comparable noise handling.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml13-code-datasets\", \"requirements\": \"The submission generates synthetic regression datasets with explicit noise (both a low-dimensional and a higher-dimensional case preferred) and creates train/test splits.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml13-code-setup\", \"requirements\": \"Experimental setup controls are consistent across kernels (same split protocol, seed handling, and optimizer restart policy), enabling fair kernel comparison.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml13-exec\", \"requirements\": \"Execution logs primary and secondary metrics for each kernel and dataset.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml13-exec-metrics\", \"requirements\": \"Execution produces a metrics artifact containing numeric test NLL and RMSE (or equivalents) for every implemented (kernel, dataset) pair.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml13-exec-seeds\", \"requirements\": \"Reported metrics are aggregated over multiple random seeds per (kernel, dataset) cell, with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Hyperparameter Tuning\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml13-results\", \"requirements\": \"Results analysis addresses H1/H2/H3 directionally with quantitative comparisons.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml13-result-h1\", \"requirements\": \"The submission reports per-dataset mean test-NLL ranking and conveys whether smooth kernels (RBF, Matern-5/2) tend to be best — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml13-result-h2\", \"requirements\": \"The submission compares the polynomial-kernel NLL against the best kernel and conveys whether polynomial is meaningfully worse on at least one dataset (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml13-result-h3\", \"requirements\": \"The submission compares kernel rankings by test NLL and RMSE on each dataset and conveys whether rankings can differ between these two metrics (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml13-result-writeup\", \"requirements\": \"The README or writeup describes implementation choices, dataset generation, metric tables, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (synthetic-only scope, kernel parameterization, seed/compute constraints). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML13.yaml", "rubric_file": "tasks/ml/rubrics/ML13.json"} |
| {"id": "ML14", "domain": "ml", "title": "Comparing split-conformal, Mondrian conformal, and CQR-style intervals on heteroscedastic regression", "topic": "Comparing conformal prediction procedures (split-conformal, Mondrian, CQR) for achieving target coverage on heteroscedastic regression tasks", "domains": ["machine-learning", "uncertainty-quantification"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "coverage_gap", "metric_direction": "minimize", "gpu_required": false, "est_wall_clock_sec": 420, "synthesis": "Conformal prediction provides finite-sample coverage guarantees with minimal\nassumptions, making it attractive for uncertainty quantification in practical\nregression. However, standard split-conformal intervals are often globally\ncalibrated and can be inefficient when noise is heteroscedastic: they may\nover-cover low-noise regions and under-cover high-noise regions. This\nmotivates conditional variants such as Mondrian conformal (group-conditional\ncalibration) and CQR-style approaches that adapt interval width through\nquantile modeling before conformal correction.\n\nIn small, CPU-only settings, one can still run a meaningful comparison by\nusing sklearn-compatible regressors and synthetic data where heteroscedastic\nstructure is controlled. A rigorous setup should evaluate not only marginal\ncoverage but also efficiency (mean interval width) and conditional behavior\n(coverage gap across strata of predicted difficulty or input regions). These\ndiagnostics reveal when methods achieve nominal coverage by producing overly\nwide intervals versus genuinely adapting to varying noise.\n\nA credible benchmark should include at least one homoscedastic control and at\nleast one heteroscedastic dataset, then compare split-conformal, a Mondrian\nvariant (binning by an auxiliary score such as |x| or predicted scale), and a\nCQR-style method based on lower/upper quantile regressors plus conformal\nadjustment. Repeated random splits (multiple seeds) are needed because\nconformal procedures can vary with calibration sample composition.\n\nThe key question is whether adaptive procedures can reduce conditional\nmiscoverage while maintaining near-target marginal coverage and reasonable\ninterval width under strict runtime constraints.\n\n*Do Mondrian and CQR-style conformal procedures achieve lower conditional coverage error than vanilla split-conformal at comparable marginal coverage on heteroscedastic regression tasks?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"On heteroscedastic datasets, at least one adaptive method (Mondrian or CQR-style) attains a lower absolute coverage gap to the 90% target than split-conformal by at least 0.02 on at least 2 of 3 datasets, averaged over ≥5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"At matched target 90% coverage, CQR-style conformal yields mean interval width no larger than split-conformal on at least 2 of 3 datasets (difference ≤ 0.00).\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Mondrian conformal reduces worst-bin conditional miscoverage (max over 4 bins of |bin_coverage-0.90|) by at least 0.03 versus split-conformal on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Do adaptive conformal procedures (Mondrian, CQR-style) improve coverage quality and efficiency over split-conformal for heteroscedastic regression under a fixed 90% target coverage?\", \"conditions\": [{\"name\": \"split_conformal_rf\", \"description\": \"Vanilla split-conformal regression using RandomForestRegressor point predictor and absolute residual conformity scores on a held-out calibration split.\"}, {\"name\": \"mondrian_conformal_rf_4bins\", \"description\": \"Mondrian split-conformal with the same base predictor, calibration and test points partitioned into 4 bins by an auxiliary score (e.g., fitted value or |x0| proxy), with per-bin quantiles.\"}, {\"name\": \"cqr_gbr\", \"description\": \"CQR-style method using GradientBoostingRegressor quantile models (alpha=0.05, 0.95) and split-conformal correction on calibration residuals of quantile bands.\"}, {\"name\": \"naive_quantile_no_conformal\", \"description\": \"Non-conformal baseline using raw quantile regression interval [q0.05, q0.95] without conformal adjustment.\"}], \"baselines\": [\"split_conformal_rf is the primary conformal baseline\", \"naive_quantile_no_conformal is the non-conformal baseline\"], \"metrics\": [{\"name\": \"coverage_gap\", \"direction\": \"minimize\", \"description\": \"Absolute difference between empirical marginal coverage and target 0.90 on the test split, averaged over seeds.\"}, {\"name\": \"mean_interval_width\", \"direction\": \"minimize\", \"description\": \"Average prediction interval width on test samples, averaged over seeds.\"}, {\"name\": \"worst_bin_miscoverage\", \"direction\": \"minimize\", \"description\": \"Maximum across 4 bins of absolute deviation between bin coverage and 0.90, measuring conditional coverage disparity.\"}], \"datasets\": [{\"name\": \"hetero_sine\", \"source\": \"synthetic: y = sin(2πx) + (0.1 + 0.5|x|)ε, x~Uniform(-1,1), n≈2000\"}, {\"name\": \"hetero_friedman1\", \"source\": \"synthetic: sklearn.datasets.make_friedman1 with multiplicative noise scale depending on x0\"}, {\"name\": \"diabetes\", \"source\": \"sklearn.datasets.load_diabetes (tabular regression benchmark)\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml14-root\", \"requirements\": \"A credible experiment comparing split-conformal, Mondrian conformal, CQR-style, and optionally naive-quantile prediction intervals for ~90% target coverage on heteroscedastic regression: methods are implemented, executed on multiple datasets with multiple seeds, and results address H1/H2/H3 directionally with coverage-gap-focused analysis.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative conformal variants (e.g., jackknife+, locally adaptive) that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml14-code\", \"requirements\": \"The conformal and baseline interval-construction conditions are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml14-code-methods\", \"requirements\": \"The submission implements multiple distinct conditions — typically including split conformal, a Mondrian/group-adaptive variant, and a CQR-style or naive-quantile baseline — as separate code paths, not a single shared interval formula.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml14-code-conformal-splits\", \"requirements\": \"Conformal methods use explicit train/calibration/test separation (or equivalent cross-fit logic) so calibration quantiles are computed on data not used to fit base regressors.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml14-code-datasets\", \"requirements\": \"The submission uses multiple datasets (including at least one heteroscedastic synthetic dataset) and prepares regression targets correctly.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml14-exec\", \"requirements\": \"Execution outputs interval-quality metrics for each condition.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml14-exec-metrics\", \"requirements\": \"Execution produces a metrics artifact containing numeric coverage gap and mean interval width (or equivalents) for each implemented condition on at least one dataset.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml14-exec-seeds\", \"requirements\": \"Reported metrics are aggregated over multiple random seeds per (condition, dataset) cell with a dispersion estimate. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml14-exec-conditional\", \"requirements\": \"Execution computes a conditional-coverage diagnostic (e.g., worst-bin miscoverage across several bins) for the conformal methods.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml14-results\", \"requirements\": \"Quantitative analysis addresses H1/H2/H3 directionally and discusses trade-offs between coverage, conditional validity, and interval efficiency.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml14-result-h1\", \"requirements\": \"The submission compares coverage gap of adaptive methods (Mondrian, CQR-style) against plain split conformal per dataset and conveys whether adaptive methods are meaningfully closer to the nominal coverage — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml14-result-h2\", \"requirements\": \"The submission evaluates whether CQR-style intervals achieve mean interval width comparable to or narrower than plain split conformal on most datasets and conveys an H2 outcome.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml14-result-h3\", \"requirements\": \"The submission compares worst-bin miscoverage for Mondrian conformal vs plain split conformal and conveys whether Mondrian yields a meaningful improvement in conditional coverage (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml14-result-writeup\", \"requirements\": \"The README or writeup describes methods and datasets, reports key metric values (coverage gap, interval width, worst-bin miscoverage), conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (finite seeds, binning choice, synthetic-to-real transfer). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 10, "manifest_file": "tasks/ml/manifests/ML14.yaml", "rubric_file": "tasks/ml/rubrics/ML14.json"} |
| {"id": "ML15", "domain": "ml", "title": "Filter-based vs embedded L1 feature selection with injected noise features", "topic": "Investigating filter-based feature selection (mutual_info_classif, chi2, f_classif) vs embedded L1-logistic selection on UCI classification datasets with injected irrelevant features", "domains": ["machine-learning", "feature-selection"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "test_accuracy", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 420, "synthesis": "Feature selection on tabular classification problems is often presented as a\nchoice between simple univariate filters and embedded sparse models.\nUnivariate filters such as chi-squared, ANOVA F-score, and mutual\ninformation are computationally cheap and easy to apply in a pipeline, but\nthey evaluate each feature independently and can miss interactions. Embedded\nL1-regularized logistic regression performs selection during model fitting,\npotentially yielding feature subsets that better align with predictive\nstructure when many irrelevant variables are present.\n\nA practical stress test is to inject synthetic irrelevant features into\notherwise standard benchmark datasets. This setup creates controlled feature\ndilution while preserving original labels and base difficulty. Under such\ndilution, methods that can ignore noise should maintain accuracy and avoid\nselecting many synthetic features. Conversely, methods sensitive to spurious\nunivariate associations may degrade as noise dimension increases.\n\nA credible CPU-only study should compare at least three filter selectors\n(mutual_info_classif, chi2, f_classif) and an embedded L1 selector in a\nshared downstream classifier pipeline, evaluate on multiple sklearn\nclassification datasets, and report both predictive performance and selection\nquality. Including multiple random seeds is important because both noise\ninjection and train/test splits can change apparent selector rankings.\n\nThe key aim is not to reproduce a single published table, but to determine\nwhether embedded sparsity offers robustness advantages over filters as\nirrelevant dimensions grow, and whether that robustness generalizes across\ndatasets with different feature scales and class structures.\n\n*Does embedded L1-logistic feature selection retain predictive performance and reject injected irrelevant features more reliably than univariate filter methods under controlled feature-noise injection?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"With 5x injected irrelevant features (relative to original feature count), L1-logistic embedded selection achieves higher mean test_accuracy than each filter method (mutual_info_classif, chi2, f_classif) on at least 2 of 3 datasets, averaged over >=5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"At fixed selected-feature budget k=20 (or all features if p<20), L1-logistic selects a lower fraction of injected irrelevant features than the average of the three filter methods on all evaluated datasets.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"From no-injection to 5x-injection settings, mean test_accuracy drop for L1-logistic is <= 3 percentage points on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Does embedded L1-logistic feature selection outperform univariate filter methods in robustness to injected irrelevant features on sklearn tabular classification datasets?\", \"conditions\": [{\"name\": \"filter_mutual_info_k20\", \"description\": \"SelectKBest(mutual_info_classif, k=20 or p if p<20) followed by LogisticRegression classifier.\"}, {\"name\": \"filter_chi2_k20\", \"description\": \"MinMax scaling to nonnegative domain, then SelectKBest(chi2, k=20 or p if p<20) followed by LogisticRegression classifier.\"}, {\"name\": \"filter_f_classif_k20\", \"description\": \"SelectKBest(f_classif, k=20 or p if p<20) followed by LogisticRegression classifier.\"}, {\"name\": \"embedded_l1_logistic\", \"description\": \"L1-penalized LogisticRegression (saga/liblinear) used as embedded selector; keep top-20 by absolute coefficient magnitude (or nonzero if <=20), retrain LogisticRegression on selected subset.\"}], \"baselines\": [\"filter_f_classif_k20 as a standard univariate linear-statistic baseline\", \"filter_mutual_info_k20 as a nonlinear dependency baseline\"], \"metrics\": [{\"name\": \"test_accuracy\", \"direction\": \"maximize\", \"description\": \"Held-out test accuracy averaged over >=5 seeds for each dataset and injection level.\"}, {\"name\": \"noise_feature_selection_rate\", \"direction\": \"minimize\", \"description\": \"Fraction of selected features that come from injected irrelevant columns.\"}, {\"name\": \"accuracy_drop_0x_to_5x_pp\", \"direction\": \"minimize\", \"description\": \"Absolute percentage-point drop in test_accuracy from 0x to 5x injection.\"}], \"datasets\": [{\"name\": \"breast_cancer\", \"source\": \"sklearn.datasets.load_breast_cancer\"}, {\"name\": \"wine\", \"source\": \"sklearn.datasets.load_wine\"}, {\"name\": \"digits\", \"source\": \"sklearn.datasets.load_digits\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml15-root\", \"requirements\": \"A credible experiment studying filter-based feature selection (mutual information, chi-square, ANOVA F) versus embedded L1-logistic selection under injected irrelevant features: conditions are implemented, runs cover multiple datasets with multiple seeds and injection levels, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative filter methods or embedded selectors (e.g., tree-based importance) that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml15-code\", \"requirements\": \"Feature-selection conditions and the noise-injection pipeline are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml15-code-conditions\", \"requirements\": \"The submission implements multiple distinct selection conditions — typically including mutual information, chi-square, ANOVA F, and L1-logistic (or a comparable embedded method) — rather than reusing one selector for all conditions.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml15-code-injection\", \"requirements\": \"The pipeline injects synthetic irrelevant features at one or more defined levels (including a no-injection baseline and a higher-injection case) and tracks which columns are injected so a noise-selection rate can be computed.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml15-code-datasets\", \"requirements\": \"The submission uses multiple datasets (sklearn built-ins or comparable) and a reasonable train/test split.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml15-exec\", \"requirements\": \"Execution produces metrics across selectors and injection levels.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml15-exec-metrics\", \"requirements\": \"Execution outputs a structured metrics artifact containing numeric test accuracy and a noise-selection-rate style measure for each implemented condition on at least one dataset with at least two injection levels.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml15-exec-seeds\", \"requirements\": \"Reported metrics are aggregated over multiple random seeds per (dataset, condition, injection level) cell with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Hyperparameter Tuning\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml15-results\", \"requirements\": \"Results address H1/H2/H3 directionally with a clear narrative.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml15-result-h1\", \"requirements\": \"The submission compares mean test accuracy at the higher injection level between the embedded L1-logistic method and each filter method per dataset and conveys whether L1 tends to be better — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml15-result-h2\", \"requirements\": \"The submission reports the noise-selection rate for each method and conveys whether the embedded method selects fewer irrelevant features than the filter-method average on most datasets (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml15-result-h3\", \"requirements\": \"The submission reports the accuracy drop from low-injection to high-injection for the embedded method and conveys whether the drop is small on most datasets (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml15-result-writeup\", \"requirements\": \"The README or writeup describes the experimental setup, reports key metric tables, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations such as synthetic-noise realism, k-choice, classifier dependence, and seed/dataset scope. No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML15.yaml", "rubric_file": "tasks/ml/rubrics/ML15.json"} |
| {"id": "ML16", "domain": "ml", "title": "Bandit algorithm robustness under stationary and drifting reward regimes", "topic": "Comparing multi-armed bandit algorithms (epsilon-greedy, UCB1, Thompson sampling, Exp3) on synthetic Bernoulli and Gaussian arms under stationary and drift regimes", "domains": ["reinforcement-learning", "bandits"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "cumulative_regret", "metric_direction": "minimize", "gpu_required": false, "est_wall_clock_sec": 420, "synthesis": "Multi-armed bandit algorithms encode different assumptions about reward\ngeneration and non-stationarity. Epsilon-greedy is simple and adaptable with\npersistent exploration; UCB1 is optimism-driven and often strong in\nstationary stochastic settings; Thompson sampling can be highly sample\nefficient when model assumptions are matched; Exp3 is designed for\nadversarial settings and may trade off stochastic efficiency for robustness.\nIn small CPU-constrained studies, these differences are often discussed\nqualitatively but not tested with matched synthetic environments.\n\nA meaningful benchmark should evaluate both Bernoulli and Gaussian reward\narms, because posterior/model assumptions and noise scale can materially\nchange outcomes. It should also include stationary and drifting regimes:\nmethods tuned for fixed means can degrade when the best arm changes over\ntime. Regret, not just final reward, is the right primary metric because it\ncaptures online learning efficiency throughout the horizon.\n\nA credible experiment here implements the four algorithms with consistent\naction/reward interfaces, runs repeated simulations over multiple seeds, and\nreports cumulative regret trajectories and endpoint statistics. To keep the\nstudy CPU-friendly, horizons and arm counts should be modest (e.g., 5-10 arms,\n500-2000 steps), with vectorized numpy updates where possible.\n\nThe key analytical value is comparative: identify where UCB1/Thompson excel\nin stationary stochastic settings, and whether epsilon-greedy or Exp3 is more\nresilient under drift. The objective is not reproducing a single paper, but\ntesting algorithm-environment fit under controlled synthetic shifts.\n\n*How do epsilon-greedy, UCB1, Thompson sampling, and Exp3 trade off cumulative regret across Bernoulli vs Gaussian arms and stationary vs drifting reward regimes?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"In stationary Bernoulli environments, either UCB1 or Thompson sampling achieves at least 10% lower mean cumulative regret than epsilon-greedy at horizon T on at least 2 of 3 datasets (averaged over >=20 seeds).\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"In drifting environments, epsilon-greedy or Exp3 achieves lower mean cumulative regret than UCB1 on at least 2 of 3 datasets at horizon T (averaged over >=20 seeds).\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Across all evaluated datasets, no single algorithm is best (lowest mean cumulative regret) on every dataset, indicating environment-dependent performance.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"How do epsilon-greedy, UCB1, Thompson sampling, and Exp3 compare in cumulative regret across stationary versus drifting synthetic bandit tasks with Bernoulli and Gaussian rewards?\", \"conditions\": [{\"name\": \"epsilon_greedy_eps0.1\", \"description\": \"Epsilon-greedy with constant epsilon=0.1 and sample-mean value estimates.\"}, {\"name\": \"ucb1\", \"description\": \"UCB1 with exploration bonus sqrt(2 log t / n_a).\"}, {\"name\": \"thompson_sampling\", \"description\": \"Thompson sampling using Beta-Bernoulli updates for Bernoulli arms and Gaussian posterior sampling (known variance assumption) for Gaussian arms.\"}, {\"name\": \"exp3_gamma0.07\", \"description\": \"Exp3 with gamma=0.07 and importance-weighted reward estimates.\"}], \"baselines\": [\"epsilon_greedy_eps0.1 as simple stochastic baseline\", \"ucb1 as canonical optimism baseline\"], \"metrics\": [{\"name\": \"cumulative_regret\", \"direction\": \"minimize\", \"description\": \"Mean cumulative regret at final horizon T relative to oracle best arm per round, averaged over seeds.\"}, {\"name\": \"instantaneous_regret_auc\", \"direction\": \"minimize\", \"description\": \"Area under per-round regret curve over time (lower is better).\"}, {\"name\": \"best_arm_selection_rate\", \"direction\": \"maximize\", \"description\": \"Fraction of rounds selecting the current optimal arm (for drift: time-varying optimum).\"}], \"datasets\": [{\"name\": \"bernoulli_stationary_k10\", \"source\": \"synthetic: 10 Bernoulli arms with fixed means sampled in [0.05, 0.95] and sorted gap >=0.05\"}, {\"name\": \"bernoulli_drift_k10\", \"source\": \"synthetic: 10 Bernoulli arms with piecewise-constant means; best arm switches at predefined changepoints\"}, {\"name\": \"gaussian_drift_k8\", \"source\": \"synthetic: 8 Gaussian arms (sigma=1) with linearly drifting means and periodic rank reversals\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml16-root\", \"requirements\": \"A credible experiment comparing bandit algorithms (epsilon-greedy, UCB1, Thompson sampling, Exp3, or equivalents) under stationary and drifting synthetic regimes: algorithms are implemented with a common interface, experiments cover multiple environments with repeated seeds, and results address H1/H2/H3 directionally using cumulative regret.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Well-motivated algorithm variants (e.g., UCB-V, linear Thompson) should be credited when they test the same scientific question.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml16-code\", \"requirements\": \"The bandit algorithms and synthetic environments are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml16-code-algos\", \"requirements\": \"The submission implements multiple distinct algorithm code paths — typically including epsilon-greedy, UCB1, Thompson sampling, and/or Exp3 — with per-round action selection and update logic that are not identical wrappers.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml16-code-envs\", \"requirements\": \"The submission defines multiple synthetic bandit environments including at least one stationary and one drifting regime, with reproducible seed control and oracle best-arm rewards per round for regret computation.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml16-code-thompson-exp3\", \"requirements\": \"If Thompson sampling and/or Exp3 are included, implementation uses sampling-based posterior decisions (Thompson) and probability-weighted action selection with importance-weighted updates (Exp3), rather than greedy mean selection.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml16-exec\", \"requirements\": \"Execution produces regret metrics across algorithms and datasets.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml16-exec-runs\", \"requirements\": \"Execution runs multiple seeds per (algorithm, dataset) cell for multiple environments and logs final cumulative-regret values with mean and dispersion. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml16-exec-artifacts\", \"requirements\": \"A machine-readable results artifact is produced containing dataset-wise metrics for each implemented algorithm, including cumulative regret.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml16-results\", \"requirements\": \"Findings address H1/H2/H3 directionally and summarize implications.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml16-result-h1\", \"requirements\": \"The submission compares stationary-regime cumulative regret between epsilon-greedy and {UCB1, Thompson} and conveys whether the principled algorithms are meaningfully better — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml16-result-h2\", \"requirements\": \"The submission evaluates drifting-regime cumulative regret and conveys whether more exploratory algorithms (epsilon-greedy or Exp3) outperform UCB1 on most drifting datasets (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml16-result-h3\", \"requirements\": \"The submission conveys whether any single algorithm dominates across all environments or whether winners are mixed (H3), with supporting tables or summaries.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml16-result-writeup\", \"requirements\": \"The README or writeup describes setup, reports key cumulative-regret results per environment, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (horizon length, hyperparameter sensitivity, synthetic-only scope). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML16.yaml", "rubric_file": "tasks/ml/rubrics/ML16.json"} |
| {"id": "ML17", "domain": "ml", "title": "Topic-model comparison on small 20newsgroups subsets: LDA vs NMF vs LSA", "topic": "Comparing topic models (LDA, NMF, LSA) on a small subset of 20newsgroups by topic coherence (c_v) and document-cluster adjusted rand score", "domains": ["natural-language-processing", "topic-modeling"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "coherence_cv", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 720, "synthesis": "Topic modeling methods are often compared using either intrinsic quality\nmeasures (such as topic coherence) or extrinsic utility (such as how well\ndocument representations align with known labels), but the two views can\ndisagree. Latent Dirichlet Allocation (LDA), Non-negative Matrix\nFactorization (NMF), and Latent Semantic Analysis (LSA/SVD) each induce\ndifferent assumptions over term-document structure and may therefore rank\ndifferently depending on metric choice.\n\nOn CPU-constrained benchmarks, a practical study should use a compact text\ncorpus and a controlled preprocessing pipeline so that runtime stays short\nwhile still yielding meaningful distinctions. A small subset of\n20newsgroups is ideal because it has accessible labels for external\nclustering evaluation and enough lexical diversity for coherence analysis.\nUsing multiple subset difficulties (well-separated vs more confusable\ncategories) helps test robustness of conclusions.\n\nA credible experiment compares LDA, NMF, and LSA under matched topic counts\nand vectorization settings, then reports both c_v-like coherence and\nadjusted rand index (ARI) from document-cluster assignments against true\nnewsgroup labels. Because exact c_v implementations are not native in\nsklearn, an explicit approximation based on sliding-window co-occurrence\nand normalized PMI should be documented and applied consistently across\nmethods. Repeated runs for stochastic models are needed to avoid\nover-interpreting single-seed noise.\n\nThe core question is whether better intrinsic coherence implies better\nlabel alignment on small real-world corpora, and which model offers the\nbest trade-off under strict CPU budgets.\n\n*Do LDA, NMF, and LSA produce different rankings on coherence versus ARI on small 20newsgroups subsets, and does NMF provide the strongest coherence-structure trade-off under CPU limits?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"NMF achieves higher mean coherence_cv than LSA on at least 2 of 3 evaluated 20newsgroups subsets when using the same number of topics and preprocessing pipeline.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"Across the three methods (LDA, NMF, LSA), the method with the highest coherence_cv is also the highest-ARI method on no more than 1 of the 3 subsets, indicating weak metric agreement.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"LDA achieves mean ARI at least 0.03 higher than LSA on at least 2 of 3 subsets when document clusters are obtained by argmax topic assignment.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"How do LDA, NMF, and LSA compare on coherence_cv versus document-cluster ARI over small 20newsgroups subsets, and do intrinsic/extrinsic rankings diverge?\", \"conditions\": [{\"name\": \"lda_k4\", \"description\": \"LatentDirichletAllocation with n_components=4, max_iter=20, learning_method='batch', CountVectorizer features, repeated over 3 seeds.\"}, {\"name\": \"nmf_k4\", \"description\": \"NMF with n_components=4, init='nndsvda', solver='cd', max_iter=300 on TF-IDF features, repeated over 3 seeds.\"}, {\"name\": \"lsa_k4\", \"description\": \"TruncatedSVD (LSA) with n_components=4 on TF-IDF features; document-topic embeddings clustered via argmax after non-negative shift or via KMeans(k=4) with fixed seed.\"}, {\"name\": \"nmf_k6_sensitivity\", \"description\": \"NMF sensitivity condition with n_components=6 to test topic-count robustness for coherence and ARI trends.\"}], \"baselines\": [\"lsa_k4 as a linear-algebra baseline without probabilistic topic assumptions\", \"lda_k4 as a probabilistic-topic baseline\"], \"metrics\": [{\"name\": \"coherence_cv\", \"direction\": \"maximize\", \"description\": \"Approximate c_v coherence computed from top words per topic using sliding-window co-occurrence and NPMI aggregation, averaged across topics and seeds.\"}, {\"name\": \"ari\", \"direction\": \"maximize\", \"description\": \"Adjusted Rand Index between predicted document clusters (topic argmax or equivalent) and true newsgroup labels.\"}, {\"name\": \"runtime_sec\", \"direction\": \"minimize\", \"description\": \"Wall-clock training + inference time per condition/dataset cell on single CPU core.\"}], \"datasets\": [{\"name\": \"20ng_easy4\", \"source\": \"sklearn.datasets.fetch_20newsgroups with categories=['sci.space','rec.autos','comp.graphics','talk.politics.misc']\"}, {\"name\": \"20ng_related4\", \"source\": \"sklearn.datasets.fetch_20newsgroups with categories=['comp.graphics','comp.os.ms-windows.misc','comp.sys.ibm.pc.hardware','comp.sys.mac.hardware']\"}, {\"name\": \"20ng_science4\", \"source\": \"sklearn.datasets.fetch_20newsgroups with categories=['sci.space','sci.med','sci.electronics','sci.crypt']\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 720}}", "requirements": "", "rubric": "{\"id\": \"ml17-root\", \"requirements\": \"A credible experiment comparing LDA, NMF, and LSA topic models on small 20newsgroups subsets: conditions are implemented, execution reports coherence and ARI on multiple subsets with repeated seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative topic-model implementations or coherence approximations that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml17-code\", \"requirements\": \"Topic-model conditions and dataset pipeline are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml17-code-models\", \"requirements\": \"The submission implements LDA, NMF, and LSA (or equivalents such as a truncated SVD for LSA) as distinct modeling code paths with a matched topic-count setting.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml17-code-data\", \"requirements\": \"The submission loads multiple 20newsgroups subsets (or comparable document corpora) with consistent text preprocessing/vectorization documented in code.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml17-code-metrics-impl\", \"requirements\": \"The code computes a coherence-style metric (e.g., c_v or a documented approximation) and ARI from document-cluster assignments for each method.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml17-exec\", \"requirements\": \"Execution produces comparable numeric outputs for all implemented methods.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml17-exec-runs\", \"requirements\": \"Execution evaluates the core methods on multiple datasets and writes per-method, per-dataset coherence and ARI values to a machine-readable artifact.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml17-exec-seeds-runtime\", \"requirements\": \"Execution repeats stochastic methods with multiple seeds per dataset, reports dispersion, and logs a wall-clock timing measure per condition. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Experimental Setup\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml17-results\", \"requirements\": \"Quantitative analysis addresses H1/H2/H3 directionally.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml17-result-h1\", \"requirements\": \"The submission compares NMF vs LSA on coherence for each evaluated dataset and conveys whether NMF tends to yield better coherence — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml17-result-h2\", \"requirements\": \"The submission checks ranking agreement between coherence and ARI across methods per dataset and conveys whether coherence-ranking and ARI-ranking tend to diverge (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml17-result-h3\", \"requirements\": \"The submission reports LDA vs LSA ARI deltas per dataset and conveys whether LDA tends to achieve meaningfully higher document-cluster ARI (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml17-result-writeup\", \"requirements\": \"The README or writeup describes methods and preprocessing, reports coherence/ARI/runtime results, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (subset choice, coherence approximation validity, seed count, clustering-assignment assumptions). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML17.yaml", "rubric_file": "tasks/ml/rubrics/ML17.json"} |
| {"id": "ML18", "domain": "ml", "title": "Post-hoc calibration methods for sklearn classifiers on tabular benchmarks", "topic": "Evaluating probability calibration methods (Platt scaling, isotonic regression, temperature scaling) applied to scikit-learn classifiers (RandomForest, GBM, SVM-RBF) on UCI benchmarks", "domains": ["machine-learning", "calibration"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "ece", "metric_direction": "minimize", "gpu_required": false, "est_wall_clock_sec": 480, "synthesis": "Many widely used scikit-learn classifiers optimize discrimination rather\nthan probability quality. Random forests, gradient boosting models, and\nRBF-kernel SVMs can achieve strong accuracy while output probabilities are\nmiscalibrated, especially on small or moderately imbalanced tabular data.\nThis matters in decision settings where thresholds, risk ranking, or cost-\nsensitive actions depend on reliable confidence estimates.\n\nPost-hoc calibration methods are attractive because they can be layered onto\nexisting models without retraining the base learner. Platt scaling fits a\nsigmoid map, isotonic regression fits a non-parametric monotone map, and\ntemperature scaling applies a single-parameter logit rescaling (implemented\nfor binary tasks via optimization on held-out logits/probabilities). These\nmethods differ in flexibility and overfitting risk, so their relative value\nmay depend on classifier family and dataset size.\n\nA credible CPU-scale study should compare uncalibrated outputs versus at\nleast three calibrators across multiple sklearn tabular datasets, using a\nproper train/calibration/test protocol and repeated seeds. Because this is a\ncalibration-centric question, expected calibration error (ECE) and log loss\nshould be primary metrics, with accuracy used as a guardrail to ensure that\ncalibration does not degrade classification utility.\n\nThe goal is not to reproduce a single paper result but to test whether any\none calibrator is consistently superior across heterogeneous models and\ndatasets under tight compute constraints. The key uncertainty is whether\nmethod ranking is stable enough to justify a default choice.\n\n*Which post-hoc calibration method (Platt, isotonic, temperature) most reliably improves ECE across RF/GBM/SVM-RBF on small tabular sklearn benchmarks without materially harming accuracy?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"At least one post-hoc calibrator (Platt, isotonic, or temperature) reduces ECE by at least 10% relative to the uncalibrated model for at least 2 of 3 classifiers on at least 2 of 3 datasets, averaged over >=3 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"Isotonic regression achieves lower mean ECE than Platt scaling on at least 2 of 3 datasets when calibration-set size is >=150 samples.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"Applying post-hoc calibration changes test accuracy by no more than 1.0 absolute percentage point (mean over seeds) for each classifier-dataset pair.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which post-hoc calibration method (Platt, isotonic, temperature) most consistently improves probability calibration for RF/GBM/SVM-RBF on sklearn tabular datasets while preserving accuracy?\", \"conditions\": [{\"name\": \"uncalibrated\", \"description\": \"Base classifier probabilities/decision scores used directly with no calibration.\"}, {\"name\": \"platt_scaling\", \"description\": \"Sigmoid calibration fit on a held-out calibration split (CalibratedClassifierCV with method='sigmoid' or equivalent).\"}, {\"name\": \"isotonic_regression\", \"description\": \"Isotonic calibration fit on a held-out calibration split (CalibratedClassifierCV with method='isotonic' or equivalent).\"}, {\"name\": \"temperature_scaling\", \"description\": \"Single temperature parameter optimized on calibration split to minimize NLL; applied to logits/scores before probability mapping.\"}], \"baselines\": [\"uncalibrated outputs are the primary baseline\", \"platt_scaling serves as a classic parametric calibration baseline\"], \"metrics\": [{\"name\": \"ece\", \"direction\": \"minimize\", \"description\": \"Expected Calibration Error (10-15 bins) on the held-out test split, averaged over seeds.\"}, {\"name\": \"log_loss\", \"direction\": \"minimize\", \"description\": \"Negative log-likelihood on test probabilities.\"}, {\"name\": \"test_accuracy\", \"direction\": \"maximize\", \"description\": \"Classification accuracy on held-out test split to verify discrimination is preserved.\"}], \"datasets\": [{\"name\": \"breast_cancer\", \"source\": \"sklearn.datasets.load_breast_cancer\"}, {\"name\": \"wine\", \"source\": \"sklearn.datasets.load_wine\"}, {\"name\": \"digits_binary\", \"source\": \"sklearn.datasets.load_digits with target transformed to binary (e.g., digit<5 vs >=5)\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 480}}", "requirements": "", "rubric": "{\"id\": \"ml18-root\", \"requirements\": \"A credible experiment studying post-hoc probability calibration methods (Platt, isotonic, temperature, or equivalents) for sklearn classifiers: calibration conditions are implemented, execution covers multiple datasets/classifiers with repeated seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Well-motivated calibrator variants or alternative base classifiers that preserve the scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml18-code\", \"requirements\": \"Calibration methods and classifier setup are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml18-code-calibrators\", \"requirements\": \"The submission implements multiple calibration conditions — typically including uncalibrated, Platt, isotonic, and temperature scaling — as distinct code paths applied post-hoc to the same base model outputs.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml18-code-models\", \"requirements\": \"The experiment includes multiple target classifiers (e.g., RandomForest, gradient boosting, SVM-RBF, or equivalents) with consistent train/calibration/test handling.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml18-code-datasets\", \"requirements\": \"The submission uses multiple datasets (sklearn built-ins or comparable) and an explicit calibration split separate from train and test.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml18-exec\", \"requirements\": \"Execution reports calibration-focused metrics across conditions.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml18-exec-metrics\", \"requirements\": \"Execution outputs numeric ECE, log loss, and test accuracy (or equivalents) for each implemented condition on at least one dataset-classifier pair in a machine-readable artifact.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml18-exec-seeds\", \"requirements\": \"Metrics are aggregated over multiple random seeds per evaluated cell with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml18-results\", \"requirements\": \"Results quantitatively address H1/H2/H3 directionally with a clear narrative.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml18-result-h1\", \"requirements\": \"The submission computes relative ECE change versus uncalibrated and conveys whether at least one calibrator yields a meaningful ECE reduction across most classifier/dataset pairs — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml18-result-h2\", \"requirements\": \"The submission compares isotonic vs Platt mean ECE per dataset and conveys the relative calibration quality plus any calibration-set size considerations (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml18-result-h3\", \"requirements\": \"The submission reports accuracy deltas between calibrated and uncalibrated outputs per classifier-dataset pair and conveys whether accuracy changes are small (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml18-result-writeup\", \"requirements\": \"The README or report conveys per-hypothesis outcomes (supported / refuted / inconclusive), cites key ECE/log-loss/accuracy numbers, and discusses limitations (dataset scope, binning sensitivity, seed count, calibration split size). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML18.yaml", "rubric_file": "tasks/ml/rubrics/ML18.json"} |
| {"id": "ML19", "domain": "ml", "title": "Comparing graph- and pseudo-label-based semi-supervised learning on small tabular datasets", "topic": "Comparing semi-supervised learning strategies (LabelPropagation, LabelSpreading, SelfTrainingClassifier) on small UCI datasets with 10-20% labeled data", "domains": ["machine-learning", "semi-supervised-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "test_accuracy", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 420, "synthesis": "Semi-supervised learning is attractive when labels are scarce but unlabeled\nexamples are cheap. In sklearn, LabelPropagation and LabelSpreading provide\ngraph-based transductive approaches, while SelfTrainingClassifier offers a\npseudo-labeling wrapper around a standard supervised base learner. These\nmethods rely on different assumptions: graph smoothness over neighborhoods\nversus confidence-thresholded iterative self-labeling. On small tabular\ndatasets, their relative behavior can change quickly as the labeled fraction\ndrops from moderate (20%) to very low (10%).\n\nA useful CPU-scale study should compare these methods under the same splits,\nfeature preprocessing, and random seeds, with explicit control of labeled\nfraction. Because the key claim of semi-supervision is label efficiency, the\nexperiment should include a supervised-only baseline trained on the same\nlabeled subset and evaluated on a fixed held-out test set. This makes gains\nattributable to unlabeled-data usage rather than favorable splits.\n\nThe study should report not only test accuracy but also balanced accuracy and\nmacro-F1, since small tabular datasets can have class imbalance and accuracy\nalone may hide minority-class degradation. Repeating runs across several seeds\nis necessary because which points are labeled strongly affects outcomes at low\nlabel rates.\n\nPractically, the benchmark should stay lightweight: sklearn datasets only,\nno downloads, and model choices that complete within minutes on a single CPU\ncore. The focus is not SOTA performance but robust relative comparisons across\nlabel regimes and datasets.\n\n*How do LabelPropagation, LabelSpreading, and SelfTrainingClassifier compare in label-efficiency versus a supervised-only baseline when only 10–20% of training labels are available on small sklearn tabular datasets?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"At 10% labeled data, at least one semi-supervised method (LabelPropagation, LabelSpreading, or SelfTrainingClassifier) achieves test accuracy at least 3 percentage points higher than the supervised-only baseline on at least 2 of 3 datasets, averaged over ≥5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"Across datasets at 10% labeled data, LabelSpreading with RBF kernel attains mean test accuracy greater than or equal to LabelPropagation with RBF kernel in at least 2 of 3 datasets (seed-averaged).\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"For SelfTrainingClassifier, increasing labeled fraction from 10% to 20% improves mean test accuracy by at least 1 absolute percentage point on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"How do graph-based (LabelPropagation/LabelSpreading) and pseudo-labeling (SelfTrainingClassifier) methods compare to supervised-only training under 10% and 20% labeled-data regimes on small tabular benchmarks?\", \"conditions\": [{\"name\": \"supervised_logreg_labeled_only\", \"description\": \"LogisticRegression trained only on the labeled subset of the training split; unlabeled points ignored.\"}, {\"name\": \"label_propagation_rbf\", \"description\": \"LabelPropagation with RBF kernel; unlabeled training labels set to -1 and transductive predictions used for test inference.\"}, {\"name\": \"label_spreading_rbf\", \"description\": \"LabelSpreading with RBF kernel under the same masked-label protocol.\"}, {\"name\": \"self_training_logreg\", \"description\": \"SelfTrainingClassifier wrapping LogisticRegression with confidence threshold (e.g., 0.8), fit on mixed labeled/unlabeled training data.\"}], \"baselines\": [\"supervised_logreg_labeled_only is the primary baseline\", \"10% labeled regime serves as a lower-label baseline versus 20% within each method\"], \"metrics\": [{\"name\": \"test_accuracy\", \"direction\": \"maximize\", \"description\": \"Classification accuracy on held-out test split, averaged over seeds.\"}, {\"name\": \"balanced_accuracy\", \"direction\": \"maximize\", \"description\": \"Balanced accuracy on held-out test split, averaged over seeds.\"}, {\"name\": \"macro_f1\", \"direction\": \"maximize\", \"description\": \"Macro-averaged F1 score on held-out test split, averaged over seeds.\"}], \"datasets\": [{\"name\": \"breast_cancer\", \"source\": \"sklearn.datasets.load_breast_cancer\"}, {\"name\": \"wine\", \"source\": \"sklearn.datasets.load_wine\"}, {\"name\": \"digits\", \"source\": \"sklearn.datasets.load_digits\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml19-root\", \"requirements\": \"A credible experiment comparing semi-supervised methods (LabelPropagation, LabelSpreading, SelfTrainingClassifier, or equivalents) against a supervised-only baseline under limited-label regimes: methods are implemented correctly, runs cover multiple datasets with repeated seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Well-motivated alternative semi-supervised methods that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml19-code\", \"requirements\": \"Semi-supervised and baseline conditions are implemented correctly with explicit label masking.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml19-code-methods\", \"requirements\": \"The submission implements multiple distinct conditions — typically including at least one graph-based method (LabelPropagation or LabelSpreading), a self-training method, and a supervised-only baseline — as separate code paths.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml19-code-masking\", \"requirements\": \"Label masking for semi-supervised runs is explicit and correct: only a defined labeled fraction retains class labels and the remaining training labels are set to the unlabeled marker.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml19-code-datasets\", \"requirements\": \"The submission uses multiple datasets (sklearn built-ins or comparable), with a held-out test split and consistent preprocessing across conditions.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml19-exec\", \"requirements\": \"Execution covers multiple label fractions and produces benchmark metrics.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml19-exec-metrics\", \"requirements\": \"Execution outputs numeric test accuracy and at least one additional metric (e.g., balanced accuracy or macro-F1) for each implemented (condition, dataset, label-fraction) cell in a machine-readable artifact.\", \"weight\": 16.6667, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml19-exec-seeds\", \"requirements\": \"Each reported cell is aggregated over multiple random seeds with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 8.3333, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml19-results\", \"requirements\": \"Results address H1/H2/H3 directionally with quantitative comparisons.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml19-result-h1\", \"requirements\": \"The submission compares each semi-supervised method versus the supervised-only baseline at a low labeled fraction and conveys whether any method provides a meaningful gain on most datasets — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml19-result-h2\", \"requirements\": \"The submission reports a direct LabelSpreading vs LabelPropagation comparison per dataset and conveys whether LabelSpreading is at least competitive on most datasets (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml19-result-h3\", \"requirements\": \"The submission reports self-training accuracy at both low and higher labeled fractions and conveys whether the gain with more labels is meaningful (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml19-result-writeup\", \"requirements\": \"The README or writeup describes setup and methods, reports per-dataset metrics, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and discusses limitations (dataset scope, seed count variance, sensitivity to masking/hyperparameters). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML19.yaml", "rubric_file": "tasks/ml/rubrics/ML19.json"} |
| {"id": "ML20", "domain": "ml", "title": "Classical forecaster robustness on synthetic seasonal time series", "topic": "Benchmarking classical time-series forecasters (AR, ARIMA, SARIMAX, ETS, Theta) on synthetic seasonal series with varying noise, trend, and seasonality amplitudes", "domains": ["time-series", "forecasting"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "smape", "metric_direction": "minimize", "gpu_required": false, "est_wall_clock_sec": 720, "synthesis": "Classical univariate forecasting methods remain widely used because they are\ninterpretable and lightweight, yet their comparative behavior can flip under\ndifferent data-generating regimes. AR models often perform well when dynamics\nare mostly autoregressive and weakly seasonal; ARIMA can absorb trend and\ndifferencing structure; SARIMAX can explicitly represent seasonal lag effects;\nETS captures error-trend-seasonal decomposition; and Theta is often a strong\nlow-variance baseline. In small CPU-constrained settings, practitioners need\npractical guidance about which method is robust to changing signal-to-noise\nand seasonality strength.\n\nA synthetic benchmark is appropriate here because we can precisely control\ntrend slope, seasonal amplitude, and observation noise while keeping compute\nmodest. By generating multiple families of monthly-like series with known\nperiodicity (e.g., period 12), we can evaluate whether model rankings are\nstable or regime-dependent. This avoids overfitting conclusions to one\nreal-world dataset and enables direct stress testing under high-noise versus\nhigh-seasonality conditions.\n\nA credible study should compare at least four of {AR, ARIMA, SARIMAX, ETS,\nTheta} on 2–3 synthetic dataset families, using rolling-origin or holdout\nforecasts with a fixed horizon. It should report scale-robust error metrics\n(especially sMAPE), include at least one baseline (e.g., seasonal naive), and\naggregate over multiple random seeds. The analysis should explicitly test\nwhether one method is consistently best overall or whether the best choice\ndepends on data regime.\n\nThe key outcome is actionable selection logic: if seasonal structure is strong\nand noise is low, do seasonal/state-space methods dominate; and if noise is\nhigh, do simpler methods become competitive? These are concrete decisions an\nautonomous forecasting pipeline can apply when only limited data and CPU are\navailable.\n\n*How do AR, ARIMA, SARIMAX, ETS, and Theta compare in sMAPE robustness across synthetic seasonal series with controlled trend, seasonality amplitude, and noise levels?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"SARIMAX or ETS achieves the lowest mean sMAPE on at least 2 of 3 synthetic dataset families when seasonality amplitude is high (amplitude >= 8) and noise is low (sigma <= 1.0), averaged over >=5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"Under high-noise settings (sigma >= 3.0), the sMAPE gap between the best advanced method (ARIMA/SARIMAX/ETS/Theta) and a seasonal_naive baseline is <= 5 percentage points on at least 2 of 3 dataset families.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"No single method among {AR, ARIMA, SARIMAX, ETS, Theta} ranks first in mean sMAPE on all evaluated dataset families, indicating regime-dependent winners.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which classical forecasting method is most robust in sMAPE across controlled synthetic regimes of trend, seasonality amplitude, and noise?\", \"conditions\": [{\"name\": \"ar_lag12\", \"description\": \"Autoregressive model with lag order selected from {3,6,12} by AIC on training split.\"}, {\"name\": \"arima_auto_small\", \"description\": \"Non-seasonal ARIMA with (p,d,q) searched over small grid p,q in [0,2], d in [0,1], selected by AIC.\"}, {\"name\": \"sarimax_seasonal12\", \"description\": \"Seasonal ARIMA/SARIMAX with seasonal period 12 and small grid over (p,d,q)x(P,D,Q), selected by AIC.\"}, {\"name\": \"ets_additive\", \"description\": \"Exponential smoothing with additive trend/seasonality (period 12), damped trend optional by AIC.\"}, {\"name\": \"theta\", \"description\": \"ThetaModel forecast with default theta decomposition for univariate series.\"}], \"baselines\": [\"seasonal_naive (forecast y_t = y_{t-12}) as a simple seasonal baseline\", \"naive_last (random walk) as a non-seasonal baseline\"], \"metrics\": [{\"name\": \"smape\", \"direction\": \"minimize\", \"description\": \"Symmetric Mean Absolute Percentage Error on forecast horizon, averaged over seeds and series instances.\"}, {\"name\": \"mae\", \"direction\": \"minimize\", \"description\": \"Mean Absolute Error on forecast horizon.\"}, {\"name\": \"fit_time_sec\", \"direction\": \"minimize\", \"description\": \"Per-method wall-clock fit+forecast runtime in seconds.\"}], \"datasets\": [{\"name\": \"syn_low_noise_high_season\", \"source\": \"Synthetic monthly series: length 180, period 12, trend slope in [0.0,0.2], season amplitude in [8,12], gaussian noise sigma in [0.5,1.0].\"}, {\"name\": \"syn_medium_noise_medium_season\", \"source\": \"Synthetic monthly series: length 180, period 12, trend slope in [0.1,0.4], season amplitude in [4,8], gaussian noise sigma in [1.5,2.5].\"}, {\"name\": \"syn_high_noise_low_season\", \"source\": \"Synthetic monthly series: length 180, period 12, trend slope in [0.0,0.3], season amplitude in [1,4], gaussian noise sigma in [3.0,4.0].\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 720}}", "requirements": "", "rubric": "{\"id\": \"ml20-root\", \"requirements\": \"A credible experiment benchmarking classical forecasters (AR, ARIMA, SARIMAX, ETS, Theta, or equivalents) on synthetic seasonal series: model conditions are implemented, execution covers multiple synthetic regimes with repeated seeds, and results address H1/H2/H3 directionally using sMAPE-centered analysis.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative forecaster implementations or substitutes that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml20-code\", \"requirements\": \"The forecasting conditions and synthetic data generation pipeline are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml20-code-methods\", \"requirements\": \"The submission implements multiple distinct forecasting methods — typically AR/ARIMA, a seasonal model (SARIMAX or ETS), and Theta or equivalents — as separate code paths, and includes at least one explicit naive/seasonal-naive baseline.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml20-code-synthdata\", \"requirements\": \"The submission generates multiple synthetic seasonal dataset families with controllable seasonality, trend, and noise, each with a train/test split suitable for forecasting.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml20-code-setup\", \"requirements\": \"Forecasting setup uses a fixed holdout horizon (or rolling-origin equivalent) and computes comparable forecasts for every (method, dataset, seed) cell.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml20-exec\", \"requirements\": \"Benchmark execution logs required metrics with repeated trials.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml20-exec-metrics\", \"requirements\": \"Execution produces a structured metrics artifact containing numeric sMAPE and MAE (or equivalents) for each implemented method on at least one dataset family.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml20-exec-seeds\", \"requirements\": \"Each reported method-dataset metric is aggregated over multiple random seeds with dispersion statistics. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml20-exec-runtime\", \"requirements\": \"The run logs per-method wall-clock timing and completes within a CPU-only budget.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml20-results\", \"requirements\": \"Results analysis addresses H1/H2/H3 directionally with quantitative comparisons.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml20-result-h1\", \"requirements\": \"The submission isolates high-seasonality low-noise results and conveys whether SARIMAX or ETS tends to attain the best mean sMAPE — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml20-result-h2\", \"requirements\": \"The submission compares the best advanced method against seasonal-naive under high-noise settings and conveys whether the sMAPE gap collapses to a small margin (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml20-result-h3\", \"requirements\": \"The submission ranks methods by mean sMAPE per dataset family and conveys whether no single method dominates across all families (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml20-result-writeup\", \"requirements\": \"The README or writeup reports key metric tables, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and discusses limitations (synthetic realism, grid-search scope, seed count, horizon sensitivity). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 10, "manifest_file": "tasks/ml/manifests/ML20.yaml", "rubric_file": "tasks/ml/rubrics/ML20.json"} |
| {"id": "ML21", "domain": "ml", "title": "PC vs GES vs NOTEARS-linear for recovering small Gaussian linear-SEM DAGs", "topic": "Comparing causal structure learning algorithms (PC, GES, NOTEARS-linear) on small synthetic linear-SEM DAGs of 5-15 nodes with Gaussian noise", "domains": ["causal-inference", "machine-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "structural_hamming_distance", "metric_direction": "minimize", "gpu_required": false, "est_wall_clock_sec": 720, "synthesis": "Causal structure learning methods often claim asymptotic guarantees, but in\npractical settings researchers usually face small graphs, finite samples, and\nimperfect hyperparameter choices. For Gaussian linear structural equation\nmodels (SEMs), three widely discussed paradigms are constraint-based (PC),\nscore-based (GES), and continuous optimization (NOTEARS-linear). Even when\ndata are generated from the model class these methods assume, finite-sample\nbehavior can differ sharply in edge recovery and orientation errors.\n\nA compact CPU-scale benchmark can isolate these differences by generating\nsynthetic DAGs with known ground truth over 5-15 nodes, sampling linear-Gaussian\nobservations, and evaluating structural recovery against the true adjacency.\nThis avoids data-download overhead and lets the study vary sample size and\ngraph density in a controlled way. The key metric is Structural Hamming\nDistance (SHD), which penalizes missing, extra, and wrongly oriented edges.\n\nA credible study should implement all three algorithms with transparent\nassumptions: PC via partial-correlation CI tests, GES via greedy score search\n(e.g., BIC), and NOTEARS-linear via a differentiable acyclicity penalty and\nL1 regularization. Multiple random seeds and at least two synthetic regimes\n(sparser/easier vs denser/harder) are needed to reduce variance and avoid\nover-interpreting one-off runs.\n\nBecause these algorithms expose different tuning knobs (significance level,\nregularization strength, score penalties), the comparison should include one\nclear default setting per method plus modest tuning on validation seeds. The\nresulting analysis should report SHD and at least one precision/recall-style\nedge metric, then map outcomes back to explicit hypotheses about relative\nranking and sample-efficiency.\n\n*On small Gaussian linear-SEM DAGs, which of PC, GES, and NOTEARS-linear most reliably minimizes structural recovery error under realistic finite-sample budgets?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"At n_samples=1000, NOTEARS-linear achieves lower mean SHD than PC on at least 2 of 3 synthetic DAG regimes (averaged over >=5 random DAG/data seeds per regime).\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"At n_samples=300, GES achieves mean SHD <= PC in all evaluated regimes, and strictly lower SHD in at least 1 regime.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"The best method at n_samples=1000 reduces mean SHD by at least 20% relative to the worst method in at least 2 of 3 regimes.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"For 5-15 node Gaussian linear-SEM DAGs, how do PC, GES, and NOTEARS-linear compare in SHD and edge recovery across sparse/medium/dense synthetic regimes and finite sample sizes?\", \"conditions\": [{\"name\": \"pc_partial_corr\", \"description\": \"Constraint-based PC using Gaussian partial-correlation conditional independence tests with significance level alpha (default 0.01, optional small grid).\"}, {\"name\": \"ges_bic\", \"description\": \"Score-based greedy equivalence search using Gaussian/BIC score with forward-backward edge operations.\"}, {\"name\": \"notears_linear_l1\", \"description\": \"Continuous optimization NOTEARS-linear with squared loss, acyclicity constraint, and L1 sparsity penalty tuned on a small lambda grid.\"}, {\"name\": \"pc_alpha_tuned\", \"description\": \"PC variant selecting alpha from {0.005, 0.01, 0.05} by lowest validation-seed SHD, then evaluated on held-out seeds.\"}], \"baselines\": [\"pc_partial_corr is the classical constraint-based baseline\", \"ges_bic is the classical score-based baseline\"], \"metrics\": [{\"name\": \"shd\", \"direction\": \"minimize\", \"description\": \"Structural Hamming Distance between learned and true DAG adjacency (lower is better), averaged over seeds.\"}, {\"name\": \"edge_f1\", \"direction\": \"maximize\", \"description\": \"F1 score on directed edge existence/orientation against ground truth adjacency.\"}, {\"name\": \"runtime_sec\", \"direction\": \"minimize\", \"description\": \"Per-run wall-clock time in seconds to assess CPU practicality.\"}], \"datasets\": [{\"name\": \"synth_linear_sem_sparse\", \"source\": \"Synthetic 8-node DAGs generated via Erdos-Renyi with expected degree ~1.5; linear weights sampled from [-2,-0.5] U [0.5,2], Gaussian noise.\"}, {\"name\": \"synth_linear_sem_medium\", \"source\": \"Synthetic 12-node DAGs with expected degree ~2.5 using same linear-Gaussian SEM generator.\"}, {\"name\": \"synth_linear_sem_dense\", \"source\": \"Synthetic 15-node DAGs with expected degree ~3.5 using same linear-Gaussian SEM generator.\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 720}}", "requirements": "", "rubric": "{\"id\": \"ml21-root\", \"requirements\": \"A credible experiment comparing causal structure learning methods (PC, GES, NOTEARS-linear, or equivalents) on small synthetic Gaussian linear-SEM DAGs: methods are implemented, runs cover multiple synthetic regimes with multiple seeds, and results address H1/H2/H3 directionally using SHD-centered evidence.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative structure-learning methods or linear-SEM variants that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml21-code\", \"requirements\": \"Core causal structure learning conditions and synthetic data generation are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml21-code-methods\", \"requirements\": \"The submission implements distinct code paths for multiple structure learning methods — typically PC, GES, and NOTEARS-linear or equivalents — rather than reusing one estimator under different names.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml21-code-synth\", \"requirements\": \"A reproducible synthetic linear-Gaussian SEM DAG generator is implemented with moderate node counts, including ground-truth adjacency export and sample generation.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml21-code-hparams\", \"requirements\": \"At least one tunable hyperparameter is exposed and used for the constraint-based method (e.g., PC's alpha) and the continuous-optimization method (e.g., NOTEARS' sparsity weight), with documented default or grid values.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Hyperparameter Tuning\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml21-exec\", \"requirements\": \"The benchmark produces comparable metrics across methods/regimes.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml21-exec-coverage\", \"requirements\": \"Execution covers multiple synthetic regimes and reports results for the primary methods at one or more sample sizes.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml21-exec-metrics\", \"requirements\": \"A machine-readable results artifact includes numeric SHD for each (method, regime) and at least one additional metric such as edge-F1 or runtime.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml21-exec-seeds\", \"requirements\": \"Each reported cell is averaged over multiple random seeds with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml21-results\", \"requirements\": \"Results analysis addresses H1/H2/H3 directionally.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml21-result-h1\", \"requirements\": \"The submission compares NOTEARS-linear vs PC at a reasonable sample size and conveys whether NOTEARS achieves meaningfully lower SHD on most regimes — judge directionally against H1.\", \"weight\": 25.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml21-result-h2h3\", \"requirements\": \"The submission conveys qualitative verdicts for H2 and H3 using SHD aggregates, including whether larger sample sizes or denser-graph regimes yield meaningful SHD improvements.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml21-result-writeup\", \"requirements\": \"The README or writeup describes setup, reports key SHD/edge-F1/runtime findings, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (synthetic-only scope, small-node regime, hyperparameter sensitivity, finite seeds). No strict word-count requirement.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML21.yaml", "rubric_file": "tasks/ml/rubrics/ML21.json"} |
| {"id": "ML22", "domain": "ml", "title": "Active learning query strategies for logistic regression on small tabular pools", "topic": "Evaluating active learning query strategies (uncertainty sampling, margin sampling, query-by-committee, expected-error reduction) for logistic regression on small UCI pools", "domains": ["machine-learning", "active-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "accuracy_at_budget", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 600, "synthesis": "Pool-based active learning is attractive when labels are expensive but a\nmodest unlabeled pool is available. For linear probabilistic models such as\nlogistic regression, classic query policies include uncertainty sampling,\nmargin sampling, query-by-committee (QBC), and expected-error reduction\nstyle lookahead approximations. These methods are often taught as broadly\nuseful, but on small tabular datasets their gains can be inconsistent due to\nmodel misspecification, class imbalance, and high variance from tiny initial\nlabeled sets.\n\nA useful CPU-bounded benchmark should compare these query strategies under a\nfixed annotation budget and shared learner, rather than mixing in model\narchitecture changes. The core signal is label efficiency: how quickly test\nperformance improves as labels are acquired. Because final accuracy alone can\nhide early-budget differences, the study should include both\naccuracy-at-budget and a budget-curve summary metric such as area under the\nlearning curve.\n\nThe experiment should therefore run multiple seeds, use at least two\nsklearn-resident tabular classification datasets, and enforce identical\ninitialization, batch size, and stopping budget across strategies. A random\nquery baseline is essential to determine whether sophisticated policies\nprovide real value beyond chance selection. A passive full-data reference is\nalso useful for contextualizing attainable performance ceilings under the\nsame logistic-regression family.\n\nPractical constraints matter: expected-error reduction can be expensive if\nnaively recomputing retrains for every candidate. A tractable variant can\nevaluate a capped candidate subset per round and use one-step hypothetical\nlabel outcomes, preserving the spirit of expected future loss minimization\nwhile staying within single-core runtime limits.\n\n*Do uncertainty, margin, QBC, and approximate expected-error reduction deliver better label efficiency than random sampling for logistic regression on small tabular pools under a fixed annotation budget?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"At 30% labeling budget, at least 2 of {uncertainty_sampling, margin_sampling, qbc, expected_error_reduction} achieve higher mean test accuracy than random_sampling by ≥1.5 percentage points, averaged over ≥5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"expected_error_reduction attains the highest mean area_under_learning_curve (AULC) on at least 1 of 3 datasets, and its AULC is not lower than random_sampling on any evaluated dataset.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"qbc outperforms uncertainty_sampling in mean test accuracy at early budget (10% labels) on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which active learning query strategy provides the best label efficiency for logistic regression on small tabular pool-based classification tasks under fixed budgets?\", \"conditions\": [{\"name\": \"random_sampling\", \"description\": \"Baseline pool-based active learning that queries unlabeled points uniformly at random each round.\"}, {\"name\": \"uncertainty_sampling\", \"description\": \"Queries samples with highest predictive entropy from current logistic regression model.\"}, {\"name\": \"margin_sampling\", \"description\": \"Queries samples with smallest top-2 class probability margin (binary: |p-0.5|).\"}, {\"name\": \"qbc\", \"description\": \"Query-by-committee with 5 bootstrapped logistic regression committee members; selects by vote entropy/disagreement.\"}, {\"name\": \"expected_error_reduction\", \"description\": \"Approximate one-step expected-error reduction over a capped candidate subset per round using hypothetical labels and expected future log-loss reduction.\"}], \"baselines\": [\"random_sampling is the primary active-learning baseline\", \"passive_full_data logistic regression reference trained once on all labels for context\"], \"metrics\": [{\"name\": \"accuracy_at_budget\", \"direction\": \"maximize\", \"description\": \"Test accuracy at 30% labeled budget, averaged over seeds.\"}, {\"name\": \"aulc\", \"direction\": \"maximize\", \"description\": \"Area under the test-accuracy-vs-labeled-fraction curve from initial seed set to 30% budget.\"}, {\"name\": \"early_accuracy_10pct\", \"direction\": \"maximize\", \"description\": \"Test accuracy when 10% of pool labels have been acquired.\"}], \"datasets\": [{\"name\": \"breast_cancer\", \"source\": \"sklearn.datasets.load_breast_cancer\"}, {\"name\": \"wine\", \"source\": \"sklearn.datasets.load_wine\"}, {\"name\": \"digits\", \"source\": \"sklearn.datasets.load_digits\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 600}}", "requirements": "", "rubric": "{\"id\": \"ml22-root\", \"requirements\": \"A credible experiment studying active-learning query strategies (random, uncertainty, margin, QBC, expected error reduction, or equivalents) for logistic regression: strategies are implemented as distinct code paths, execution covers multiple datasets with repeated seeds under a fixed budget, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Alternative acquisition functions or base classifiers that preserve the scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml22-code\", \"requirements\": \"Active-learning strategies and shared logistic-regression pipeline are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml22-code-strategies\", \"requirements\": \"The submission implements multiple distinct query conditions — typically random plus several of {uncertainty, margin, QBC, expected error reduction} — with genuinely different acquisition logic.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml22-code-loop\", \"requirements\": \"There is a pool-based active-learning loop with an initial labeled seed set, iterative querying, model retraining/update, and stopping at a defined label budget.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml22-code-datasets\", \"requirements\": \"The submission uses multiple datasets (sklearn built-ins or comparable) with consistent train/pool/test handling for all strategies.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml22-exec\", \"requirements\": \"Execution emits budget-aware performance metrics.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml22-exec-metrics\", \"requirements\": \"Execution produces a machine-readable metrics artifact containing numeric accuracy-at-budget and area-under-learning-curve (or equivalents) for each implemented strategy on at least one dataset.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml22-exec-seeds\", \"requirements\": \"Each reported (strategy, dataset) result is aggregated over multiple random seeds, with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml22-exec-budget-curve\", \"requirements\": \"The run logs performance across multiple budget checkpoints so early-budget accuracy and AULC are computable from recorded traces.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml22-results\", \"requirements\": \"Quantitative analysis addresses H1/H2/H3 directionally.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml22-result-h1\", \"requirements\": \"The submission compares non-random strategies to random sampling at the final budget and conveys whether informative strategies meaningfully outperform random — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml22-result-h2\", \"requirements\": \"The submission reports per-dataset AULC rankings and conveys whether expected-error-reduction (or a comparable principled strategy) is at least competitive, never falling below random (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml22-result-h3\", \"requirements\": \"The submission compares QBC vs uncertainty sampling at an early-budget checkpoint per dataset and conveys a qualitative verdict (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml22-result-writeup\", \"requirements\": \"The README or report describes setup, reports key numeric outcomes, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (runtime approximations, seed count, dataset scope, strategy sensitivity). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 10, "manifest_file": "tasks/ml/manifests/ML22.yaml", "rubric_file": "tasks/ml/rubrics/ML22.json"} |
| {"id": "ML23", "domain": "ml", "title": "Pointwise vs pairwise vs listwise learning-to-rank on synthetic query-document benchmarks", "topic": "Comparing learning-to-rank approaches (pointwise linear, pairwise RankNet-lite, listwise ListMLE-lite) on synthetic query-document pairs with known relevance grades", "domains": ["information-retrieval", "machine-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "ndcg_at_10", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 480, "synthesis": "Learning-to-rank (LTR) objectives are often grouped into pointwise, pairwise,\nand listwise families, each optimizing a different surrogate of retrieval\nquality. Pointwise methods reduce ranking to regression/classification on\nindividual query-document pairs; pairwise methods optimize relative order\nbetween document pairs; listwise methods directly shape permutations or top-k\nemphasis. In production IR systems, metric alignment (especially with NDCG)\nis often cited as a reason to prefer pairwise/listwise objectives, but this\nclaim is difficult to test cleanly on public collections with many confounds.\n\nA compact synthetic benchmark with known latent relevance structure allows a\ncontrolled comparison. If query-specific utility functions generate graded\nrelevance labels (0-4), we can produce query groups with explicit train/test\nsplits and evaluate whether objectives aligned with ranking structure obtain\nbetter NDCG@10 than a strong linear pointwise baseline. Because the data are\nsynthetic, we can vary noise and query heterogeneity while keeping runtime low\nenough for single-core CPU execution.\n\nA credible study should implement three distinct training paradigms in the\nsame feature space: (1) pointwise linear regression on grades, (2) a\nRankNet-lite pairwise logistic objective over within-query document pairs,\nand (3) a ListMLE-lite listwise objective using per-query permutations sorted\nby true grades. Evaluation should report NDCG@10 as primary, plus at least one\nsecondary ranking metric and a sanity metric such as training loss. Results\nshould be averaged across multiple seeds and at least two synthetic dataset\nregimes.\n\nThe key scientific goal is not reproducing a specific paper, but testing\nwhether increasing objective-level ranking awareness yields measurable gains\nin top-k ranking quality under controlled conditions and limited compute.\n\n*Do pairwise and listwise lightweight objectives consistently outperform a pointwise linear baseline on NDCG@10 across synthetic query-document datasets with graded relevance?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"ListMLE-lite achieves higher mean NDCG@10 than pointwise linear regression on at least 2 of 3 synthetic datasets, with an absolute improvement of at least 0.03 on each winning dataset (averaged over >=3 seeds).\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"At least one of {RankNet-lite, ListMLE-lite} outperforms pointwise linear regression in mean NDCG@10 on all evaluated datasets.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"The best ranking-aware method (max of RankNet-lite and ListMLE-lite per dataset) improves mean MAP@10 over pointwise linear by at least 0.02 on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Do pairwise/listwise lightweight ranking objectives provide consistent top-k retrieval gains over a pointwise linear baseline on synthetic grouped query-document data?\", \"conditions\": [{\"name\": \"pointwise_linear\", \"description\": \"Linear model trained pointwise on query-document feature vectors to predict graded relevance via squared error; ranking by predicted score within each query.\"}, {\"name\": \"ranknet_lite\", \"description\": \"Pairwise logistic ranking objective on within-query document pairs using a linear scorer; optimized with mini-batch gradient descent in numpy.\"}, {\"name\": \"listmle_lite\", \"description\": \"Listwise ListMLE-style objective with a linear scorer, optimizing likelihood of ground-truth relevance-sorted document permutations per query.\"}, {\"name\": \"pointwise_ridge_tuned\", \"description\": \"Pointwise linear baseline with L2 regularization tuned over a small grid to ensure a competitive non-ranking-aware baseline.\"}], \"baselines\": [\"pointwise_linear is the primary baseline\", \"pointwise_ridge_tuned is a strengthened linear baseline\"], \"metrics\": [{\"name\": \"ndcg_at_10\", \"direction\": \"maximize\", \"description\": \"Mean NDCG@10 across test queries; primary metric.\"}, {\"name\": \"map_at_10\", \"direction\": \"maximize\", \"description\": \"Mean Average Precision@10 across test queries using relevance>=1 as relevant.\"}, {\"name\": \"pairwise_accuracy\", \"direction\": \"maximize\", \"description\": \"Fraction of correctly ordered document pairs within test queries (based on graded relevance order).\"}], \"datasets\": [{\"name\": \"synth_ltr_easy\", \"source\": \"Synthetic generator: 80 queries x 25 docs/query, 20 dense features, low noise, relevance grades 0-4 from latent linear utility.\"}, {\"name\": \"synth_ltr_noisy\", \"source\": \"Synthetic generator: 100 queries x 30 docs/query, 25 features, higher Gaussian noise and feature corruption, grades 0-4.\"}, {\"name\": \"synth_ltr_sparse\", \"source\": \"Synthetic generator: 90 queries x 20 docs/query, 40 features with sparsity mask, query-specific weight drift, grades 0-4.\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 480}}", "requirements": "", "rubric": "{\"id\": \"ml23-root\", \"requirements\": \"A credible experiment comparing pointwise, pairwise (RankNet-style), and listwise (ListMLE-style) ranking approaches on synthetic query-document datasets: methods are implemented as distinct objectives, execution reports ranking metrics on multiple datasets with multiple seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Well-motivated ranking-objective substitutes (e.g., LambdaRank-style, softrank) that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml23-code\", \"requirements\": \"The ranking methods and synthetic grouped datasets are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml23-code-methods\", \"requirements\": \"The submission implements multiple distinct conditions — typically a pointwise baseline plus at least one pairwise and one listwise method — with genuinely different objective computations, not merely different hyperparameters of one loss.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml23-code-data\", \"requirements\": \"A synthetic query-document generator is implemented with grouped queries, per-query document lists, and graded relevance labels, and multiple dataset regimes are instantiated.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}, {\"id\": \"ml23-code-setup\", \"requirements\": \"The experimental setup includes query-level train/test splitting (no leakage of documents from the same query across splits) and a shared scoring-function interface across methods.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml23-exec\", \"requirements\": \"Execution logs ranking metrics per method and dataset.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml23-exec-metrics\", \"requirements\": \"Execution produces a metrics artifact containing numeric NDCG@k and MAP@k (or equivalents) for each implemented condition on at least one dataset.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml23-exec-seeds\", \"requirements\": \"Reported metrics are averaged over multiple random seeds per (condition, dataset) cell with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml23-exec-tuning\", \"requirements\": \"At least one non-trivial hyperparameter search or documented default is executed (e.g., learning rate/regularization) and the chosen value is logged.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Hyperparameter Tuning\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml23-results\", \"requirements\": \"Results analysis addresses H1/H2/H3 directionally.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml23-result-h1\", \"requirements\": \"The submission compares the listwise method vs the pointwise baseline in mean NDCG@k per dataset and conveys whether listwise meaningfully improves ranking quality — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml23-result-h2h3\", \"requirements\": \"The submission conveys whether at least one ranking-aware method beats the pointwise baseline on NDCG across all datasets (H2) and whether the best ranking-aware method yields a meaningful MAP improvement on most datasets (H3).\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml23-result-writeup\", \"requirements\": \"The README or writeup describes methods, datasets, and metric outcomes; conveys per-hypothesis outcomes (supported / refuted / inconclusive); and discusses limitations (synthetic-data realism, scorer capacity, seed count, metric sensitivity). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 9, "manifest_file": "tasks/ml/manifests/ML23.yaml", "rubric_file": "tasks/ml/rubrics/ML23.json"} |
| {"id": "ML24", "domain": "ml", "title": "Online binary classification under concept drift: SGD, PA, NB, and FTRL", "topic": "Benchmarking online learning algorithms (SGD logistic, Passive-Aggressive, online Naive Bayes, Follow-the-Regularized-Leader) on streaming binary classification with concept drift", "domains": ["online-learning", "machine-learning"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "prequential_accuracy", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 420, "synthesis": "In streaming classification, data arrive sequentially and model updates must\nbe cheap and immediate. Under concept drift, the relationship between\nfeatures and labels changes over time, so static train/test evaluation can\nbe misleading. Prequential (test-then-train) evaluation better reflects\ndeployment: each incoming point is first predicted, then used for an update.\nThis setting creates practical trade-offs among stability, adaptability, and\ncomputational cost.\n\nSeveral lightweight online learners in sklearn-style workflows represent\ndifferent adaptation biases. SGD logistic regression can track drift via\ngradient updates but may require careful learning-rate choices. Passive-\nAggressive updates can react strongly to mistakes and often adapt quickly to\nabrupt shifts. Online Naive Bayes is extremely cheap and robust but may lag\nwhen feature dependence structure changes. A simple FTRL-style proximal\nlogistic update (implemented with diagonal accumulators) offers adaptive\nper-feature learning rates and implicit regularization.\n\nA credible CPU-only benchmark should compare these methods on at least one\nsynthetic abrupt-drift stream and one real sklearn dataset converted to a\nstream with induced drift (e.g., blockwise class-prior or feature transform\nshift). It should report prequential accuracy as the primary metric, plus at\nleast one drift-sensitive secondary metric such as post-drift recovery and\ncumulative log loss. Multi-seed runs are important because stream order and\ndrift-point randomness can materially change outcomes.\n\nThe study should also verify that drift actually hurts models by comparing a\nno-drift control stream against drifted streams, and should discuss which\nalgorithms recover fastest after drift while maintaining overall accuracy.\nThis makes the benchmark more diagnostic than a single aggregate score.\n\n*Which lightweight online learner provides the best trade-off between overall prequential accuracy and adaptation speed after concept drift in binary streams?*", "num_hypotheses": 3, "hypotheses": "[{\"id\": \"H1\", \"statement\": \"On drifted streams, Passive-Aggressive or FTRL achieves higher prequential_accuracy than online Naive Bayes by at least 0.02 absolute on at least 2 of 3 datasets, averaged over >=5 seeds.\", \"measurable\": true}, {\"id\": \"H2\", \"statement\": \"For at least 2 of 3 drifted datasets, the best post-drift recovery accuracy in a fixed window of 200 samples after each drift point is achieved by either Passive-Aggressive or FTRL.\", \"measurable\": true}, {\"id\": \"H3\", \"statement\": \"For SGD logistic and Passive-Aggressive, prequential_accuracy on a drifted version of each stream is at least 0.03 lower than on its matched no-drift control, on at least 2 of 3 datasets.\", \"measurable\": true}]", "experiment_design": "{\"research_question\": \"Which online method among SGD logistic, Passive-Aggressive, online Naive Bayes, and FTRL-style logistic best balances prequential accuracy and post-drift recovery on binary data streams with concept drift?\", \"conditions\": [{\"name\": \"sgd_logistic\", \"description\": \"SGDClassifier with log_loss and partial_fit in prequential test-then-train loop.\"}, {\"name\": \"passive_aggressive\", \"description\": \"PassiveAggressiveClassifier with partial_fit in the same loop.\"}, {\"name\": \"online_naive_bayes\", \"description\": \"GaussianNB updated incrementally via partial_fit.\"}, {\"name\": \"ftrl_prox_logistic\", \"description\": \"Custom numpy FTRL-proximal logistic regression with diagonal accumulator and L1/L2 regularization.\"}], \"baselines\": [\"online_naive_bayes as a simple low-cost baseline\", \"sgd_logistic as a standard linear online-learning baseline\"], \"metrics\": [{\"name\": \"prequential_accuracy\", \"direction\": \"maximize\", \"description\": \"Mean test-then-train accuracy over the full stream, averaged over 5 seeds.\"}, {\"name\": \"post_drift_recovery_accuracy\", \"direction\": \"maximize\", \"description\": \"Accuracy in the first 200 samples after each drift point, averaged across drift events and seeds.\"}, {\"name\": \"prequential_log_loss\", \"direction\": \"minimize\", \"description\": \"Cumulative mean log loss computed prequentially when probabilistic outputs are available (or clipped score-to-prob mapping).\"}], \"datasets\": [{\"name\": \"synthetic_abrupt_drift\", \"source\": \"numpy synthesis (piecewise make_classification with coefficient/sign flips at fixed indices)\"}, {\"name\": \"breast_cancer_stream\", \"source\": \"sklearn.datasets.load_breast_cancer transformed into a repeated/shuffled stream with blockwise feature scaling shift\"}, {\"name\": \"spambase_like_synthetic\", \"source\": \"numpy synthesis with sparse informative features and class-prior shift across blocks\"}], \"compute_requirements\": {\"gpu_required\": false, \"estimated_wall_clock_sec\": 420}}", "requirements": "", "rubric": "{\"id\": \"ml24-root\", \"requirements\": \"A credible experiment studying online learners (SGD logistic, Passive-Aggressive, online Naive Bayes, FTRL-style logistic, or equivalents) under concept drift: streaming prequential setup is implemented, execution covers multiple datasets with multiple seeds, and results address H1/H2/H3 directionally.\", \"judging_note\": \"Score on scientific substance and directional correctness of evidence, not on exact numeric thresholds. Well-motivated online-learner substitutes or drift-stream variants that test the same scientific question should be credited.\", \"weight\": 1, \"sub_tasks\": [{\"id\": \"ml24-code\", \"requirements\": \"Online-learning conditions and drifted streaming setup are implemented correctly.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml24-code-methods\", \"requirements\": \"The submission implements multiple distinct online methods, each as a real incremental update path (e.g., partial_fit or equivalent), not batch retraining.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Method Implementation\"}, {\"id\": \"ml24-code-streaming\", \"requirements\": \"A prequential test-then-train loop is implemented: each sample (or mini-batch) is predicted before being used for model update, with ordered stream processing.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Experimental Setup\"}, {\"id\": \"ml24-code-datasets\", \"requirements\": \"The submission uses multiple streams including at least one drifted stream and one no-drift or matched-control stream, generated from sklearn datasets or numpy synthesis.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Development\", \"finegrained_task_category\": \"Dataset and Model Acquisition\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml24-exec\", \"requirements\": \"Execution reports prequential metrics for each condition.\", \"weight\": 2, \"sub_tasks\": [{\"id\": \"ml24-exec-metrics\", \"requirements\": \"Execution produces a metrics artifact containing numeric prequential accuracy (or equivalent) for each implemented method on at least one dataset.\", \"weight\": 12.5, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml24-exec-recovery\", \"requirements\": \"Execution includes a post-drift recovery measure (e.g., fixed-window post-drift accuracy) with a documented computation referencing known drift points.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml24-exec-seeds\", \"requirements\": \"Each reported (dataset, method) metric is averaged over multiple random seeds, with a dispersion measure. Honest small-seed runs with variance reported are preferable to a single run.\", \"weight\": 6.25, \"sub_tasks\": [], \"task_category\": \"Code Execution\", \"finegrained_task_category\": \"Hyperparameter Tuning\"}], \"task_category\": null, \"finegrained_task_category\": null}, {\"id\": \"ml24-results\", \"requirements\": \"Quantitative analysis addresses H1/H2/H3 directionally.\", \"weight\": 3, \"sub_tasks\": [{\"id\": \"ml24-result-h1\", \"requirements\": \"The submission compares prequential accuracy of margin-based online methods (Passive-Aggressive, FTRL) against online Naive Bayes across drifted datasets and conveys whether margin-based methods are meaningfully better — judge directionally against H1.\", \"weight\": 20.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml24-result-h2\", \"requirements\": \"The submission reports post-drift recovery comparisons and conveys whether Passive-Aggressive or FTRL tends to lead on most datasets (H2).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Evaluation, Metrics & Benchmarking\"}, {\"id\": \"ml24-result-h3\", \"requirements\": \"The submission compares drifted vs no-drift control performance for the implemented methods and conveys whether drift produces a meaningful accuracy degradation (H3).\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}, {\"id\": \"ml24-result-writeup\", \"requirements\": \"The README or writeup describes setup, key metrics, conveys per-hypothesis outcomes (supported / refuted / inconclusive), and notes limitations (stream realism, drift design choices, seed count, metric assumptions). No strict word-count requirement.\", \"weight\": 10.0, \"sub_tasks\": [], \"task_category\": \"Result Analysis\", \"finegrained_task_category\": \"Logging, Analysis & Presentation\"}], \"task_category\": null, \"finegrained_task_category\": null}], \"task_category\": null, \"finegrained_task_category\": null}", "rubric_num_leaves": 10, "manifest_file": "tasks/ml/manifests/ML24.yaml", "rubric_file": "tasks/ml/rubrics/ML24.json"} |
| {"id": "ML25", "domain": "ml", "title": "Reservoir computing vs MLP vs GP for short-horizon Lorenz-63 forecasting", "topic": "Testing whether a random-feature echo state network (reservoir computing) predicts Lorenz-63 short-horizon trajectories more accurately than a matched-parameter MLP and a Gaussian Process at small training-data sizes (N=200-2000 points)", "domains": ["dynamical-systems", "machine-learning", "reservoir-computing"], "arxiv_id": null, "venue": "ARC-Bench 2026", "metric_key": "valid_prediction_time", "metric_direction": "maximize", "gpu_required": false, "est_wall_clock_sec": 720, "synthesis": "Reservoir computing (a.k.a. echo state networks, ESN) is a lightweight\napproach for learning dynamical systems: a fixed random recurrent network\nprojects the input into a high-dimensional reservoir state, and only a\nlinear readout layer is trained. For chaotic systems, ESNs have repeatedly\nbeen reported to match or outperform trainable recurrent nets on\nshort-horizon trajectory prediction, despite training only a ridge\nregression on the readout.\n\nA credible CPU-scale study of this claim must compare (i) a reservoir\nnetwork with a random sparse internal matrix and fixed spectral radius,\n(ii) a parameter-matched fully-connected MLP that reads a fixed window of\npast states, and (iii) a Gaussian Process regressor with an RBF kernel on\nthe same windowed input. The target signal is Lorenz-63 integrated at\ndt=0.02 with the classical (σ=10, ρ=28, β=8/3) parameters. Training-set\nsizes N ∈ {200, 500, 1000, 2000} are small enough that the standard\nreservoir-computing folklore — \"ESNs win when data is scarce\" — can\nactually be tested rather than assumed.\n\nThe key dynamics-aware metric is the *valid prediction time* (VPT):\nthe first time step at which the normalized prediction error exceeds a\nthreshold (commonly 0.4 of the attractor standard deviation). Unlike\nper-step RMSE, VPT tracks how long a model's forecast remains useful on\nthe chaotic attractor. Reporting both RMSE and VPT reveals whether any\nobserved \ is numerical or dynamical.\n\nThe research question is: *does a small reservoir (N_res ≤ 500 units) beat\na matched-parameter MLP and a Gaussian Process on valid prediction time\nfor Lorenz-63, and at which training-set size does the gap appear?*num_hypotheseshypotheses[{\: \, \: \, \: true}, {\: \, \: \, \: true}, {\: \, \: \, \: true}]experiment_design{\: \, \: [{\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}, {\: \, \: \}], \: [\, \], \: [{\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}, {\: \, \: \, \: \}], \: [{\: \, \: \}, {\: \, \: \}], \: {\: false, \: 720}}requirementsrubric{\: \, \: \, \: \, \: 1, \: [{\: \, \: \, \: 2, \: [{\: \, \: \, \: 8.3333, \: [], \: \, \: \}, {\: \, \: \, \: 8.3333, \: [], \: \, \: \}, {\: \, \: \, \: 4.1667, \: [], \: \, \: \}, {\: \, \: \, \: 4.1667, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 2, \: [{\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 5.0, \: [], \: \, \: \}], \: null, \: null}, {\: \, \: \, \: 3, \: [{\: \, \: \, \: 20.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}, {\: \, \: \, \: 10.0, \: [], \: \, \: \}], \: null, \: null}], \: null, \: null}rubric_num_leavesmanifest_filetasks/ml/manifests/ML25.yamlrubric_filetasks/ml/rubrics/ML25.json |
| |
