Hugging Face's logo

Datasets:
AIMING-Lab-UNC
/
ARC-Bench

Tasks:
Other
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
autonomous-research
ai-agents
llm-agents
scientific-discovery
benchmark
machine-learning
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ARC-Bench / tasks /quantum /README.md
StarThomas1002's picture
StarThomas1002
Add ARC-Bench: 55-topic autonomous-research benchmark across 5 domains
7ce68e5 verified
preview code
|
raw
history blame contribute delete
7.45 kB

ARC-Bench Quantum Domain

Ten open-ended quantum machine learning and variational quantum algorithm topics that run on the autoclaw native sandbox (Qiskit 2.x stack, statevector and finite-shot simulation, no GPU, no IBM Quantum hardware, no external agent).

Added in branch feat/quantum-domain. Same rc_full runner that drives the ML domain (run_bench.py), same stages 10 to 14, same paperbench_finalize + judge post-processing.

Layout 

config/quantum/
  topics.yaml              registry of 10 topics (Q01-Q10)
  manifests/Q0N.yaml       per-topic manifest: synthesis, hypotheses, experiment_design, rubric_path
  rubrics/Q0N.json         per-topic 3-bucket rubric (code / exec / results)
  README.md                this file

results/rc_full/Q0N/<run_id>/
  bench_meta.json          run config and bench inputs
  judge_result.json        per-leaf grades + overall_score
  claims.json              extracted quantitative claims
  RESULTS_README.md        auto-generated run summary
  submission/              trimmed agent output (code + figures + writeup)

log/rc_full/Q0N/<run_id>/full_run/
  full stage-07..14 archive. Gitignored.

Topic registry

ID Theme Primary metric Direction
Q01 Data encoding strategies for VQC test_accuracy maximize
Q02 CNOT ablation in VQC test_accuracy maximize
Q03 Classical optimizers for VQE on H2 shots_to_chemical_accuracy minimize
Q04 Data re-uploading depth scaling test_mse minimize
Q05 Barren plateau cost locality log_gradient_variance maximize
Q06 NN warm-start for QAOA MaxCut iterations_to_target_ratio minimize
Q07 MPS classifier vs NN baselines test_accuracy maximize
Q08 Layerwise vs end-to-end VQC test_accuracy maximize
Q09 Noise-aware VQC training test_accuracy maximize
Q10 Quantum autoencoder fidelity reconstruction_fidelity maximize

Full topic strings live in topics.yaml. Per-topic hypotheses H1 / H2 / H3, conditions, baselines, and seed counts live in manifests/Q0N.yaml.

Latest single-run scores (rc_full, gpt-5.3-codex + gpt-4o judge)

Methodology: one latest run per topic, no best-of-N cherry picking.

ID Score Run timestamp
Q01 0.709 20260517-192423
Q02 0.430 20260517-192423
Q03 0.757 20260517-212933
Q04 0.571 20260517-212933
Q05 0.576 20260517-212933
Q06 0.317 20260517-212933
Q07 0.513 20260517-212933
Q08 0.421 20260517-212720
Q09 0.145 20260517-192423
Q10 0.421 20260517-192422
Mean 0.486

Stage-15 PROCEED gate requires at least 2 baselines plus the proposed method, so every quantum manifest declares two or more baselines (random / random_init_control / classical MLP / logistic-regression as appropriate).

How to run

Single topic:

python experiments/arc_bench/scripts/run_bench.py \
  --mode rc_full \
  --topic Q03 \
  --runs 1

All quantum topics:

python experiments/arc_bench/scripts/run_bench.py \
  --mode rc_full \
  --domain quantum \
  --runs 1

The runner reads base_config.yaml, expands Q0N to the matching manifests/Q0N.yaml and rubrics/Q0N.json, registers Q -> quantum in prepare_run.py, and drops outputs under results/rc_full/Q0N/<run_id>/.

Skill

Agents pick up Qiskit-specific patterns from a domain skill:

researchclaw/skills/builtin/domain/quantum-qiskit/SKILL.md

Triggered on stages 10 and 13 whenever the topic synthesis or experiment plan mentions Qiskit, VQE, VQC, QAOA, EfficientSU2, ZZFeatureMap, MPS, noise model, etc. The skill is intentionally generic. It contains no Q-topic narratives and no bench-specific signal that would leak rubric information to the agent.

Key patterns it documents:

  1. Imports for Qiskit 2.x (StatevectorEstimator, StatevectorSampler, BackendSamplerV2, AerSimulator).
  2. Data-encoding feature maps (ZFeatureMap, ZZFeatureMap, StatePreparation).
  3. Variational ansatze (EfficientSU2, n-local).
  4. VQC training with qiskit_machine_learning.VQC.fit.
  5. Manual VQE loop pattern. qiskit_algorithms.VQE is broken under Qiskit 2.x because qiskit_nature.second_q.algorithms imports the removed BaseEstimator. The skill shows the StatevectorEstimator + optimizer.minimize() replacement.
  6. MPS-structured circuits via AerSimulator(method='matrix_product_state', matrix_product_state_max_bond_dimension=chi).
  7. Noise model wiring. qiskit_machine_learning.Sampler does not accept a noise_model kwarg, so noisy training routes through BackendSamplerV2(backend=AerSimulator(noise_model=...)).
  8. Qiskit 2.x compatibility notes and a table of common errors with fixes.

Dependencies

Pinned in config/base_config.yaml allowed_imports:

qiskit, qiskit_aer, qiskit_algorithms, qiskit_machine_learning, qiskit_nature, pyscf
numpy, scipy, matplotlib, sklearn, pandas, statsmodels, networkx, skimage

The agent runs inside a sandbox so it cannot install new packages mid-run. Anything outside this list fails the import gate at stage 10.

Caveats

  • LLM nondeterminism. Single-run scores carry std around 0.2 to 0.4. Compare topics with caution. The committed runs are one snapshot. A rerun on the same manifest can swing materially.
  • Q09 is hard. Noise-aware training under our sandbox time budget regularly bottoms out near the random-prediction floor. The 0.145 score reflects this. The skill already documents the correct noise model wiring, so the failure mode is not a missing pattern but a genuinely thin training signal under the configured noise rates.
  • Multi-file consistency. Topics such as Q08 require the agent to keep main.py, model.py, evaluate.py, etc. in sync. When the agent gets out of sync, the run fails at stage 12 with a missing-import error. This is an agent capability ceiling and is not fixable by skill updates.
  • Code bucket weight. Rubrics carry a Code Development bucket (weight 2 of 7), but the autoclaw default judge.py operates in results_only mode and skips code-leaf grading. Code leaves are kept in the rubric so that a future code-aware judge can score them without re-authoring.

Adding a new quantum topic

  1. Add an entry to topics.yaml with a Q11+ id, a verbose topic string, domains, metric_key, metric_direction.
  2. Write manifests/Q11.yaml with synthesis (research question + skill pointer + protocol), hypotheses (H1 / H2 / H3), experiment_design (conditions, at least 2 baselines, metrics, datasets, compute_requirements), and a rubric_path pointing to rubrics/Q11.json.
  3. Write rubrics/Q11.json as a 3-bucket tree (code / exec / results), 25 / 25 / 50 weights, with 9 leaves total.
  4. Make sure scripts/prepare_run.py _PREFIX_MAP and scripts/judge.py _PREFIX_MAP both map Q -> quantum. They already do for Q01-Q10.
  5. Run end-to-end with --runs 1 first to shake out manifest typos before committing.