huminclu/autoscirub-rcb-main-exp

AutoSciRub Main Experiments on ResearchClawBench

This repository is an artifact archive for the main ResearchClawBench experiments in:

Learning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents
Xuehai Wang, Haowei Qin, Tongxin Liu, Junkai Li, Buqiang Xu, Jintian Zhang, Yijun Chen, Zirui Xue, and Shumin Deng.
arXiv:2608.31076

It contains generated scientific reports, figures, code, supporting outputs, run metadata, and evaluator records for the paper's six paired vanilla versus AutoSciRub configurations on the 40-task ResearchClawBench benchmark. It is an experiment-output archive, not a replacement for the benchmark's task dataset.

The source benchmark is available from the ResearchClawBench GitHub repository and the official Hugging Face dataset.

What is ResearchClawBench?

ResearchClawBench evaluates whether an autonomous research agent can carry out an end-to-end scientific workflow from an underspecified research instruction and task-visible resources. The benchmark contains 40 tasks, with four tasks in each of ten domains:

• astronomy;
• chemistry;
• earth science;
• energy science;
• information science;
• life science;
• materials science;
• mathematics;
• neuroscience; and
• physics.

Each benchmark task is grounded in a published target paper and provides the agent with a research instruction, related literature, raw data, and a writable workspace. The agent explores the materials, implements and runs analyses, produces figures, and writes report/report.md. The target paper and official benchmark checklist remain hidden from the research agent during execution.

The official evaluator compares the generated report and supporting artifacts against a weighted, task-specific checklist containing textual and image criteria. ResearchClawBench scores range from 0 to 100; under the benchmark protocol, a score around 50 is intended to represent approximately target-paper-level re-discovery.

What is AutoSciRub?

AutoSciRub is an evaluation-first control framework for autonomous research agents. It turns an underspecified instruction into a task-specific executable rubric and then uses that rubric during execution, verification, and revision. Its two main stages are:

1. Automatic Rubric Induction
   • derive an instruction-faithful scientific-goal skeleton;
   • ground the goals in relevant scientific literature;
   • inspect task-visible data and execution constraints; and
   • synthesize executable criteria specifying experiments, comparisons, metrics, evidence artifacts, and satisfaction conditions.
2. Rubric-Guided Iterative Revision
   • verify the current report and supporting artifacts criterion by criterion;
   • diagnose missing experiments, evidence, analyses, or unsupported claims; and
   • perform targeted revisions, with a maximum of three revision rounds and adaptive early stopping.

The + workspace conditions below denote the corresponding agent/model configuration augmented with the AutoSciRub plugin layer. Vanilla and +AutoSciRub members of a pair received the same task inputs, task-visible resources, workspace interface, and available tools. They were not compute-matched because AutoSciRub adds rubric-induction, verification, and revision calls.

Experimental design represented in this archive

The paper evaluates generalization on ResearchClawBench along two controlled axes:

Cross-model generalization

The agent harness is fixed to Codex while the backbone model is varied:

Directory	Harness	Backbone model	Condition	Paper overall score
workspaces-codex/	Codex	GPT-5.4	vanilla	18.66
workspaces-codex+/	Codex	GPT-5.4	+AutoSciRub	21.04
workspaces-codex-glm52/	Codex	GLM-5.2	vanilla	20.86
workspaces-codex+-glm52/	Codex	GLM-5.2	+AutoSciRub	22.73
workspaces-codex-minimax-m3/	Codex	MiniMax-M3	vanilla	19.05
workspaces-codex-minimax-m3-skills+/	Codex	MiniMax-M3	+AutoSciRub	21.04

Cross-harness generalization

The backbone is fixed to DeepSeek-V4-Flash while the surrounding agent harness is varied:

Directory	Harness	Backbone model	Condition	Paper overall score
workspaces-cc/	Claude Code	DeepSeek-V4-Flash	vanilla	16.60
workspaces-cc+/	Claude Code	DeepSeek-V4-Flash	+AutoSciRub	18.74
workspaces-openclaw/	OpenClaw	DeepSeek-V4-Flash	vanilla	17.25
workspaces-openclaw+/	OpenClaw	DeepSeek-V4-Flash	+AutoSciRub	20.36
workspaces-openscience/	OpenScience	DeepSeek-V4-Flash	vanilla	14.49
workspaces-openscience+/	OpenScience	DeepSeek-V4-Flash	+AutoSciRub	18.09

These are the ResearchClawBench main-result conditions from Table 1 of the paper. This archive does not claim to contain the complete AstaBench experiment archive or every analysis and ablation reported elsewhere in the paper.

Additional leaderboard run

The repository also includes a later, separate ResearchClawBench leaderboard run at workspaces-leaderboard0903/. This directory contains one completed run for each of the 40 benchmark tasks. The retained _meta.json records identify the configuration as Codex+AutoSciRub with the GPT-5.6-sol model; all 40 workspaces contain _meta.json, _score.json, and report/report.md. Run timestamps range from 2026-08-30 through 2026-09-01.

On 2026-09-03, this configuration reached 33.2 Pass@1 on the latest ResearchClawBench leaderboard. The workspaces-leaderboard0903/ directory preserves the reports, scores, code, and supporting artifacts associated with that AutoSciRub + Codex + GPT-5.6-Sol leaderboard result.

This additional run postdates the paper's main-result conditions listed above and must not be included when reproducing the Table 1 aggregates. It was packaged from the local reduced-size workspaces-codex-subfinal0902-womodel copy; that local suffix is a packaging label and does not mean that the experiment was executed without a model. Its score records should be interpreted using the same evaluation and aggregation rules described below.

Evaluation and score aggregation

Each task/configuration pair produces one scientific report and one set of supporting artifacts. Every selected submission is independently evaluated three times by the official ResearchClawBench evaluator using GPT-5.1.

For one evaluation attempt, the task score is the weighted sum of checklist-item scores. The saved _score.json records include item types, weights, per-item scores, evaluator reasoning, generated/target image references where applicable, the weighted total_score, and the multi-attempt score_history.

The reported quantities are aggregated as follows:

1. Task score: arithmetic mean of the three independent evaluator attempts.
2. Domain score: arithmetic mean of the four task scores in that domain.
3. Overall score: arithmetic mean of all 40 task scores.
4. Improvement: absolute difference between paired vanilla and +AutoSciRub scores.

The paper's reported aggregate values should be treated as authoritative. Several directories retain retries or additional raw runs:

• workspaces-codex-minimax-m3/ contains 43 run directories and 41 _score.json files, including two scored runs for Earth_001;
• workspaces-codex-minimax-m3-skills+/ contains 42 run directories and 40 scored task records;
• workspaces-openscience/ contains 46 run directories and 40 scored task records; and
• workspaces-openscience+/ contains 43 run directories and 40 scored task records.

Do not naively average every run directory. workspaces-openclaw+/manifest.json explicitly records its one-per-task checkpoint selection. Equivalent selection manifests were not preserved at the root of every other condition, which is a limitation of this release candidate.

Repository structure

.
├── README.md
├── workspaces-codex/
├── workspaces-codex+/
├── workspaces-codex-glm52/
├── workspaces-codex+-glm52/
├── workspaces-codex-minimax-m3/
├── workspaces-codex-minimax-m3-skills+/
├── workspaces-cc/
├── workspaces-cc+/
├── workspaces-openclaw/
├── workspaces-openclaw+/
├── workspaces-openscience/
├── workspaces-openscience+/
└── workspaces-leaderboard0903/  # later 40-task leaderboard run

A typical run directory has the following layout:

<condition>/<TaskID>_<timestamp-or-checkpoint>/
├── INSTRUCTIONS.md        # task instruction supplied to the agent
├── _meta.json             # harness/model, command, status, timing, usage
├── _score.json            # checklist scores and three-attempt history
├── report/
│   ├── report.md          # final scientific report
│   └── images/            # report figures
├── code/                  # agent-authored analysis/experiment code
└── outputs/               # numerical results, models, caches, and other artifacts

Not every run contains every path. AutoSciRub runs can additionally contain induced rubrics, literature-grounding records, experiment plans, verifier outputs such as report_rubric_check.json, and revision metadata. OpenScience runs can contain _openscience_session.json session exports.

Artifact semantics

The following classification should be used when interpreting files:

Class	Paths	Interpretation
Final research artifact	report/report.md, report/images/	The report and figures presented for evaluation.
Evaluation record	_score.json	Official checklist-based judgments and score history.
Run provenance	_meta.json, INSTRUCTIONS.md, manifest.json	Agent/model configuration, task identity, timing, and selection metadata.
Supporting reproducibility artifact	code/, report-referenced files in outputs/	Code and machine-readable evidence used to support reported results.
Mixed/intermediate artifact	other files in outputs/	May include processed inputs, temporary forecasts, model checkpoints, graph caches, optimizer state, and exploratory results. Presence alone does not mean that a file was used in the final report.

The archive intentionally preserves much of outputs/ because the benchmark evaluates evidence beyond prose. However, outputs/ is heterogeneous. Files and directories named cache, graph_cache, processed, tmp, prepared_data, *_states.*, or downloaded model copies should normally be treated as intermediate unless the corresponding report or code explicitly cites them.

Large files

At packaging time, the archive contained 28 regular files larger than 100 MB and eight larger than 1 GB. Examples include:

• trained checkpoints such as understanding_trained.pt;
• processed molecular benchmark tensors such as muv.pt and hiv.pt;
• forecast arrays such as cascade_forecast.npy, clim_forecast.npy, and persist_forecast.npy; and
• serialized graph caches and baseline models.

Forecast arrays directly cited by a report may be primary numerical evidence. Trained weights, processed tensors, caches, and serialized models are generally supporting or intermediate reproducibility artifacts rather than the final human-readable result. Users should trace a large artifact through the run's report and code before treating it as a reported result.

Dates, code version, and dependency provenance

• Experiment run timestamps: 2026-06-09 through 2026-07-17, according to retained _meta.json records.
• Evaluation timestamps: 2026-06-09 through 2026-07-18, according to retained _score.json records.
• Archive packaging date: 2026-09-03.
• Packaging checkout: b76fdeca0164148d5be5283665ba31a67f021061 from XuehaiWang/ResearchClawBench.
• Canonical benchmark: InternScience/ResearchClawBench.

The packaging checkout is recorded for provenance, but the exact Git commit used for every historical run was not consistently written into run metadata. It must not be interpreted as proof that all runs used precisely that commit.

The benchmark repository documents Python 3.10 or newer. At the packaging checkout, the top-level evaluation requirements used version ranges rather than a complete lockfile:

flask>=3.0
flask-cors>=4.0
openai>=1.0
python-dotenv>=1.0
structai                # unpinned
openpyxl                # unpinned

Only three run-local dependency manifests remain in this curated archive, and environment directories were deliberately excluded. Consequently, exact package-level environment reconstruction is not guaranteed. Consult any run-local requirements*.txt, pyproject.toml, code imports, and _meta.json fields before attempting to rerun an artifact.

Packaging and exclusions

The main-experiment release candidate is approximately 24 GB. The additional workspaces-leaderboard0903 leaderboard directory is approximately 1.1 GB. The main archive was curated from a roughly 50 GB working archive by excluding common execution environments, caches, benchmark inputs, and raw agent logs. The main excluded categories are:

• nested Git repositories;
• virtual environments, local package installations, Python bytecode, and common caches;
• top-level downloaded models/ directories;
• task data/ inputs;
• task related_work/ papers; and
• _agent_output.jsonl raw agent-output logs.

The original task data, related literature, hidden target papers, and official checklists should be obtained from the canonical ResearchClawBench distribution subject to its access and licensing terms. Their omission also means this archive is not, by itself, a fully standalone rerun bundle.

Agent traces, model outputs, and sensitive information

This archive does contain model-generated content: reports, code, plots, numerical outputs, induced rubrics, verifier judgments, and evaluator reasoning.

Raw _agent_output.jsonl logs were excluded. Nevertheless, partial trace-like material remains, including:

• 43 _openscience_session.json files containing OpenScience session messages and metadata;
• induced-rubric, grounding, experiment-plan, and verifier JSON files; and
• run commands, session identifiers, token-usage statistics, local paths, and workspace metadata in some _meta.json, _score.json, manifest, and session files.

No credentials are intentionally included, but this release candidate has not been certified as fully anonymized or secret-free. An automated pattern scan still reports possible matches in a small number of files; several appear to be vendored-code or placeholder false positives, but they require manual review before making the repository public. Keep the first Hub upload private until that review and any desired path/session redaction are complete.

The archive does not intentionally include private human conversations or human-subject data. Model-generated scientific claims may be incorrect and must not be treated as expert-validated findings merely because they received a benchmark score.

License and reuse

No single dataset-wide license is asserted for this experiment archive at this time.

The canonical ResearchClawBench code repository and its Hugging Face dataset identify an MIT license, but this archive also contains generated reports, generated code, derived numerical artifacts, model outputs, and possible derivatives of task-provided or third-party materials. The benchmark's MIT license must not be assumed to override upstream dataset, model, paper, or artifact-specific terms.

Before redistribution or commercial reuse, users should:

1. consult the canonical ResearchClawBench license;
2. inspect any file-level license or attribution notices;
3. verify the terms of the original task data and pretrained models; and
4. contact the archive authors when the applicable rights are unclear.

Limitations

• The archive reflects the paper's ResearchClawBench main experiments, not every experiment in the paper.
• Some condition directories retain retries, incomplete runs, or duplicate task runs.
• Only the OpenClaw +AutoSciRub directory includes a root selection manifest.
• Exact historical Git revisions and fully pinned Python environments are unavailable for some runs.
• Raw task data and related-work papers were excluded, so rerunning requires the canonical benchmark resources.
• outputs/ mixes final supporting evidence with intermediate artifacts.
• Generated scientific content and evaluator judgments can contain errors or unsupported claims.
• Local absolute paths and session-level metadata remain in some records.

Citation

If you use these experimental artifacts, please cite the AutoSciRub paper:

@article{wang2026learning,
  title   = {Learning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents},
  author  = {Wang, Xuehai and Qin, Haowei and Liu, Tongxin and Li, Junkai and Xu, Buqiang and Zhang, Jintian and Chen, Yijun and Xue, Zirui and Deng, Shumin},
  journal = {arXiv preprint arXiv:2608.31076},
  year    = {2026}
}

Please also cite the benchmark:

@article{xu2026researchclawbench,
  title   = {ResearchClawBench: A Benchmark for End-to-End Autonomous Scientific Research},
  author  = {Xu, Wanghan and Li, Shuo and Ye, Tianlin and Cao, Qinglong and Chen, Yixin and others},
  journal = {arXiv preprint arXiv:2606.07591},
  year    = {2026}
}

Links

• AutoSciRub paper: https://arxiv.org/abs/2608.31076
• ResearchClawBench paper: https://arxiv.org/abs/2606.07591
• ResearchClawBench code: https://github.com/InternScience/ResearchClawBench
• ResearchClawBench dataset: https://huggingface.co/datasets/InternScience/ResearchClawBench