Add files using upload-large-folder tool

667c0a2 verified 10 days ago

11.1 kB

ResearchBench

ResearchBench is the official code release for the ACL Findings 2026 paper ResearchBench: Benchmarking LLMs in Scientific Discovery via Inspiration-Based Task Decomposition.

This GitHub repository contains the evaluation toolkit, command-line interface, prompts, metrics, tests, and a 12-sample tiny smoke-test split. The full benchmark dataset is released on Hugging Face: ankilok/ResearchBench.

ResearchBench evaluates large language models on scientific discovery through three inspiration-based subtasks:

Inspiration retrieval: select the papers that can inspire a target research hypothesis.
Hypothesis composition: compose a research hypothesis from a research question and gold inspirations.
Hypothesis ranking: choose the better hypothesis in pairwise comparisons between a gold hypothesis and negative hypotheses.

The full release contains 1,369 paper records, 1,367 retrieval tasks, 1,367 generation tasks, and 1,323 ranking tasks.

🚀 Quick Start

Clone the code repository and run all three subtasks on the bundled data/tiny/*.jsonl smoke-test split with an OpenAI-compatible provider:

git clone https://github.com/ankitala/ResearchBench.git
cd ResearchBench

python -m pip install -e ".[openai]"

export OPENAI_API_KEY="<your-api-key>"
export OPENAI_BASE_URL="<your-base-url>"
MODEL="<model-name>"

researchbench run-retrieve --data data/tiny/retrieve.jsonl --model-name "$MODEL" --concurrency 15 --out outputs/retrieve_tiny.jsonl
researchbench score-retrieve --pred outputs/retrieve_tiny.jsonl --data data/tiny/retrieve.jsonl

researchbench run-generate --data data/tiny/generation.jsonl --model-name "$MODEL" --num-mutations 2 --num-itr-self-refine 1 --concurrency 15 --out outputs/generation_tiny.jsonl
researchbench score-generate --pred outputs/generation_tiny.jsonl --data data/tiny/generation.jsonl --judge-model "$MODEL" --concurrency 15

researchbench run-rank --data data/tiny/ranking.jsonl --model-name "$MODEL" --order both --concurrency 15 --out outputs/ranking_tiny.jsonl
researchbench score-rank --pred outputs/ranking_tiny.jsonl

📊 Benchmark Results

The figures below summarize model performance on the three ResearchBench subtasks. Higher scores indicate better performance.

🔍 Inspiration Retrieval

💡 Hypothesis Composition

🏆 Hypothesis Ranking

⚙️ Installation

ResearchBench requires Python 3.10 or later.

Core local tooling:

python -m pip install -e .

Development and tests:

python -m pip install -e ".[dev]"

OpenAI-compatible provider calls:

python -m pip install -e ".[openai]"
export OPENAI_API_KEY="<provider-api-key>"
export OPENAI_BASE_URL="<provider-base-url>"

You can also pass provider settings per command:

researchbench run-retrieve \
  --data data/tiny/retrieve.jsonl \
  --model-name "<provider-model-name>" \
  --api-key "<provider-api-key>" \
  --base-url "<provider-base-url>" \
  --concurrency 15 \
  --out outputs/retrieve_provider.jsonl

pyproject.toml is the authoritative package configuration. requirements.txt is a convenience file for installing the editable package with development and OpenAI-compatible provider extras.

🗂️ Data Layout

This GitHub repository is code-first. It tracks only the small tiny JSONL files needed for smoke tests:

data/tiny/papers.jsonl: 12-sample paper records.
data/tiny/retrieve.jsonl: 12-sample inspiration retrieval tasks.
data/tiny/generation.jsonl: 12-sample hypothesis composition tasks.
data/tiny/ranking.jsonl: 12-sample hypothesis ranking tasks.

The full benchmark files are hosted on Hugging Face and are not uploaded to GitHub:

papers/papers.jsonl: standardized paper records.
retrieve/retrieve.jsonl: inspiration retrieval tasks.
generation/generation.jsonl: hypothesis composition tasks.
ranking/ranking.jsonl: hypothesis ranking tasks.
validation/*.csv and validation/*.json: build and validation reports.

Download any subset as needed:

huggingface-cli download ankilok/ResearchBench \
  --repo-type dataset \
  --include "retrieve/*.jsonl" \
  --include "generation/*.jsonl" \
  --include "ranking/*.jsonl" \
  --local-dir data

Validate a local data directory:

researchbench validate data

Validation checks candidate counts, gold-label coverage, empty fields, local absolute paths, API key leakage, and task-specific shape constraints.

🛠️ Basic Usage

All run-* commands support --concurrency, defaulting to 15. run-generate also supports --inner-concurrency for independent within-sample branches. Use lower values if your provider has strict rate limits.

Before running the full benchmark, download the corresponding JSONL files from Hugging Face into data/ as shown in the Data Layout section.

🔍 Inspiration Retrieval

researchbench run-retrieve \
  --data data/retrieve/retrieve.jsonl \
  --model-name "<model>" \
  --concurrency 15 \
  --out outputs/retrieve.jsonl

researchbench score-retrieve \
  --pred outputs/retrieve.jsonl \
  --data data/retrieve/retrieve.jsonl

Default settings match the original experiment: window size 15, keep size 3, two rounds, no background survey, no similarity-only selection. Round 1 keeps 15 of 75 candidates; round 2 keeps 3 of 75.

💡 Hypothesis Composition

researchbench run-generate \
  --data data/generation/generation.jsonl \
  --model-name "<model>" \
  --num-mutations 2 \
  --num-itr-self-refine 2 \
  --max-inspiration-search-steps 3 \
  --concurrency 15 \
  --inner-concurrency 1 \
  --out outputs/generation.jsonl

researchbench score-generate \
  --pred outputs/generation.jsonl \
  --data data/generation/generation.jsonl \
  --judge-model "<judge-model>" \
  --concurrency 15

Generation uses main_hypothesis as the gold hypothesis. fine_grained_hypothesis is retained only as optional metadata. The default generation run uses gold inspirations, two mutation lines, two refinement iterations, same-inspiration recombination, cross-inspiration recombination up to three inspiration steps, inter-EA screening window size 12, keep size 3, and recombination beam size 15.

🏆 Hypothesis Ranking

researchbench run-rank \
  --data data/ranking/ranking.jsonl \
  --model-name "<model>" \
  --order both \
  --concurrency 15 \
  --out outputs/ranking.jsonl

researchbench score-rank \
  --pred outputs/ranking.jsonl

--order both evaluates both comparison directions. You may also run --order gold-first or --order negative-first。

📋 Example CLI Output

Retrieve scoring prints a readable summary:

Retrieve score summary
- Round 1: kept 15 of 75 candidates
  - Gold inspiration hit ratio: 0.7419 (74.19%)
  - Negative tier 1 selection ratio: 0.5463 (54.63%)
  - Negative tier 2 selection ratio: 0.0833 (8.33%)
  - Negative tier 3 selection ratio: 0.0189 (1.89%)
- Round 2: kept 3 of 75 candidates
  - Gold inspiration hit ratio: 0.3548 (35.48%)

Generation scoring:

Hypothesis generation score summary
- Scored samples: 12
- Average matched score (0-5): 4.1667
- Generation accuracy (average score / 5): 0.8333 (83.33%)

Ranking scoring:

Hypothesis ranking score summary
- Overall directional accuracy: 0.3750 (37.50%)
- Gold-first order accuracy (gold is candidate 1): 0.3333 (33.33%)
- Negative-first order accuracy (gold is candidate 2): 0.4167 (41.67%)
- Sample-level two-order outcomes
  - Gold wins in neither order: 0.5000 (50.00%); count=6
  - Gold wins in exactly one order: 0.2500 (25.00%); count=3
  - Gold wins in both orders: 0.2500 (25.00%); count=3

The full machine-readable metric objects are written next to prediction files as *.score.json unless --out is specified.

📊 Metrics

Retrieve gold hit ratio: selected gold inspirations divided by total gold inspirations. Reported for round 1 and round 2.
Retrieve negative selection ratio: selected negative candidates in a distance tier divided by total candidates in that tier. This supports the negative-distance analysis.
Generation matched score: judge score from 0 to 5 measuring how well the generated hypothesis covers the gold hypothesis key points.
Generation accuracy: average matched score divided by 5.
Ranking directional accuracy: whether the gold hypothesis wins under a single candidate order.
Ranking sample two-order outcomes: whether gold wins in neither, exactly one, or both of the two reversed orders.
Ranking pair-level order consistency: whether both orders choose gold, both choose negative, or the two orders disagree.

📝 Notes

Full benchmark JSONL files are released through the Hugging Face dataset repository, not through this GitHub repository.
Retrieve candidates are read from merge.json; d0.json, d1.json, d2.json, and d3.json label candidates as gold, negative tier 1, negative tier 2, and negative tier 3.
Keep API keys in environment variables or pass them through CLI arguments.

⚖️ License

Code is released under the MIT License. The dataset is released on Hugging Face under CC BY-NC 4.0 for non-commercial research use; see DATA_LICENSE.md.

📑 Citation

If you use ResearchBench, please cite:

@misc{liu2026researchbenchbenchmarkingllmsscientific,
      title={ResearchBench: Benchmarking LLMs in Scientific Discovery via Inspiration-Based Task Decomposition}, 
      author={Yujie Liu and Zonglin Yang and Tong Xie and Jinjie Ni and Ben Gao and Yuqiang Li and Shixiang Tang and Wanli Ouyang and Erik Cambria and Dongzhan Zhou},
      year={2026},
      eprint={2503.21248},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.21248}, 
}