ResearchClawBench: Benchmark for Automated Scientific Research Agents

ResearchClawBench is a benchmark that measures whether AI coding agents can independently conduct scientific research from reading raw data to producing publication-quality reports. The benchmark, created by InternScience, was last updated on May 19, 2026. It evaluates agent outputs against real human-authored papers.

Use Cases

• Benchmarking AI agents' ability to read raw data and produce reports based on the described research workflow.
• Evaluating the quality of AI-generated scientific reports against human-authored papers as described.
• Training AI agents for tasks spanning from literature re-discovery to new discovery as mentioned in the description.

Strengths

• Benchmark is designed to evaluate a complete research workflow from data reading to report generation.
• Evaluation is performed against real human-authored papers, providing a concrete ground truth.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

InternScience on Hugging Face.

Collection Method

Likely compiled for benchmarking AI research agents; specific collection method is not detailed.

Freshness

Last updated 2026-05-19 11:59:57; freshness should be verified.