ARC-Bench: 55 Open-Ended Research Questions Across Five Scientific Domains

ARC-Bench is a 55-topic benchmark for evaluating autonomous research agents. Each topic presents an open-ended research question with a structured briefing, spanning five distinct scientific domains. The dataset was created by AIMING-Lab-UNC and released in May 2026.

Use Cases

• Benchmarking autonomous research agents based on their ability to address open-ended scientific questions.
• Evaluating human-AI collaboration workflows based on structured research briefings.
• Training AI systems on multi-domain scientific reasoning based on the five covered domains.

Strengths

• Contains 55 distinct research topics for evaluation.
• Covers five different scientific domains, suggesting breadth.
• Provides structured briefings for each open-ended question.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Sample data and file formats are unavailable for preview.

Provenance

Source

AIMING-Lab-UNC

Collection Method

Released as part of the AutoResearchClaw project for autonomous research.

Freshness

Last updated 2026-05-22 03:33:50; freshness should be verified.