New PRL-Bench benchmark evaluates how well AI language models can autonomously conduct end-to-end physics research tasks without human intervention.

PRL-Bench is constructed from 100 curated papers from Physical Review Letters since August 2025 and validated by domain experts

The benchmark addresses gaps in current scientific evaluation by testing exploratory capabilities and procedural complexity, not just domain knowledge comprehension

Focuses on theoretical and computational physics as a testbed with verifiable workflows that don't require physical experiments

Designed to map capability boundaries of LLMs in executing real-world research tasks with long-horizon, autonomous exploration

Covers five major theory- and computation-intensive domains to comprehensively assess agentic science capabilities