PRL-Bench: New AI Benchmark Tests LLMs on Physics Research Tasks

A new benchmark called PRL-Bench evaluates large language models' abilities to conduct frontier physics research autonomously. Developed from 100 curated papers published in Physical Review Letters since August 2025, the benchmark assesses AI systems on theory- and computation-intensive tasks. Domain experts validated the benchmark, which focuses on theoretical and computational physics as a testbed requiring comprehensive domain knowledge and complex reasoning. PRL-Bench systematically maps LLM capabilities in executing end-to-end physics research workflows without experimental reliance. The benchmark addresses limitations of current scientific evaluations that fail to assess exploratory nature and procedural complexity. It aims to advance agentic science paradigms where AI systems engage in long-horizon autonomous exploration. The work introduces research-oriented evaluations that move beyond domain knowledge comprehension toward verifiable end-to-end workflows.