TSBench: A physics-grounded benchmark for evaluating LLM understanding of chemical reaction mechanisms

RESUMO

Understanding a chemical reaction requires mapping a symbolic reactant-product description onto the three-dimensional pathway through which atoms rearrange, yet chemistry benchmarks for large language models (LLMs) largely probe factual knowledge and text-based reasoning. Here we introduce TSBench, a benchmark in which an LLM agent uses structure-editing tools to construct three-dimensional transition-state (TS) guesses verified by an automated quantum-chemical pipeline, yielding a physics-grounded pass/fail verdict. Across 546 evaluations of seven frontier LLMs on 78 elementary reactions, the aggregate success rate rose from 50.4% to 66.8% under diagnosis-driven revision; the best models approached 90% on the simplest reactions, yet performance dropped sharply with mechanistic complexity. Most failed attempts produced locally plausible saddle points whose reaction paths led to the wrong reactant-product pair, revealing local geometric intuition without a robust grasp of the global reaction coordinate. TSBench establishes a mechanism-level yardstick for LLM agents in mechanism-sensitive tasks such as synthesis planning and autonomous experimentation.