LLM Agents Pass Numerical Tests But Misidentify Physics Models

In a new paper accepted to the AID-Wild workshop at CAIS 2026, researchers Wanyu Zhao and Wanbing Zhao probe whether LLM-based AI agents can truly understand theoretical physics at a structural level. They introduce StatMechBench-v0, a benchmark composed of six Ising-type problems that require skills like transfer-matrix methods, gauge-removable disorder, and planar/Pfaffian structure identification. The agents operate in a propose-verify-revise loop across multiple large language models, attempting to transform raw partition functions into tractable representations.

The results are a cautionary tale for AI-assisted science. While numerical feedback (computed partition function values) often helps agents fix bugs and recover correct outputs, the agents can still pass all numerical checks while misidentifying the true mathematical structure—e.g., labeling a problem as exactly solvable when it is not, or underestimating its computational complexity. This reveals a critical gap: current LLMs lack the deep reasoning to recognize known model classes beyond surface-level numerical agreement. The authors argue that future verification stacks must go beyond numbers to include symbolic checks and structural invariants, especially for AI agents aimed at discovery in theoretical physics.