Exploring Structures in Physics Problems: Can AI Agents Discover Statistical Mechanical Mappings?

Short summary

This study introduces StatMechBench-v0, a benchmark of six Ising-type problems testing whether LLM-based agents can discover statistical mechanical mappings from raw partition functions. A propose-verify-revise agent shows that numerical feedback helps repair code but agents often misidentify the underlying tractable class while passing numerical checks. The findings reveal limitations in LLM reasoning and call for verification stacks incorporating symbolic checks and structural invariants beyond numerical agreement.

• •StatMechBench-v0 benchmarks LLM agents on discovering statistical mechanical mappings in physics
• •Agents pass numerical checks while misidentifying tractable structure, revealing reasoning gaps
• •Calls for verification stacks combining symbolic checks and structural invariants

Generated with AI, which can make mistakes.