Machine Heart

Jun 19, 2026 · Artificial Intelligence

LLMs Finally Derive Formulas: FunctionEvolve Boosts LLM‑SRBench Accuracy 3.6× and Scores Perfect on AI‑Feynman

The FunctionEvolve framework represents formulas as abstract syntax trees, letting large language models guide symbolic regression; this yields a 3.6‑fold improvement on the LLM‑SRBench benchmark (55.8% SA@1) and a perfect 120/120 score on AI‑Feynman, with detailed component ablations confirming the value of structure‑aware generation, selection, mutation, and optimization.