LLM-SR: Scientific Equation Discovery via Programming with Large Language Models
What this work does
A frozen LLM proposes executable equation skeletons; numerical optimization fits constants, and scored programs are retained in a clustered experience buffer for iterative refinement.
Connection to discovery
A useful frozen-policy baseline against which to measure whether RL actually improves equation search efficiency or generalization.
Evidence
Inspected the fitness definition, program sampling rules, in/out-of-domain error comparisons and ablations removing scientific priors and program representations.
Scope and limitations
Descriptions supply substantial scientific priors; low prediction error does not establish a unique mechanism or causal identifiability. The evaluator and available operators delimit discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.