LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models
What this work does
Combines transformed known equations and synthetic scientific terms to reduce simple formula recitation, evaluating both data fidelity and symbolic agreement.
Connection to discovery
Separates memorized scientific priors from fitting less familiar mechanisms and motivates hidden mechanism families in discovery evaluation.
Evidence
Inspected synthetic-task generation/novelty checks, numeric versus symbolic scoring, and differing ID/OOD performance.
Scope and limitations
LLMs participate in synthetic construction and some symbolic judgments. Transformed equations are not entirely unfamiliar science, and symbolic equivalence can differ from numerical agreement.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.