NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents
What this work does
Mutates canonical physical-law expressions and embeds them in increasingly complex interactive model systems to test identification beyond familiar formulas.
Connection to discovery
A control against memorization and a way to evaluate active experimental design under known hidden equations.
Evidence
Evaluations separate symbolic accuracy, predictive fidelity and token cost; the noise analysis shows that these metrics can diverge.
Scope and limitations
Equation mutations and simulator construction define the hypothesis family. Counterfactual physics is intentionally virtual, and numerical fit is not equivalent to correct symbolic recovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.