# Explore, Then Commit: Measurement-Efficient Scientific Law Discovery with Language Models

arXiv: https://arxiv.org/abs/2610.07620 · HTML: https://mlsanity.com/paper/2610.07620/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 9 of its 18 references that are in the map.
- First posted on arXiv 6 Oct 2026.

## Builds on (most cited first, 9 of 9)

- [Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks](https://mlsanity.com/paper/2211.12588/) (2022)
- [PAL: Program-aided Language Models](https://mlsanity.com/paper/2211.10435/) (2022)
- [Contemporary Symbolic Regression Methods and their Relative Performance](https://mlsanity.com/paper/2107.14351/) (2021)
- [LLM-SR: Scientific Equation Discovery via Programming with Large Language Models](https://mlsanity.com/paper/2404.18400/) (2024)
- [DISCOVERYWORLD: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents](https://mlsanity.com/paper/2406.06769/) (2024)
- [Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl](https://mlsanity.com/paper/2305.01582/) (2023)
- [NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents](https://mlsanity.com/paper/2510.07172/) (2025)
- [BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design](https://mlsanity.com/paper/2508.21184/) (2025)
- [BoxingGym: Benchmarking Progress in Automated Experimental Design and Model Discovery](https://mlsanity.com/paper/2501.01540/) (2025)

## Abstract (arXiv)

Scientific law discovery requires selecting measurements and converting evidence into a governing equation. We evaluate an explore-then-commit protocol in which a large language model proposes hypotheses, a programmatic planner gathers measurements, and a fresh prompt synthesizes the final law from fixed observations. The protocol combines structured probes, automatic numerical diagnostics, restricted measurement batches, and optional interpreter access. Across 576 NewtonBench trials, we compare eight configurations on 12 physics modules using GPT-4.1-mini and a medium-difficulty GPT-4.1 replication. On medium tasks, interpreter-enabled planners use 8.6 versus 22.5 measurements per trial for GPT-4.1-mini and 8.9 versus 43.0 for GPT-4.1. Their mean magnitude-based root-mean-squared logarithmic error falls from 2.514 to 0.202 and from 0.626 to 0.149, respectively. An additional audit retains incomplete and invalid submissions in a coverage-sensitive analysis. Observed symbolic-accuracy gains are less consistent across modules, and random acquisition is competitive with disagreement scoring. Measurement savings occur in every module, but unequal batch constraints prevent attributing them solely to acquisition quality. These results support the complete protocol as a promising measurement-efficient configuration, while leaving its causal components and generalization beyond noiseless direct-equation tasks unresolved.