CARE: Controlling LLM-Generated Policies through Auditable Review of Evidence in Scientific Experimentation

2026-06-12

AI Summary

This work addresses the challenge of safely harnessing large language models (LLMs) in high-throughput experimental optimization, where direct LLM use risks unsafe exploration yet complete exclusion forfeits their optimization potential. To reconcile this trade-off, the authors propose the CARE framework, which employs a non-LLM default optimizer as the primary pathway while leveraging the LLM to generate candidate strategies. Adoption of these candidates is governed by an evidence-based intervention gating mechanism that audits proposals against publicly available evidence, ensuring decisions are auditable, controllable, and traceable. By synergistically integrating LLM-driven creativity with evidence-guided safety constraints, CARE achieves state-of-the-art performance on the Minerva/Olympus and ChemLex benchmarks, improving peak scores from 80.0 to 88.5 and from 83.9 to 92.1, respectively.

Abstract

Granting LLMs direct control over costly, irreversible scientific experiments leads to unsafe exploration and unstable performance, but discarding LLM creativity entirely sacrifices significant optimization potential. We introduce CARE (Controlling LLM-Generated Policies through Auditable Review of Evidence in Scientific Experimentation), an auditable controller for high-throughput experimentation (HTE) optimization that keeps a non-LLM incumbent optimizer as the default action path while using LLMs to revise challenger ranking policies. Before each outcome is revealed, a public-evidence intervention gate compares the challenger with the incumbent. It authorizes the challenger's selection only when the evidence available before selection supports the change, with the decision recorded in the audit log. CARE outperforms all other evaluated methods on Minerva/Olympus and ChemLex benchmarks, with final-best improving from 80.0 to 88.5 on Minerva/Olympus and from 83.9 to 92.1 on ChemLex, relative to the public incumbent. Our experiments indicate that LLM self-evolution is more reliable when it expands the proposal space under an auditable controller, rather than directly choosing experiments.

Problem

• LLM control
• scientific experimentation
• unsafe exploration
• high-throughput experimentation
• auditable decision-making

Innovation

• auditable control
• LLM-guided optimization
• evidence-based intervention
• high-throughput experimentation
• challenger-incumbent framework