Agentic large language models are increasingly used to support scientific prediction by grounding their decisions in external evidence. In biology, such models promise accurate and interpretable predictions for important problems by integrating complex experimental data. But verifying what the model actually relied on remains a challenge.
The arXiv paper "Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing" addresses this gap. It describes two complementary approaches: source localization, which aims to pinpoint which external sources or tool traces contributed to a prediction, and counterfactual dependence testing, which probes whether the prediction would change if those sources were different.
Together, these methods are intended to make agentic LLM reasoning more auditable. The abstract emphasizes interpretability as a key goal, suggesting that the work is aimed at users who need to trust model predictions in high-stakes scientific settings. Because the abstract is brief, the specific techniques and results are not detailed here, but the stated direction is clear: move from black-box outputs to traceable, testable evidence use.