This study presents a distributional theory on how large language models (LLMs) integrate external evidence into their decision-making process. The theory predicts that more probable candidates are more persuasive, models tend to favor their own characteristic errors, and weaker models can benefit from evidence while stronger models may be harmed.
The research confirms these predictions through over ten million trials involving twelve LLMs across four domains, including scientific discovery tasks in physics, genetics, and molecular biology. Results show that evidence integration is a receiver-specific control policy, influenced by model properties rather than trust in the evidence source.
Additionally, the study finds that LLMs can integrate evidence even after verifying its invalidity, with high success rates. Candidate integration occurs late in the network as a structured sequence of steps, with representations of verification being decodable but having little causal impact on answers. The findings highlight the importance of understanding evidence integration mechanisms for model reliability and behavior.
Source: https://arxiv.org/abs/2609.04290