Researchers Unveil How Large Language Models Integrate External Evidence

A new research preprint published on arXiv examines how large language models (LLMs) integrate external evidence from tools, agents, and retrieval-augmented systems into their decision-making processes. The authors proposed a distributional framework showing that models are more persuaded by candidate answers that match their own initial probability distributions and characteristic errors.

The framework was evaluated across more than ten million trials involving twelve LLMs from four families across eight domains, including quantum mechanics, physics, genetics, and molecular biology. The findings show that identical evidence can improve the performance of weaker models while degrading the output of stronger ones.

Additionally, researchers discovered that models frequently integrate external candidate answers even after internally verifying their invalidity (reaching 93% to 100% under propositional constraints). Causal interventions revealed that candidate integration occurs late in the network architecture and operates largely separately from internal verification representations.