LLM Evidence Integration Theory and Empirical Validation

A new distributional theory explains how LLMs integrate external evidence, finding that identical evidence can improve weaker models while harming stronger ones. Experiments across twelve LLMs and eight domains, including quantum mechanics and genetics, confirm that receivers are more susceptible to evidence that aligns with their prior probability.