大语言模型中的证据整合机制研究

AI 摘要

该研究提出了一种关于大语言模型（LLM）如何整合外部证据的分布理论，认为证据会通过接收者先验权重和候选证据倾斜来改变模型的初始答案分布。基于超过一千万次试验、十二个模型和八个领域的验证，研究发现模型更易接受自身倾向的候选，且相同证据可能提升弱模型而损害强模型。因果干预表明候选整合发生在网络后期，而验证表征对最终答案几乎没有因果影响。

正文节选

Evidence Integration in Large Language Models Abstract Despite increasing reliance on LLMs that reason with external evidence supplied by tools, retrieval-augmented generation (RAG) systems, other agents, and users, how LLMs integrate such evidence into decisions they have already begun to form remains largely unclear. We present a distributional theory in which evidence shifts the receiver’s distribution of initial answers, driven by a receiver prior weight and a candidate evidence tilt, leadin

发布时间：2026-09-07 12:00

抓取时间：2026-09-07 12:06

来源机构：arXiv