Evidence Integration in Large Language Models
作者: Sebastien Kawada, Manolis Kellis
分类: cs.CL, cs.AI, cs.LG
发布日期: 2026-09-03
备注: 114 pages, 16 figures, 38 tables
💡 一句话要点
提出证据整合理论以提升大型语言模型的决策能力
🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)
关键词: 大型语言模型 证据整合 推理机制 科学发现 决策理论
📋 核心要点
- 现有大型语言模型在整合外部证据时缺乏清晰的机制，导致决策过程不透明。
- 论文提出了一种分布理论，强调接收者的先验权重和候选证据的影响，形成新的推理框架。
- 实验结果显示，接收者更倾向于整合自身错误，且相同证据对不同模型的影响存在显著差异。
📝 摘要（中文）
尽管大型语言模型（LLMs）越来越依赖外部证据进行推理，但它们如何将这些证据整合到已形成的决策中仍不清楚。本文提出了一种分布理论，认为证据会改变接收者的初始答案分布，基于接收者的先验权重和候选证据的倾斜程度，形成三项预测。通过对十百万次试验的验证，涵盖了来自四个家族的十二个LLMs和八个领域的任务，结果表明，接收者更容易整合自身的特征性错误，而相同的证据对弱模型有提升作用，对强模型则可能产生负面影响。
🔬 方法详解
问题定义：本文旨在解决大型语言模型在整合外部证据时的决策不透明性问题，现有方法未能有效解释证据如何影响模型的初始答案分布。
核心思路：论文提出的分布理论认为，证据通过接收者的先验权重和候选证据的倾斜程度来影响决策，强调接收者的特性在整合过程中的重要性。
技术框架：整体架构包括接收者的初始答案分布、候选证据的整合过程和决策输出，分为证据输入、验证、整合和输出四个主要阶段。
关键创新：最重要的创新在于提出了接收者相对可靠性边界的概念，揭示了接收者特性如何影响错误整合的深度和效果。
关键设计：在实验中，使用了多种参数设置和损失函数，确保模型在处理不同类型的证据时能够保持高效的整合能力。
🖼️ 关键图片
📊 实验亮点
实验结果表明，接收者更倾向于整合自身的特征性错误，且相同证据对弱模型的提升幅度可达93-100%，而对强模型则可能造成负面影响，显示出证据整合的复杂性和模型间的差异。
🎯 应用场景
该研究的潜在应用领域包括科学发现、医疗诊断和智能问答系统等，能够帮助模型更好地利用外部证据进行推理，从而提升决策质量和准确性。未来，随着模型能力的提升，该理论可能会在更广泛的领域中发挥重要作用。
📄 摘要（原文）
Despite increasing reliance on LLMs that reason with external evidence supplied by tools, retrieval-augmented generation, 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, leading to three predictions. First, candidates more probable to the receiver are more persuasive. Second, receivers more readily integrate characteristic errors of their own than foreign errors from different sources. Third, identical evidence can improve weaker models and harm stronger ones. We confirm these over ten million trials, twelve LLMs from four families, and eight domains, four of them scientific discovery tasks in the physical and life sciences: quantum mechanics, physics, genetics, and molecular biology. The law also yields a receiver-relative reliability frontier: receiver-congruent errors depress performance more steeply than random errors of the same rate. LLMs also integrate candidates even after internally verifying their invalidity (93-100% with propositional constraints; up to 99.4% on held-out physical and life-sciences reasoning), demonstrating evidence integration is a receiver-specific control policy over existing distributions, determined by receiver properties rather than scalar trust in the evidence source. Causal interventions show candidate integration is implemented late in the network, as a structured sequence of steps admitting external candidate answers, promoting them, and transporting them into the answer state. Representations of verification are decodable but have little causal impact on answers. A J-lens decomposition shows the state underlying verbalized verification is fully dissociable from that underlying candidate integration.