Evidence Integration in Large Language Models

大语言模型中的证据整合

Abstract: 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.

摘要： 尽管人们越来越依赖大语言模型（LLM）利用工具、检索增强生成（RAG）、其他智能体及用户提供的外部证据进行推理，但大语言模型如何将这些证据整合到其已经开始形成的决策中，目前仍不明确。

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.

我们提出了一种分布理论，在该理论中，证据会改变接收者（模型）对初始答案的分布，这种改变由接收者的先验权重和候选证据的倾向性驱动，并由此得出三个预测。第一，接收者认为概率更高的候选答案更具说服力。第二，接收者更容易整合自身特有的错误，而非来自不同来源的外部错误。第三，相同的证据可能会提升较弱模型的表现，却损害较强模型的表现。

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.

我们通过一千万次试验、来自四个系列的十二个大语言模型以及八个领域验证了这些结论，其中四个领域涉及物理和生命科学的科学发现任务：量子力学、物理学、遗传学和分子生物学。该定律还揭示了一个“接收者相对可靠性边界”：与接收者自身逻辑一致的错误比同等比例的随机错误对性能的抑制作用更为显著。

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.

即使在内部验证出候选答案无效后，大语言模型仍会将其整合（在命题约束下整合率为 93-100%；在物理和生命科学推理的留出测试集中高达 99.4%）。这表明证据整合是一种针对现有分布的、接收者特定的控制策略，其结果由接收者的属性决定，而非取决于对证据来源的标量信任度。

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.

因果干预研究表明，候选答案的整合是在网络后期实现的，表现为一个结构化的步骤序列：接纳外部候选答案、提升其权重，并将其传输至答案状态。验证过程的表征虽然可被解码，但对最终答案几乎没有因果影响。J-lens 分解显示，语言化验证背后的状态与候选答案整合背后的状态是完全可分离的。