文章背景与核心概要
随着大语言模型（LLM）越来越多地依赖来自工具、检索增强生成（RAG）、其他智能体以及用户的外部证据来进行推理，LLM 如何将这些外部证据整合到其已经开始形成的决策中，在很大程度上仍然不够清晰。本文提出了一种分布理论，认为证据会根据“接收者先验权重”和“候选证据倾斜度”来改变接收者对初始答案的分布，并通过涵盖超 1000 万次试验、12 个大模型和 8 个领域（包括量子力学、物理学、遗传学和分子生物学）的实验进行了验证。
研究发现，尽管大模型内部能够验证错误或无效的候选答案并不准确，但它们往往仍然会将其整合进最终结果中。因果干预进一步证明，这种整合发生在神经网络的深层阶段，是由接收者特定的控制策略驱动的，而不是对来源的简单标量信任。该研究为理解大模型在复杂推理任务中的外部信息处理方式提供了重要的理论与实验洞察。
Evidence Integration in Large Language Models
Summary
This paper investigates how Large Language Models (LLMs) integrate external evidence (from tools, retrieval-augmented generation, agents, or users) into ongoing reasoning processes. The authors propose a distributional theory where evidence shifts the receiver's initial answer distribution based on a "receiver prior weight" and a "candidate evidence tilt." Validated across over 10 million trials, 12 LLMs, and 8 domains (including quantum mechanics, physics, genetics, and molecular biology), the findings reveal that LLMs often integrate erroneous or invalid candidate answers despite internally verifying their inaccuracy. Causal interventions further demonstrate that this integration occurs late in the network through a structured, receiver-specific control policy rather than simple scalar trust in the source.
本文研究了大语言模型（LLM）如何将外部证据（来自工具、检索增强生成、智能体或用户）整合到正在进行的推理过程中。作者提出了一种分布理论：证据基于“接收者先验权重（receiver prior weight）”和“候选证据倾斜度（candidate evidence tilt）”来改变接收者对初始答案的分布。经过对超过 1000 万次试验、12 个 LLM 以及 8 个领域（包括量子力学、物理学、遗传学和分子生物学）的验证，研究结果表明，尽管 LLM 在内部能够验证候选答案的不准确性，但它们往往仍会整合错误或无效的候选答案。因果干预进一步表明，这种整合发生在网络后期的结构化、接收者特定的控制政策中，而不是源于对来源的简单标量信任。
Metadata
- arXiv ID: arXiv:2609.04290
- Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) - Authors: Sebastien Kawada, Manolis Kellis
- Submission Date: September 3, 2026
- Comments: 114 pages, 16 figures, 38 tables
元数据
- arXiv ID: arXiv:2609.04290
- 研究主题: 计算与语言 (
cs.CL); 人工智能 (cs.AI); 机器学习 (cs.LG)- 作者: Sebastien Kawada, Manolis Kellis
- 提交日期: 2026年9月3日
- 评论: 114页，16幅图表，38个表格
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. 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.
- Candidates more probable to the receiver are more persuasive.
- Receivers more readily integrate characteristic errors of their own than foreign errors from different sources.
- 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.
摘要
尽管人们越来越依赖利用工具、检索增强生成（RAG）、其他智能体和用户提供的外部证据进行推理的 LLM，但 LLM 如何将此类证据整合到它们已经开始形成的决策中，在很大程度上仍然不明确。我们提出了一种分布理论，其中证据由接收者先验权重和候选证据倾斜度驱动，从而改变接收者对初始答案的分布，该理论导出了三个预测结果。
- 对接收者而言概率更高的候选答案更具说服力。
- 接收者比来自不同来源的外来错误更容易整合其自身的特征错误。
- 相同的证据可以改进较弱的模型，同时损害较强的模型。
我们通过 1000 多万次试验、来自 4 个家族的 12 个 LLM 以及 8 个领域（其中 4 个是物理学和生命科学领域的科学发现任务：量子力学、物理学、遗传学和分子生物学）证实了这些预测。该规律还产生了一个“接收者相对可靠性边界”（receiver-relative reliability frontier）：与接收者相符的错误比相同比率的随机错误对性能的压制更为严重。
此外，LLM 甚至在内部验证了候选答案无效之后，仍然会将其整合（在命题约束下为 93-100%；在留出的物理和生命科学推理中高达 99.4%），这表明证据整合是针对现有分布的、接收者特定的控制策略，由接收者自身的属性而非对证据来源的标量信任所决定。因果干预表明，候选答案的整合是在网络后期实现的，表现为接纳外部候选答案、提升它们并将其传输到答案状态的结构化步骤序列。验证的表征是可解码的，但对答案几乎没有因果影响。J-lens 分解表明，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解耦的。
Links & Resources
链接与资源