Evidence Integration in Large Language Models
Sebastien Kawada, Manolis Kellis
cs.CL · cs.AI · cs.LG
Quick verdict
声明 LLM 证据整合遵循接收者相对法则，在 12 个模型上通过超过 1000 万次试验进行了验证；结果严谨，尽管合成任务限制了部署。
Summary
该论文提出了一种分布理论，研究大型语言模型如何将外部证据——工具输出、检索到的文档、其他代理、用户陈述——整合到已经开始形成的决策中。证据通过接收者先验权重 (1+a) 和候选特定倾斜度对模型预先存在的答案分布进行重新加权；贝叶斯接收者对应于 a=0。由此得出三个预测：接收者已经偏好的候选者更具说服力；接收者比外部错误更容易吸收自己的特征错误；相同的证据对弱接收者有帮助，对强接收者有害。测试跨越了 1000 多万次试验、来自四个家族的 12 个指令调优模型，以及包括量子力学、物理学、遗传学和分子生物学在内的 8 个领域。在算术工具上测得的先验权重在 0.20 到 0.65 之间；在所有九个算术模型中，配对吸引子效应均为正，在 Gemma-4-31B 中高达 +0.590；模型在 93%–100% 的情况下采用提供的候选者，即使是在刚刚判定其无效之后，在保留的科学推理中这一比例达到 99.4%。从机制上讲，集成在网络深处起作用；一个可解码的验证方向可以引导 Gemma，但不能引导 Llama，且 J-lens 可言语化的工作空间不包含任何仲裁状态。主要局限性：所有任务都是具有精确预言机的合成任务，机制覆盖范围比行为测试范围窄，且在部署系统中的普遍性仍未得到测量。
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Evidence integration follows a receiver-relative law p(y|e) ∝ q(y)^(1+a)·exp(φe(y)); measured prior weights 1+a lie between 0.20 and 0.65 on the arithmetic instrument, all below the rational value of one.
- Models adopt externally supplied candidates at 93–100% in propositional-constraint tasks and up to 99.4% in held-out scientific reasoning even immediately after verifying them invalid, dissociating verification from decision control.
- Receiver-congruent errors depress performance more steeply than random errors of the same rate (Proposition 3.5), and the resulting competence-dependent reliability frontier implies no source-only deference rule is optimal.
- Causally, candidate integration is implemented late in the network (handoff at 79–81% of depth); verification representations are decodable but causally inert in Llama while actively steering Gemma, and the J-lens verbalizable workspace holds none of the arbitration state.
- Arbitration weights are properties of model–domain pairs: model–domain interaction explains 49–80% of between-cell variance, and parameter count fixes neither the model effect nor the interaction.
Strengths
Unusually disciplined identification for LLM behavior work
Candidate support A(v) is the exact no-reference sequence log-probability measured before any candidate is shown (Eq. 24); attractors and competence are cross-fitted over disjoint halves of 48 temperature-0.7 samples (Section 4.2); matched quadruples cross certificate status with support under a 0.5-nat caliper (Section 4.3), and Appendix F derives the direction of measurement-error bias in the recruitment contrast (Props. F.2–F.4).
Theory-behavior closure, not just curve fitting
The frontier predicted from the adoption law fitted on other error classes agrees with the directly estimated frontier within 0.06 in six models and the harm crossing within 0.04 in five (Section 5.3); composition predictions (sub-additivity slope −0.39 to −0.71 vs predicted a of −0.47 to −0.91, Section 5.6) are derived before being tested.
Behavioral–mechanistic triangulation
Held-out scientific validation with matched candidate pools
Confound auditing is explicit
Weaknesses
All tasks are synthetic with exact oracles
Mechanistic coverage narrower than the behavioral fleet
Thin headline cells and estimator disagreement
Ceiling compression and uneven fleet reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.04290, MIT CSAIL) asks how LLMs integrate external evidence — tool outputs, retrieval, other agents, user remarks — into decisions they have already begun to form. It contributes (i) a distributional theory in which evidence reweights the receiver's pre-existing answer distribution q via a prior weight 1+a and a candidate-local tilt (Eqs. 4–6), with a Bayesian receiver as the a=0 benchmark and scalar source trust recovered only under uniform error kernels (Section 3); (ii) an empirical law measured over ten million-plus trials on twelve instruction-tuned models from four families across eight domains, four of them scientific; and (iii) causal-mechanistic analyses (patching, attention knockout, steering, J-lens). Central claims: measured prior weights 0.20–0.65, all below the rational value of one; adoption rises with prior support (paired attractor effect positive in all nine arithmetic models, up to +0.590 in Gemma-4-31B); correlated errors are disproportionately dangerous (Proposition 3.5); models adopt candidates at 93–100% right after verifying them invalid; and verification representations are decodable yet causally inert in Llama while steering-active in Gemma.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is evidence integration in LLMs according to this paper?
It is the transformation of a model's already-formed answer distribution when external evidence (a tool output, a retrieved document, another agent, or a user statement) proposes a candidate answer. The paper models it as reweighting the model's own distribution by a prior weight (1+a) plus a tilt toward the proposed candidate, with a Bayesian-style rational receiver as the a=0 special case.
What is the paper's key finding?
LLMs weigh evidence relative to their own priors rather than rationally: measured prior weights are 0.20–0.65 (below the rational 1), errors aligned with the model's own biases are adopted more readily than random errors, and models adopt supplied answers at 93–100% even immediately after judging them invalid — meaning verification does not control use.
Which models were tested?
Twelve instruction-tuned models from four families: Llama-3.1/3.3 (8B and 70B), Qwen3 (4B to 32B), Gemma-4 (E4B and 31B), and Ministral 3 (3B, 8B, 14B), across eight reasoning domains including quantum mechanics, physics, genetics, and molecular biology, over more than ten million trials.
What are the practical implications for RAG and agent systems?
Source accuracy alone does not determine reliability: the same evidence helps a weak receiver and harms a strong one; errors correlated with the receiver's own characteristic mistakes are more damaging than random ones at the same rate; provenance premiums communicated only through writable text can be spoofed; and verification results need an authenticated causal route to actually change the model's answer.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt: five marked [...sampling gap...] regions skip parts of Sections 3.2–3.3, 5.1–5.3, 6.3–7.1, the middle of the references, and Appendices C–F; judgments touching those portions rely on the surrounding visible text.
- No explicit code release statement appears in the provided text; the only availability mention is an accompanying CSV of whole-corpus figures for Table 9 (Appendix C).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.