arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.04918cs.LGcs.CLcs.CY
收藏

模拟采纳:在大语言模型上下文冲突解决中解耦幅度与方向

Simulated Adoption: Decoupling Magnitude and Direction in LLM In-Context Conflict Resolution

South China University of Technology(华南理工大学);The University of Hong Kong(香港大学)

Long Zhang, Fangwei Lin

AI总结:

研究揭示大语言模型在解决上下文冲突时通过几何位移机制实现模拟采纳,而非单纯抑制知识幅度。

摘要:

大型语言模型(LLMs)经常优先考虑上下文中的冲突信息,而非先前的参数记忆,这种现象通常被称为趋顺或服从。然而,这种行为的机制性实现仍不明确,特别是模型如何通过服从解决这些知识冲突,以及这种抑制是否源于残差流中的信号幅度稀释或方向几何改变。为此,我们对Qwen-3-4B、Llama-3.1-8B和GLM-4-9B进行了逐层几何分析,将反事实上下文引起的残差流更新分解为径向(基于范数)和角向(基于余弦)成分。我们的实证结果驳斥了

英文摘要:

Large Language Models (LLMs) frequently prioritize conflicting in-context information over pre-existing parametric memory, a phenomenon often termed sycophancy or compliance. However, the mechanistic realization of this behavior remains obscure, specifically how the model resolves these knowledge conflicts through compliance, and whether this suppression arises from signal magnitude dilution or directional geometric alteration within the residual stream. To resolve this, we conducted a layer-wise geometric analysis across Qwen-3-4B, Llama-3.1-8B, and GLM-4-9B, decomposing the residual stream updates induced by counter-factual contexts into radial (norm-based) and angular (cosine-based) components. Our empirical results reject the universality of the "Manifold Dilution" hypothesis, as two of the three architectures maintained stable residual norms despite exhibiting significant performance degradation on factual queries. Instead, we observed that compliance is consistently characterized by "Orthogonal Interference," where the conflicting context injects a steering vector that is quasi-orthogonal to the ground-truth direction, effectively rotating the hidden state representation. This suggests that models do not "unlearn" or suppress the magnitude of internal truths but rather employ a mechanism of geometric displacement to bypass the correct unembedding vector, effectively simulating adoption while preserving the original structural magnitude. These findings challenge scalar confidence metrics for detecting hallucinations and underscore the necessity of vectorial monitoring to distinguish between genuine knowledge integration and superficial in-context mimicry.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑