arXiv cs.AI· Vardhan Dongre, Joseph Hsieh, Viet Dac Lai, Seunghyun Yoon, Trung Bui, Dilek Hakkani-T\"ur·· 10 天前AI 评分36
LLM 为何在多轮对话中"断线"：注意力关闭与残差表征的通道转移机制
When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction
AI 导读
研究提出"通道转移"机制解释 LLM 多轮交互中的指令遗忘：目标定义 token 通过注意力变得难以访问，但目标信息可能仍保留在残差表征中。团队提出 Goal Accessibility Ratio（GAR）指标，结合滑动窗口消融与残差流探针，在 Mistral 上强制关闭注意力通道后，20 项事实回忆任务准确率从接近满分骤降至 11%，且角色约束违规率超过对抗压力基线。
正文 · 原文
Abstract:Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instructions, persona, and rules. This degradation has been measured behaviorally but not mechanistically explained. We propose a channel-transition account: goal-defining tokens become less accessible through attention, while goal-related information may persist in residual representations. We introduce the Goal Accessibility Ratio (GAR), measuring attention from generated tokens to task-defining goal tokens, and combine it with sliding-window ablations and residual-stream probes. When attention to instructions closes, what survives reveals architecture. Across architectures, the transition yields qualitatively distinct failure modes: some models preserve goal-conditioned behavior at vanishing attention, others fail despite decodable residual goal information, and the layer at which this encoding emerges varies from 2 to 27. A within-model causal ablation that force-closes the attention channel in Mistral collapses recall from near-perfect to 11% on a 20-fact retention task and raises persona-constraint violations above an adversarial-pressure baseline without user pressure, with both effects emerging at the predictable crossover turn. Linear probes recover per-episode recall outcomes from residual representations with AUC up to 0.99 across all four primary architectures, while input embeddings remain at chance. Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure. We contribute GAR as a diagnostic, the channel-transition framework as a controlled mechanistic account, and a parametric prediction of failure timing under windowed attention closure.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2605.12922 [cs.AI] |
| (or arXiv:2605.12922v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2605.12922 arXiv-issued DOI via DataCite |
Submission history
From: Vardhan Dongre [view email]
[v1]
Wed, 13 May 2026 02:58:18 UTC (12,450 KB)
[v2]
Tue, 6 Oct 2026 16:57:10 UTC (12,450 KB)
来源：arXiv cs.AI · arxiv.org