arXiv：cs.AI· Fan Yang, Yue Song·· 22 小时前
研究揭示注意力汇形成前的隐藏状态几何特征，并可通过扰动该几何影响其形成
Before Attention Sinks Form: Hidden-State Geometry Reveals and Modulates Their Formation
AI 导读
研究对比同一模型同一层不同提示词发现，注意力汇形成前约 2 至 3 层会出现短暂的隐藏状态频谱重组，该信号强于当前注意力信息。在 Qwen2.5、GLM-Z1 和 GPT-OSS 上，仅修改当前注意力计算后层隐藏状态的奇异值，即可相对范数匹配的随机频谱扰动改变后续注意力汇行为；在 Qwen2.5 中该抑制效应伴随下一层对位置 0 的 query-key 偏好下降。
正文
Abstract:Attention sinks are usually studied after they have formed. We ask whether a measurable hidden-state change appears beforehand and whether changing that geometry alters what happens later. To remove network depth as a shortcut, we compare prompts in the same model and at the same layer. When sink timing varies across prompts, simple spectral summaries of the hidden states add information beyond current attention, with the clearest prompt-specific signal in a short window about two to three layers before sink formation. The predictive strength varies across cohorts, so we treat the spectrum as a readout of a pre-sink state change rather than a universal warning statistic. The signal is broader than one measured extreme activation. We then modify only the singular values of the post-layer hidden states after current attention has already been computed. These interventions change later sink behavior in Qwen2.5, GLM-Z1, and GPT-OSS relative to norm-matched random spectral perturbations. In Qwen2.5, the immediate suppressive effect is accompanied by a reduction in next-layer query--key preference for position~0, and the learned spectral directions remain stable when estimated from substantially less training data. Together, the results show that attention sinks are preceded by a short spectral reorganization of hidden states, and directly perturbing this geometry can causally alter subsequent sink formation.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.16745 [cs.LG] |
| (or arXiv:2610.16745v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.16745 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Fan Yang [view email]
[v1]
Mon, 12 Oct 2026 13:14:10 UTC (444 KB)
来源：arXiv：cs.AI · arxiv.org