arXiv cs.CL· Haoyan Luo, Mateo Espinosa Zarlenga, Mateja Jamnik·· 10 小时前AI 评分34
Persistent SAEs:为语言模型特征学习专属时间尺度
Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations
AI 导读
研究者提出 Persistent Sparse Autoencoders(Persistent SAEs),在标准 SAE 基础上为每个特征学习一个持久系数,使特征专属时间尺度从重建中自然涌现。实验显示其重建质量保持竞争力,短时间尺度特征局部可解释,长时间尺度特征则累积更高层上下文信息。在提示词注入监控案例中,慢特征能在长上下文中保留注入信号并因果影响监控判断。
正文 · 原文
Abstract:Sparse autoencoders (SAEs) decompose language model activations into sparse features, yet these models traditionally encode each token independently, failing to expose information that persists across a sequence. We first show that temporal persistence can naturally emerge in standard SAE features: after a feature activates, the hidden state remains aligned with its direction, and past activations reduce the reconstruction error of later hidden states. How long this lasts varies widely across features. We therefore introduce Persistent Sparse Autoencoders (Persistent SAEs), an extension of standard SAEs that learns a persistence coefficient for each feature, allowing feature-specific timescales to emerge from reconstruction. Our experiments show that Persistent SAEs retain competitive reconstruction quality while learning a spectrum of timescales: short-timescale (fast) features remain locally interpretable, whereas long-timescale (slow) features accumulate higher-level contextual information. Moreover, we show in a prompt-injection monitoring case study that slow features retain injection signals over long contexts and causally affect monitor judgements. These results suggest that Persistent SAEs offer new opportunities for interpreting and monitoring language models via persistent features.
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2607.17117 [cs.LG] |
| (or arXiv:2607.17117v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2607.17117 arXiv-issued DOI via DataCite |
Submission history
From: Haoyan Luo [view email]
[v1]
Sun, 19 Jul 2026 08:07:32 UTC (1,036 KB)
[v2]
Wed, 2 Sep 2026 11:15:48 UTC (912 KB)
[v3]
Thu, 19 Nov 2026 08:02:31 UTC (2,067 KB)
来源:arXiv cs.CL · arxiv.org