Persistent Sparse Autoencoders: Learning Feature Timescales in Language Models

• URL: http://arxiv.org/abs/2607.17117v1
• Date: Sun, 19 Jul 2026 08:07:32 GMT
• Title: Persistent Sparse Autoencoders: Learning Feature Timescales in Language Models
• Abstract summary: We introduce Persistent Sparse Autoencoders (Persistent SAEs)<n>Persistent SAEs learn a persistence coefficient for each feature, allowing the model to learn which features should persist and for how long.<n>Our experiments show that they retain competitive reconstruction quality while learning a spectrum of feature timescales.
• Score: 23.93187817296646
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders (SAEs) decompose language model activations into sparse features, but standard SAEs encode each token independently and do not expose information that persists across a sequence. We introduce Persistent Sparse Autoencoders (Persistent SAEs), which extend standard SAEs by learning a persistence coefficient for each feature, allowing the model to learn which features should persist and for how long. Our experiments show that they retain competitive reconstruction quality while learning a spectrum of feature timescales: fast features behave as locally interpretable detectors, whereas slow features concentrate topic-level information in a persistent state. Moreover, as shown in a prompt-injection monitoring case study, slow features preserve detection signals and remain causally effective over long contexts. These results suggest that Persistent SAEs open up new opportunities for interpreting and monitoring language models through persistent semantic representations.

Related papers

• Tracking the Truth: Object-Centric Spatio-Temporal Monitoring for Video Large Language Models [154.39583176906893]
  Large language models (MLs) have advanced understanding, they remain highly prone to hallucinations in scenes.<n>We argue this stems from a failure intemporal monitoring, the ability to persistently track object dynamic identities, states, and relations over time.<n>Existing benchmarks obscure this deficit relying on single final-answer queries that can often be resolved via local visual cues or statistical priors.
  arXiv  Detail & Related papers  (2026-05-09T14:32:36Z)
• Beyond Visual Cues: Semantic-Driven Token Filtering and Expert Routing for Anytime Person ReID [10.58405139566836]
  Any-Time Person Re-identification (AT-ReID) necessitates the robust retrieval of target individuals under arbitrary conditions.<n>Existing methods are highly relying on pure visual features, which are prone to change due to environmental and time factors.<n>We propose a novel framework that leverages the ability of Large Vision-Language Models (LVLMs) to generate identity consistency text.
  arXiv  Detail & Related papers  (2026-04-16T14:49:30Z)
• In-Sync: Adaptation of Speech Aware Large Language Models for ASR with Word Level Timestamp Predictions [51.236148875603185]
  We extend an existing speech-aware language model to predict timestamps directly alongside transcripts.<n>We introduce a set of novel lightweight training strategies that improve alignment robustness while preserving recognition quality.
  arXiv  Detail & Related papers  (2026-04-14T20:56:24Z)
• TS-Haystack: A Multi-Scale Retrieval Benchmark for Time Series Language Models [4.387988928531881]
  Time Series Language Models (TSLMs) are emerging as unified models for reasoning over continuous signals in natural language.<n>Existing models are typically trained and evaluated on short sequences, while real-world time-series sensor streams can span millions of datapoints.<n>We introduce TS-Haystack, a long-context temporal retrieval benchmark comprising ten task types across four categories.
  arXiv  Detail & Related papers  (2026-02-15T15:50:02Z)
• Priors in Time: Missing Inductive Biases for Language Model Interpretability [58.07412640266836]
  We show that Sparse Autoencoders impose priors that assume independence of concepts across time, implying stationarity.<n>We introduce a new interpretability objective -- Temporal Feature Analysis -- which possesses a temporal inductive bias to decompose representations at a given time into two parts.<n>Our results underscore the need for inductive biases that match the data in designing robust interpretability tools.
  arXiv  Detail & Related papers  (2025-11-03T18:43:48Z)
• Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability [31.30541946703775]
  Translating internal representations and computations of models into concepts that humans can understand is a key goal of interpretability.<n>Recent dictionary learning methods such as Sparse Autoencoders provide a promising route to discover human-interpretable features.<n>But they exhibit a bias towards shallow, token-specific, or noisy features, such as "the phrase 'The' at the start of sentences"
  arXiv  Detail & Related papers  (2025-10-30T17:59:30Z)
• Temporal-Aware Iterative Speech Model for Dementia Detection [0.0]
  Current methods for automated dementia detection using speech rely on static, time-agnostic features or aggregated linguistic content.<n>We introduce TAI-Speech, a Temporal Aware Iterative framework that dynamically models spontaneous speech for dementia detection.<n>Our work provides a more flexible and robust solution for automated cognitive assessment, operating directly on the dynamics of raw audio.
  arXiv  Detail & Related papers  (2025-09-26T01:56:07Z)
• Dense SAE Latents Are Features, Not Bugs [86.50389855919292]
  We show that dense latents serve functional roles in language model computation.<n>We identify classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction.
  arXiv  Detail & Related papers  (2025-06-18T17:59:35Z)
• StPR: Spatiotemporal Preservation and Routing for Exemplar-Free Video Class-Incremental Learning [79.44594332189018]
  Class-Incremental Learning (CIL) seeks to develop models that continuously learn new action categories over time without previously acquired knowledge.<n>Existing approaches either rely on forgetting, raising concerns over memory and privacy, or adapt static image-based methods that neglect temporal modeling.<n>We propose a unified and exemplar-free VCIL framework that explicitly disentangles and preserves information.
  arXiv  Detail & Related papers  (2025-05-20T06:46:51Z)
• On the Consistency of Video Large Language Models in Temporal Comprehension [57.985769348320616]
  Video large language models (Video-LLMs) can temporally ground language queries and retrieve video moments.<n>We conduct a study on prediction consistency -- a key indicator for robustness and trustworthiness of temporal grounding.
  arXiv  Detail & Related papers  (2024-11-20T00:47:17Z)
• Temporal Blind Spots in Large Language Models [20.631107338678234]
  Large language models (LLMs) have recently gained significant attention due to their unparalleled ability to perform various natural language processing tasks. This study investigates the underlying limitations of general-purpose LLMs when deployed for tasks that require a temporal understanding.
  arXiv  Detail & Related papers  (2024-01-22T16:20:14Z)
• Multi-scale temporal network for continuous sign language recognition [10.920363368754721]
  Continuous Sign Language Recognition is a challenging research task due to the lack of accurate annotation on the temporal sequence of sign language data. This paper proposes a multi-scale temporal network (MSTNet) to extract more accurate temporal features. Experimental results on two publicly available datasets demonstrate that our method can effectively extract sign language features in an end-to-end manner without any prior knowledge.
  arXiv  Detail & Related papers  (2022-04-08T06:14:22Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.