- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Persistent Sparse Autoencoders (Persistent SAEs), which extend standard SAEs by introducing learnable feature-specific persistence coefficients to capture temporal dependencies in language model representations. The method automatically learns different feature timescales, with fast features capturing local information and slow features retaining higher-level contextual information. Experiments on Pythia and Gemma demonstrate competitive reconstruction quality, improved retention of subject information, and the potential of slow features for long-context prompt-injection monitoring.
Critical strengths
The paper proposes learning an independent persistence coefficient for each SAE feature, allowing the model to automatically distinguish between fast and slow features. Experiments demonstrate that slow features can effectively retain contextual information, such as the current topic, and show promising potential for prompt-injection monitoring.
Critical weaknesses
- Limited Novelty and Missing Baselines
Prior works, including Temporal Sparse Autoencoders (Bhalla et al., ICLR 2026) and Priors in Time (Lubana et al., ICLR 2026), have already explored temporal dependencies in SAE representations, including the separation of context-dependent and newly emerging information. The main architectural contribution of this paper is introducing learnable feature-wise exponential moving average (EMA) coefficients. Although the authors provide ablations on shared persistence coefficients, comparisons with simpler baselines, such as standard SAEs with fixed or post-hoc learned EMA, are missing. Without these comparisons, it is unclear whether jointly learning feature-specific timescales is necessary.
- Unclear Mechanism of Long-Term Information Retention
The authors train Persistent SAEs on sequences of 256 tokens but evaluate prompt-injection signals up to 350 tokens later. According to the state update equation, historical information decays exponentially. Therefore, it remains unclear whether the observed long-term retention comes from genuinely preserved historical information or repeated activation of the same features by subsequent tokens. Additional experiments are needed to distinguish these mechanisms.
- Degradation in Reconstruction Quality
Persistent SAEs show some degradation in reconstruction quality. For example, on Gemma-2B, the FVE decreases from 0.72 (BatchTopK) to 0.66 (Persistent SAE). Furthermore, sparsity is imposed on token-wise activations rather than directly on the accumulated states. Comparisons under matched state sparsity, reconstruction quality, and computational budgets would provide a clearer assessment of the trade-offs.
- Limited Practical Validation
The prompt-injection experiments are conducted using an external monitor and only failed attack trajectories. Although the results demonstrate that persistent features can influence the monitor's judgments, they do not establish whether Persistent SAEs improve end-to-end attack detection or reduce attack success rates in actual agents.
Refs [1] Bhalla, U., Oesterling, A., Verdun, C. M., Lakkaraju, H., & Calmon, F. (2026). Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability. International Conference on Learning Representations (ICLR 2026). [2] Lubana, E. S., Rager, C., Hindupur, S. S. R., et al. (2026). Priors in Time: Missing Inductive Biases for Language Model Interpretability. International Conference on Learning Representations (ICLR 2026).
Questions for authors
Please ref the Weaknesses.
5 comments
Author Response (2/3)
W1 (Part II): Missing baselines
Regarding the missing baselines, we ran a new ablation study in Appendix G.4. As shown in Table 9, we designed a baseline that keeps the trained BatchTopK encoder fixed and accumulates its activations over time with a fixed lambda. We test two types of simple baselines: a fixed EMA, where all features share one , and a post-hoc learned EMA, where each feature learns its own with the same parametrisation as ours. We report FVE and the sparse probing AUC from Section 5.3. The results are shown below (for details, please see Appendix G.4 in the updated manuscript).
| Method | FVE | AUC, 0.5% features | 1% | 5% | 10% | 100% |
|---|---|---|---|---|---|---|
| BatchTopK | 0.73 | 0.52 | 0.52 | 0.58 | 0.62 | 0.79 |
| + fixed EMA, | 0.33 | 0.62 | 0.64 | 0.74 | 0.78 | 0.91 |
| + post-hoc learned EMA | 0.73 | 0.55 | 0.56 | 0.61 | 0.65 | 0.81 |
| Persistent SAE (slowest first) | 0.73 | 0.79 | 0.80 | 0.81 | 0.81 | 0.83 |
We show that neither type of simpler baseline matches the performance of Persistent SAE. A fixed EMA smooths every feature with the same timescale, so it loses much of the reconstruction quality. This uniform smoothing also explains its high AUC when the probe uses all features (0.91): every feature becomes an average over the preceding context, which helps predict the subject but no longer describes the current token (FVE 0.33). The post-hoc learned EMA maintains reconstruction quality, but it loses the meaningful ranking based on features' temporal persistence. The Persistent SAE maintains the same FVE as the baselines and provides meaningful temporal profiles of features, showing that only the slowest 5% of features capture most of the subject information. Therefore, we show that jointly learning the features and their timescales is necessary for this result.
We have added the full results to Appendix G.4 (Table 9) and refer to them in Section 5.1.
W2: Unclear mechanism of long-term information retention
Our reading of your second concern is that, because the state decays exponentially and our models are trained on 256-token sequences by default, it is unclear what contributes to the results on longer sequences. If this is a correct interpretation of your concern, we believe the following points may help clarify our work.
First, the 256-token training sequence length does not impose a hard limit on inference. The learned recurrence is time-homogeneous and applies to any longer sequence. In Appendix G.2 (Figure 14) we show that the learned persistence distributions remain broadly stable across training lengths from 32 to longer tokens.
Second, we would like to clarify that the benefits of Persistent SAEs come from learning to integrate information over time. The persistent state combines previously accumulated information with new evidence from subsequent tokens. Therefore, long-term signal retention can arise from both historical carry-over and continued activation of relevant features. We argue that these mechanisms are complementary, as subsequent activations can themselves reflect historical context that is already encoded and carried in the LLM's hidden states, which Persistent SAEs rediscover without misidentifying it as a new feature.
We have clarified this distinction in Section 5.4 in the updated manuscript. Nevertheless, if we misinterpreted your concern, we would sincerely appreciate further clarification so we can address it properly.
Author Response (3/3)
W3: Degradation in Reconstruction Quality
After we corrected the mismatch in our evaluation, the Persistent SAE no longer shows a drop in reconstruction quality for Gemma (we note that there was no significant drop in reconstruction for Pythia). Please see our reply to W1 from Reviewer WqJT for details on why this was the case and how we addressed it. Accordingly, we updated Table 1 with the corrected results, and we thank the reviewer for identifying this mismatch.
Regarding your second point, as you note, because the state accumulates earlier activations, it has more non-zero activations per token. In Appendix B.3, we show that the state has about 500 non-zero activations, compared with about 20 for . To test whether reconstruction quality relies on these extra entries, we ran a new ablation study in Appendix F.1, where we keep only the largest activations the decoder can read and measure FVE to see whether reconstruction quality depends on the additional non-zero activations. We show the headline results here (the full tables are provided in Table 7, Appendix F.1).
| Model | Method | Full | ||||
|---|---|---|---|---|---|---|
| Pythia-160M | BatchTopK | 0.929 | 0.931 | 0.931 | 0.931 | 0.931 |
| Persistent SAE | 0.921 | 0.931 | 0.933 | 0.933 | 0.933 | |
| Gemma-2-2B | BatchTopK | 0.717 | 0.729 | 0.729 | 0.729 | 0.729 |
| Persistent SAE | 0.689 | 0.718 | 0.725 | 0.725 | 0.725 |
There we see that, with the same budget as standard SAEs, Persistent SAE's reconstruction quality stays within a similar range to BatchTopK. About 40 activations for reconstruction already recover full reconstruction quality on both models. Since most non-zero state activations are small, decayed traces of earlier activations, Persistent SAE does not gain reconstruction quality directly from them. The benefit of each feature keeping a decaying trace of its earlier activations is that it lets the model learn from reconstruction which features should carry context over longer ranges. For this reason, we impose sparsity on as in standard SAEs.
To take this feedback into account, we have added the matched-budget comparison results, including all baselines and TFA, to Appendix F.1 and refer to them in Section 5.1.
W4: Limited practical validation
We would like to clarify that we explicitly argue this precise point in our limitations when we also discuss potential future work (lines 499-501: “Finally, we note that, in our prompt-injection case study, we use an external replay monitor of agent trajectories; therefore, we do not establish that persistent states have a causal effect on the original agent’s actions.”). Nevertheless, we include prompt-injection monitoring as a case study (Section 5.4) to show that fast and slow features behave as expected in practice: fast features are good local detectors, while slow features are causally relevant to higher-level monitor decisions. The main contribution of the paper is the Persistent SAE itself, which learns feature-specific timescales and provides a clean axis for empirically analysing these temporal roles. Exploring more complex uses of these models in agentic settings is something we believe would merit a whole new paper, so we left this for future work (as stated in Section 6).
Author Response (1/3)
Thank you for taking the time to go over our manuscript and provide some helpful feedback. We address each question below and have updated our paper accordingly, marking new changes in a different colour for easy detection. If you have further concerns, please let us know. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our responses below.
Our work’s novelty is twofold. First, we observe that existing SAE features capture some temporal information. Second, motivated by this observation, we build a method that learns representations that explicitly expose this temporal structure, allowing more detailed analysis and interpretation of learned features. We elaborate on both aspects below.
This results in a method in which timescales emerge naturally from the data, enabling us to easily distinguish features that contribute to longer contexts from those that are short-lived. Our results show that the temporal coefficients we learn for features span a broad range and agree with each feature’s measured empirical persistence at a much finer granularity than Temporal SAE (Figure 3). More importantly, our formulation provides a simple, interpretable ranking of all features by timescale without any labels, which neither prior method provides. This motivates our later analysis of their division of labour, which shows clearer and better results than these temporal baselines. Finally, we emphasise that our method is much simpler than TFA (e.g., it uses orders of magnitude fewer parameters, see Appendix B.3), and, even with fewer parameters and without complex modules, our features outperform TFA’s features in subject separation (see new results in Appendix F.2, Figure 8) and in reconstruction under equivalent budgets (see new results in Appendix F.1, Table 7).
In addition to the comparisons against TFA described above, we have taken this feedback into account and updated the paragraph discussing our approach in light of related work (last paragraph of Section 3) to make this distinction clearer.