acceptodds
Under review as a conference paper at ICLR 2027

Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations

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.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_97SqOpenReview ↗
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
  1. 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.

  1. 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.

  1. 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.

  1. 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
AuthorsOpenReview ↗

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.

W1 (Part I): Limited novelty

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.

  1. Prior work motivates temporal structure in different ways. Bhalla et al. (2026) start from the intuition that high-level information changes slowly, so they assign 20% of the features to a high-level group before training (i.e., slow features are hard-coded). Then, they use an additional loss to make these features change slowly across nearby tokens. Lubana et al. (2026) study temporal dependencies in hidden states and demonstrate temporal correlations between them. Unlike both works, we study the temporal structure of the units interpretability work focuses on, i.e., individual SAE features. Specifically, we show three important new observations in Section 3: (1) standard SAE features stay active much longer on natural text (Figure 1b), (2) past feature activations help reconstruct later hidden states (Figure 1c), and (3) temporal persistence varies widely from feature to feature. These observations strongly suggest that individual features need their own learned timescale, motivating our method.
  2. Persistent SAEs learn a persistence coefficient for every feature from reconstruction alone, without (1) a predefined split of features or an additional loss to model temporal structure (e.g., Temporal SAE from Bhalla et al.), or (2) introducing additional complex prediction modules (TFA from Lubana et al.).

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.

AuthorsOpenReview ↗

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 λ∈{0.5,0.9,0.99}\lambda \in \{0.5, 0.9, 0.99\}, and a post-hoc learned EMA, where each feature learns its own λj\lambda_j 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).

MethodFVEAUC, 0.5% features1%5%10%100%
BatchTopK0.730.520.520.580.620.79
+ fixed EMA, λ=0.9\lambda=0.90.330.620.640.740.780.91
+ post-hoc learned EMA0.730.550.560.610.650.81
Persistent SAE (slowest first)0.730.790.800.810.810.83

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.

AuthorsOpenReview ↗

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 a(t)\mathbf{a}^{(t)}. 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 KK 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).

ModelMethodK=20K=20K=40K=40K=80K=80K=160K=160Full
Pythia-160MBatchTopK0.9290.9310.9310.9310.931
Persistent SAE0.9210.9310.9330.9330.933
Gemma-2-2BBatchTopK0.7170.7290.7290.7290.729
Persistent SAE0.6890.7180.7250.7250.725

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 a(t)\mathbf{a}^{(t)} 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).

Reviewer_WqJTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Persistent SAEs: each feature keeps an exponential moving average of its activations with a learned persistence coefficient λj\lambda_j, and the decoder reconstructs from this state. Unlike prior temporal SAEs, timescales are learned rather than pre-assigned, giving a continuous spectrum. A motivating study shows that SAE features persist far longer on natural text than on a sentence-shuffled control. On Pythia-160M and Gemma-2-2B, slow features are smoother and cluster by subject, fast features are more interpretable, and reconstruction matches baselines on Pythia but drops on Gemma. Downstream, λ\lambda-ranking selects subject-predictive features without supervision, and in an AgentDojo prompt-injection case study slow-feature states retain the injection signal and give the only consistently positive steering effect on a monitor.

Critical strengths
  1. Learned per-feature persistence. A simple, well-motivated step beyond fixed slow/fast partitions (Bhalla et al.; Lubana et al.), supported by the shuffled-control persistence study (Sec. 3) and ablations over λ\lambda granularity, initialization, context length and scale.
  2. Long-context causal validation. In Fig. 6, slow-feature directions are the only ones with a consistently positive effect on the monitor's logit margin at every distance up to 350 tokens (+0.1, with CIs); all baseline SAE directions are near zero or negative.
Critical weaknesses
  1. Low reconstruction on Gemma-2-2B. FVE is 0.66 vs 0.71–0.72 for the baselines trained on the same data (Table 1), despite Sec. 3 motivating persistence partly by its contribution to reconstruction.
Questions for authors
  1. Smoothness for your model is computed on h(t) (App. E.1), which is smooth by construction, but on token-wise activations for baselines. How do slow features score on a(t)?
  2. Bhalla et al. report Gemma FVE of 0.75–0.76 for all SAEs at the same layer, width and k, versus 0.62–0.72 here, while Pythia numbers agree. Could you please explain this? Please also report L0.
2 comments
AuthorsOpenReview ↗

Author Response

Thank you for the review! We address your concerns and questions 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.

W1 & Q2: Low reconstruction quality

We sincerely appreciate your keen observation about the low reconstruction quality of our method in Gemma. After careful analysis, we identified a small issue in our Table 1 evaluation that affected only the reconstruction metrics for Gemma in that table, and nothing else. We have now fixed that issue (and updated Table 1 accordingly) and describe the problem below.

  1. Setup. We first built the evaluation pipeline for Table 1 on Pythia-160M and then applied the same pipeline to Gemma-2-2B for both reconstruction metrics (FVE and cosine similarity).
  2. Role of Table 1. We intended the reconstruction metrics as a sanity check to ensure that adding persistence does not severely impair reconstruction quality. Therefore, we did not consider the original small gap on Gemma particularly negative. However, we agree that the disagreement with Bhalla et al. on Gemma needed an explanation, so we investigated the issue.
  3. Cause. Gemma-2's tokeniser adds a beginning-of-sequence (BOS) token by default, whereas Pythia's tokeniser does not. During SAE training, we tokenised using the default settings and ran the language model with the BOS token. However, our evaluation code for the reconstruction metrics disabled special tokens by default (because Pythia's tokeniser does not include a BOS token), so Gemma-2 received no BOS token. Without it, Gemma-2 uses the first real token as the attention sink instead: that token gets a very large activation norm, and the activations of later tokens also shift. Every SAE was therefore evaluated on Gemma activations that differed from those it was trained on. For Pythia, the inputs were identical with or without this setting, so training and evaluation matched. This explains why our Pythia numbers agreed with Bhalla et al. while our Gemma numbers did not.
  4. Corrected results. Since the problem affected all baselines for Gemma-2 in Table 1 for the FVE, cosine similarity, and smoothness columns, we reran the evaluations and provided the updated Table 1 in the revised paper. The corrected results are shown below.
Gemma-2-2BFVE (↑\uparrow)Cosine sim. (↑\uparrow)L0L_0
BatchTopK0.730.8919.7
Matryoshka0.720.8819.9
Temporal0.700.8819.9
Persistent0.730.8922.6

With the corrected evaluation, our method now matches the reconstruction quality of the strongest baseline, and the results are consistent with those of Bhalla et al. (2026). Note that the results are not identical to those of Bhalla et al. (2026) because they did not report the total number of training tokens needed to match their setup. We provide training details in Appendix B.3 and train each baseline using the same total training token budget for a fair comparison. We also report L0L_0 for sparse activations in Table 1. We also include an additional ablation study of reconstruction quality in Appendix F.1 to report L0L_0 for the state and to show that Persistent SAEs reconstruct as well as standard SAEs with the same L0L_0 budget.

  1. Code. To support reproducibility, we share the code and training scripts at this link.

Q1: Smoothness of h(t)h(t) and a(t)a(t)

We evaluated smoothness on h(t)h^{(t)} because it plays the same role in reconstruction as the baseline SAEs' activations. The Temporal SAE is evaluated in the same way, and its high-level features are also smooth by design, because its contrastive loss explicitly encourages them to be. To address your question, we evaluated temporal smoothness on both the token-wise activations a(t)a^{(t)} and persistent states h(t)h^{(t)}, using the same metric from Table 1 (lower is smoother).

RepresentationPythiaGemma
Slow 1%, on h(t)0.0470.034
Slow 1%, on a(t)0.2060.263
Fast 50%, on h(t)0.1950.175
Fast 50%, on a(t)0.2150.202

These additional results show that the smoothness of slow features arises from temporal integration. The token-wise activations of slow features are not smooth. They are in fact less smooth than the baseline activations (0.13–0.15 in Table 1), so the smoothness of slow features comes from learned temporal integration of these activations. In contrast, fast features show similar smoothness in activations and states, which is consistent with their smaller learned persistence coefficients. We have included this table in Appendix E.2 in our updated manuscript.

AuthorsOpenReview ↗

Follow-up before the discussion period ends

Dear Reviewer,

Thank you again for all your feedback! As the discussion period is coming to an end, we wanted to ask whether you have any further questions or concerns in light of our replies above? If so, we would be happy to discuss and provide further clarifications whilst the discussion period is open. Otherwise, if our rebuttal addressed your main concerns and questions, we would sincerely appreciate it if you could let us know. Thank you in advance!

Reviewer_dEQ7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper tackles the static nature of standard SAE decomposition by proposing persistent SAEs, a technique that extends SAEs by learning a persistence coefficient for features, which quantifies the effect that an activating feature has on future tokens. The authors first demonstrate that SAEs features show persistence across a context and consistent decay patterns. They then show that the persistent SAE architecture can capture this decay, while retaining good reconstruction. Moreover, the authors analyze the "fast" vs. "slowly" decaying features according to their learned persistence value, finding interpretable differences between the two. Finally, they use a prompt injection experiment to show that slow features maintain their importance even after several hundred tokens. Taken together, the results support the authors' case for the importance of incorporating temporal dynamics into current SAE architectures.

Critical strengths
  • The paper addresses a topic that is of importance to the mechanistic interpretability community: the failure of standard SAEs, and how they could be enhanced to better capture internal states. I believe the paper will be of interest to the community.
  • The paper is very clearly written, has a logic flow, and contains nice figures that help in explaining the concepts and conveying the authors' main points. While I would have mentioned this regardless, it is particularly notable given the prevalence of bad LLM writing in this year's ICLR submissions, and so I wanted to thank the authors for their efforts to make their paper clear to read.
  • Most importantly, the paper identifies a novel and interesting phenomenon, the persistence and decay of feature presence across tokens, incorporates this finding as an inductive bias into an SAE architecture, and shows this inductive bias allows for new discoveries (e.g., distinguishing slow and fast features and the things they activate for) while retaining reconstruction performance. Overall, I found the analysis quite compelling and interesting.
Critical weaknesses
  • The biggest weakness of the paper in my opinion, is that given the work of Bhalla (2026) and Lubana (2026), it's contribution of adding temporal structure into SAEs is not very novel. While the paper identifies a new phenomenon that justifies its SAE architecture, I think the case for a substantial contribution beyond that of Bhalla (2026) and Lubana (2026) needs to be made more strongly (see my first question for example).
  • I find it quite unlikely that the decay coefficients for features would be static across contexts, and thus the new architecture, despite incorporating more temporal structure, may still be too static. Have you considered adding cross feature interactions (as you mention in the discussion) or another context-dependent way that could capture context-specific decay rates? The somewhat static nature of the current architecture may also prevent more interesting slow features from appearing.
Questions for authors
  • I am not sure I understand the distinction you made between the work of Lubana (2026) characterizing the hidden state level vs. your work focusing on the feature level. Could we not think of features as hidden states?
  • Small comment: If you are defining the two feature types as fast and slow, please consistently use this terminology rather than change it to high vs. low (as is done in table 1 for example).
4 comments
AuthorsOpenReview ↗

Author Response (1/2)

Thank you for the very thoughtful review! Your feedback was very helpful in improving our manuscript, and we are glad you found that our work addresses a problem “of importance”, is “clearly written”, and identifies a “novel and interesting phenomenon”. We address your concerns and questions 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.

W1: Contribution and novelty

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.

  1. Prior work motivates temporal structure in different ways. Bhalla et al. (2026) start from the intuition that high-level information changes slowly, so they assign 20% of the features to a high-level group before training (i.e., slow features are hard-coded). Then, they use an additional loss to make these features change slowly across nearby tokens. Lubana et al. (2026) study temporal dependencies in hidden states and demonstrate temporal correlations between them. Unlike both works, we study the temporal structure of the units interpretability work focuses on, i.e., individual SAE features. Specifically, we show three important new observations in Section 3: (1) standard SAE features stay active much longer on natural text (Figure 1b), (2) past feature activations help reconstruct later hidden states (Figure 1c), and (3) temporal persistence varies widely from feature to feature. These observations strongly suggest that individual features need their own learned timescale, motivating our method.
  2. Persistent SAEs learn a persistence coefficient for every feature from reconstruction alone, without (1) a predefined split of features or an additional loss to model temporal structure (e.g., Temporal SAE from Bhalla et al.), or (2) introducing additional complex prediction modules (TFA from Lubana et al.).

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.

Q1: Distinction between hidden states and features

The hidden state is a single dense vector that mixes many features and concepts. Lubana et al. (2026) measure temporal dependence on this whole vector, while this temporal dependence has also been studied in many works (Section 2). However, an SAE decomposes the hidden state into many sparse, interpretable directions (i.e., different “concepts” or “features”). We agree that features can be viewed also this way, however, we want to argue that the meaningful difference is resolution: a hidden-state temporal correlation may aggregate the effect of all features, so two sequences can show the same correlation even when very different features produce it. Our experiments provide strong evidence that persistence varies widely across features (Figure 1b). This leads our approach to learn representations that capture a wide range of temporal concepts. Finally, as we argue in our replies above, we achieve this without the complex, memory-intensive machinery required by Lubana et al.’s approach. We have added this point to Section 3.

AuthorsOpenReview ↗

Author Response (2/2)

W2: Static feature decay coefficients

If we understand the point correctly, the reviewer is concerned that, once trained, each feature has a fixed decay rate, regardless of the context in which it appears. We would like to address this concern from two perspectives: (1) the static nature of the decay coefficients and (2) the potential need for cross-feature persistence.

(1) Fixed decay coefficients do not imply static representations.

Fixed decay coefficients do not imply a fixed or unresponsive representation. Specifically, we note that Persistent SAEs already allow representations to adapt to context through two mechanisms: (i) context-dependent evidence (e.g., feature-specific sparse activations) and (ii) feature specialisation, where different features learn different persistence coefficients during training.

Under this design, the model cannot dynamically accelerate the decay of an individual slow feature when the context changes (i.e., there is no inhibition). However, our experiments show that slow-feature representations adapt well to changing semantic contexts, better than any baseline (e.g., our token-similarity experiments in Appendix F.2, Figure 8). Our results therefore suggest that when the context changes and the previous feature no longer receives supporting evidence, its state gradually decays, while new features corresponding to the current context can become active, thereby enabling easy context adaptation. We have added this point to Section 6 and Appendix F.2.

(2) Why feature-specific persistence?

A fixed, learned coefficient gives each feature an explicit, interpretable temporal profile while (a) preserving the simplicity and feature-level attribution of standard SAEs, and (b) facilitating post-hoc analysis and interpretation of features. Cross-feature interactions could introduce additional expressivity, but whether this leads to more informative or interpretable features remains an open empirical question. We believe our work provides a strong prerequisite and proof of concept for future work exploring cross-feature interactions, but it will require further empirical evidence to justify the added complexity.

Q2: Terminology of fast vs. slow features

Thank you so much for catching this small inconsistency! We have updated the paper to use "fast" and "slow" to describe features and revised the caption.

I thank the authors for their detailed response to my review and for engaging with my suggestions! I continue to think that experimenting with cross feature interactions (perhaps as an appendix experiment) would strengthen the paper, but I understand this added complexity is not necessarily justified by the data at the moment. I will maintain my weak acceptance score since I am not sure the paper is quite at the level of a 4, but I will advocate for its acceptance at the end of the review stage.

AuthorsOpenReview ↗

Follow-up Response

Thank you for the follow-up and for offering to support the paper. We first wanted to clarify why we do not train new SAEs that explicitly include cross-feature terms for rebuttal. First, adding this explicit dependency would give an SAE significantly more parameters than our Persistent SAE, so evaluating against our proposed method would not be fair. Second, and perhaps more challenging to address, all other evaluations in the paper that are not about reconstruction, and therefore the evaluations that showcase the benefits of learning persistence in SAE features, are built around single-feature persistence and cannot be directly adapted to cross-feature terms.

That being said, to get a better understanding of the potential impact that explicitly capturing cross-feature interactions may have, we have included a set of new experiments in Appendix H. We summarise our findings of these new experiments below.

Finding #1: Persistent SAE features still interact meaningfully (Appendix H.1).

Although each feature persists on its own for Persistent SAE, features on the same topic (e.g., sports, clinical studies, family) activate together far more often than chance. These topic groups also follow the subject: a feature’s set of co-activating features (what we call a feature's “partners”) stay more active when a new document continues the same subject than when it changes. Intuitively, this can be understood by noticing that the evidence used to update features in a Persistent SAE is derived from the model's latent space, which contains information about all features and can therefore already implicitly capture cross-feature interactions. This helps explain why a small set of slow features tracks the subject in Section 5.3.

Finding #2: Initial exploration suggests that cross-feature modelling needs evidence beyond reconstruction (Appendix H.2).

We extend the future-hidden-state reconstruction experiments summarised in Figure 1c to perform the reconstructions with groups of features. We compare four intuitive ways to choose these feature groups. Given Finding #1, we expected co-activating partners to help most. However, we find that the past information that helps reconstruction lies mainly along each feature's own direction: nearest directions help more than co-activating partners. This supports our choice to model persistence using one coefficient per feature, while the small gain from co-activation motivates future work explicitly modelling cross-feature interactions, which may need an objective beyond reconstruction to demonstrate its utility.


In summary, the interaction among features already emerges in the Persistent SAE without explicit cross-feature terms, and we motivate future work with an initial exploration of cross-feature modelling. We have included these results in Appendix H of the updated manuscript and refer to them in Sections 5.3 and 6. Thank you again for this suggestion; it really helped us clarify the scope of our design.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.