acceptodds
Under review as a conference paper at ICLR 2027

LineageSAE: Versioned Feature Vocabularies across SAE Retraining and Model Updates

Abstract

A feature index in a sparse autoencoder (SAE) is useful only if it keeps denoting the same feature after the SAE is retrained or the language model changes. Existing methods either encourage reproducibility or align independently learned dictionaries after training. Neither decides, for a saved parent and a changed view, which of its addresses to retain and which to free for new features. We introduce LineageSAE, a parent–child SAE that builds a versioned feature vocabulary from a saved parent and current activations without replaying the parent's data. A corrected marginal-utility test marks each parent address as inherited or released, and a parent-guided decoder handoff transfers reconstruction to the child. A temporary ranking loss uses the parent's pre-activation scores to align released addresses across independently trained children, including when ReLU outputs are zero. On SynthSAEBench and toy superposition with feature deletion, addition, and frequency changes, LineageSAE reaches child–child support Jaccard, the overlap of firing samples at the same address, of and without post-training assignment. The strongest matched baselines reach and . Reconstruction is close to -regularized TopK on SynthSAEBench; on Toy the NMSE is , against and for RA-SAE and -regularized TopK. A trained child keeps its index assignments and can parent the next update.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Mechanistic Interpretability, Representation Learning, Feature Consistency, Model Updating

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_AfR2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a method for maintain a sparse autoencoder's (SAE) vocabulary across retraining, including across distributional shifts in its inputs. This work is motivated by the need for a shared vocabulary when SAEs are retrained due to shifts in the LLM input datasets (which is evaluated on a shift from wiki text to code), and when the LLM changes, for example during fine-tuning. Given a parent SAE and the goal of training a child SAE on a different data distribution, LineageSAE first classifies latents as inherited or released using a correct marginal-utility criterion, before retraining the released latents in the child SAE using a parent-child decoder handoff mechanism. To maintain retrained latent ordering between children, a temporary response-ordering mechanism is used. The method is evaluated on two language model distribution shifts, and on toy models where known dataset features are removed, modified, and added.

Critical strengths

The toy model and SynthSAEBench model interesting distributional shifts, where features are replaced, or have their activation frequencies changed. On these synthetic tasks, LineageSAE outperform baselines and other methods for maintaining a shared feature vocabulary.

Critical weaknesses

Confusingly, the paper seems to present results from two distinct methods as equivalent in the main body results tables. The method described in the main body applies the response ordering approach to all released addresses, whereas the residual-learning method described in the appendix only applies response ordering to zero-response addresses. However, this alternative method is not restricted to the appendix - Tables 1 and 2 in the main body differ in which method they use. Table 2 and 28 present very different results as a consequence of this method split on SynthSAEBench. I don't think I can currently determine whether each method outperforms baselines on both synthetic and LLM tasks, or whether this is only the case when choosing the best of the variants to display in the main body.

An additional basic control that needs including is using the parent SAE unchanged on the shifted distribution. My suspicion is that on the new-domain and updated-model experiments, the distribution shift is not so large that the parent SAE is invalidated - see results from Gemmascope on evaluating base model SAEs on fine-tuned instruct models. Table 25 has related (but not identical) baselines that empirically support this concern: parent-copy continuation dominates LineageSAE across similarity and reconstruction metrics, and even the frozen decoder baseline is competitive with much better similarity scores and only reducing CE recovery from 95.80% to 95.69% - I'm not sure whether that change is significant as only point estimates are provided.

Table 4 demonstrates better detection but not end-to-end utility for correct utility over uncorrected utility. Corrected utility is much better at detecting deleted-feature addresses, however uncorrected-utility performs better on every address-coordination metric, and all variants recover the 51 new features. The end-to-end benefits of the corrected utility are therefore not clear to me.

If LineageSAE is finding important new features in the LLM distributional shift experiments, case studies and qualitative examples of such would be very useful, otherwise I can't assess whether this method is actually useful for interpretability practitioners.

Questions for authors
  1. Clarify which configuration of the method is used for all the main body tables, and provide results for each variant on all the tables (or commit to a single variant).
  2. How does the basic parent-copy control I propose compare with the results present in tables 1 and 3?
  3. What is the end-to-end benefits of corrected utility?
3 comments
AuthorsOpenReview ↗

Response to Reviewer AfR2 (R3)

Thank you for cross-checking the main text against the appendix. Below we answer W1-W4 and Q1-Q3 directly.

1. Method clarification and unified main-body commitment (W1, Q1)

The table mismatch in our initial submission was our mistake. Initial Tables 2 and 4 ran Section 2 (Sections 2.1-2.5: corrected utility release in Eqs. 3-4, parent-guided decoder handoff in Eq. 5, and pre-activation response ordering on all released addresses in Eq. 6, zero post-training permutation), whereas initial Tables 1, 3, 25, and 28 mistakenly reported an early Appendix A.5-A.8 prototype (residual learning + zero-response ordering + post-hoc training-data assignment).

In the revision, all main-body tables (Tables 1-5) commit to the single Section 2 native method (r = 0 on single-step shifts, zero post-hoc assignment), while Appendix A.5-A.8 and Appendix F.2 (Tables 30-31) keep the residual prototype as a labeled ablation. Answering Q1, both variants independently outperform all 13 baselines across synthetic and LLM tasks:

Method / VariantCode P-C / C-C JUpdate P-C / C-C JSynth C-C JToy C-C J
LineageSAE (residual ablation, App. F.2)0.1176 / 0.34940.6163 / 0.69590.4622 (0.629 3-seed)0.5260 (0.758 3-seed)
LineageSAE (native Sec. 2, Tables 1-3)0.1054 / 0.47660.5804 / 0.80130.8700.853
Stage-Wise (Parent copy + continue)0.0807 / 0.19280.5240 / 0.68720.4640.601
Stable Pool0.0818 / 0.21670.3902 / 0.67460.55660.4229
Best other baseline (native / matched)0.0138 / 0.00440.3401 / 0.01240.014 / 0.6070.036 / 0.483

2. Unchanged parent SAE control, Parent-copy continuation, and variance (W2, Q2)

Answering Q2 on Tables 1 and 3: initial Table 25 (Appendix Table 28) retrained on fixed static Wiki data with zero shift. Under shifts, LineageSAE beats Stage-Wise on Table 1 Wiki-to-Code (C-C J: 0.4766 vs. 0.1928; C-C cosine: 0.6905 vs. 0.3453) and Table 3 Pythia update (C-C J: 0.8013 vs. 0.6872; NMSE: 0.0154 vs. 0.0162; CE: 95.13% vs. 94.60%, a +0.53 pp gain vs. +/-0.01% 3-seed SD in Table 23), and cuts causal logit-shift error by 42% (Table 4, Figure 5: 0.1606 vs. 0.2782; silent 1.55% vs. 7.01%).

For the zero-shot unchanged parent control: on Wiki-to-Code (Section 3.1, Appendix H.2), zero-shot CE recovery on Code drops from 97.99% to 43.04%, as 1,433 / 4,096 (35.0%) Wiki features never fire on Code and 538 (13.1%) have negative marginal utility. On model updates, consistent with your GemmaScope point, a mild 300-step Pythia update induces a moderate shift (frozen parent old-vs-new J = 0.8318, steering cosine 0.9424), so Eq. 4 releases only 48 / 4,096 (1.2%) slots (Table 21). On larger shifts like Pythia 100k-to-143k (Table 43), unchanged-parent NMSE degrades 2.46x (0.0206 to 0.0507), and on step-900 unseen Wiki articles adaptation lowers NMSE from 0.03584 to 0.03390.

3. End-to-end benefits of corrected marginal utility (W3, Q3; revised Table 5, Appendix Table 44)

Uncorrected covariance had higher raw J (0.885 vs. 0.866) in Table 5 because it misses 116 / 153 deletions (37 / 153 vs. 130 / 153) and leaves dead slots frozen as inherited, causing correlated false-positive activations on neighbors that inflate raw overlap.

Corrected utility gives three end-to-end gains (Section 3.3, Table 5; Appendix H.4, Table 44): (1) higher active-set TPR (82.37% to 86.63%) with half the false-positive firing rate FPR (0.223% to 0.113%), better rare recovery (125 / 204 vs. 120 / 204), and lower NMSE (0.1075 vs. 0.1077); (2) an accurate retirement mask (85.0% vs. 24.2% obsolete detection); and (3) under matched-count controls (Table 44), TopK activation count falsely deletes 64 / 102 (62.7%) rare survivors vs. 3 / 102 (2.9%) for corrected utility (mean TopK amplitude catches only 26 / 153 deletions).

4. Qualitative case studies on LLM shifts (W4; Appendix H.1 Figure 9, Appendix H.3 Table 42)

We added steering transcripts (Appendix H.1, Figure 9f-g) and Python AST tracking (Appendix H.3, Table 42). On updated Pythia-1.4B, removing feature 1188 (QPR vs. Birmingham) drops clean P(" of") from 67.33% to 1.29% (parent) and 1.45% / 1.46% in both LineageSAE children at native index 1188 with identical 32-token continuations (native Stable Pool stays at 67.33%). Removing feature 3182 (St Ceidio's) drops P(" the") from 25.09% to 3.91% (parent) and 4.07% / 4.25% (LineageSAE), vs. 16.65% / 17.10% (Stage-Wise) and 20.47% / 20.51% (matched Stable Pool). On Python updates (Table 42), inherited "with" / "async with" features retain 0.766 / 0.868 causal cosine (6.94% silent vs. 43.40% Stage-Wise), while released slots learn new comprehension detectors at 0.680 AUPRC (+0.266 over parent).

Thanks for the additional results, they represent substantial progress. Residual concerns:

  • I think Figure 2 still uses the residual learning method: E.3 uses the A.5 config for Table 25 results, which Fig 2 summarise.
  • The reported 74.9% SynthSAEBench rare-feature recovery result in the main paper doesn't seem to match the reported results in Table 17.
  • I'm still not sure why the corrected utility is preferred to the uncorrected utility. The differences in some of those metrics are small and uncertainty estimates aren't provided.
  • I'd like a more detailed example of a new feature that walks me through the parent address, reason for release, child feature examples, and a comparison with what ordinary continuation learns.

I'll maintain my score for now as I'm confused by the inconsistencies between the appendix and main body, but will bump it up if those can be addressed.

Thank you for reading the revised paper so closely. We updated Figure 2, Appendix E.3, and Appendix Table 17 in the revised PDF, and address all four points below.

1. Figure 2 and Appendix E.3 / Table 25 (four-generation Wiki trajectory)

In the last upload, Tables 1-5 and Figure 3 used the Section 2 method, whereas Figure 2 and Appendix E.3 (Tables 25-27) still plotted the earlier Appendix A.5 run. We updated Figure 2, Section 3.2, and Appendix E.3 to the Section 2 trajectory (3 chains, 12 children, zero post-hoc assignment; the A.5 run is kept alongside Table 28 controls):

GenAdj. J / CosRoot J / CosC-C J / CosNMSECERel.
G10.6590 / 0.96830.6590 / 0.96830.8474 / 0.99760.0153995.83%46.0
G20.7479 / 0.98680.5686 / 0.94130.8092 / 0.99420.0153995.83%55.7
G30.7728 / 0.99180.5199 / 0.92200.7821 / 0.99100.0154395.82%52.0
G40.7411 / 0.99240.4790 / 0.90570.7175 / 0.98720.0155395.74%52.7

Local parent-child J stays high (0.6590 to 0.7411) while root-relative J drifts (0.6590 to 0.4790) at 95.74%-95.83% CE recovery.

2. Match between 74.9% SynthSAEBench rare recovery and Appendix Table 17

Appendix Table 17 still held the pre-revision counts (0/153, 89/153, 153/153 = 242/459 = 52.7%). Updated to the Table 2 and Figure 3 checkpoints, Table 17 matches the main paper:

  • SynthSAEBench: Rare recovery across the three reduced-frequency conditions is 54/153, 137/153, and 153/153, totaling 344/459 = 74.9% (Figure 3), with 100% new and ordinary recovery (918/918 and 7839/7839) and mean NMSE 0.111 (Table 2).
  • Toy (roots 101-103): Rare, new, and ordinary recovery are 54/72 (75.0%), 65/72 (90.3%), and 876/936 (93.6%), with mean NMSE 0.097 (Table 2).

3. Why corrected utility is preferred and uncertainty estimates (Table 5, Appendix D.4, Table 44)

We agree that NMSE (0.1075 vs. 0.1077) and rare counts (125/204 vs. 120/204; diff +2.45%, 95% CI [-1.96%, +7.35%]) are close, and uncorrected covariance gives higher raw J (0.885 vs. 0.866). Corrected utility is preferred for two reasons (20k-draw paired feature bootstrap 95% CIs in Appendix D.4):

  1. Retiring dead features and cutting false positives: Because surviving features correlate with deleted directions, uncorrected covariance releases only 37/153 (24.2%) deleted features, vs. 130/153 (85.0%, +60.8 pp; 5/5 vs. 0/5 correlated deletions in Fig. 4a) for corrected utility. Leaving 116/153 dead features marked as inherited keeps monitors on obsolete slots and raises false firing on deleted slots (-0.155 pp, 95% CI [-0.255, -0.061] pp). Across the 102 rare survivors, corrected utility also raises TPR by +4.27 pp (86.63% vs. 82.37%, 95% CI [+1.02, +7.98] pp) and halves FPR (-0.110 pp, 0.113% vs. 0.223%, 95% CI [-0.211, -0.022] pp).
  2. Saturated LLM dictionaries: SynthSAEBench (512 features in 1,024 slots) leaves about 500 unused slots even if deleted slots stay frozen. On Pythia Wiki-to-Code (N = 4096), 99.4% (4,073/4,096) of parent slots fire on Wiki, and negative corrected utility frees 538 firing-but-obsolete Wiki slots (13.1% of width) plus 1,433 zero-firing slots.

4. Walkthrough of a newly learned LLM feature vs. ordinary continuation (Appendix H.3)

In the 1,000-step Pythia CodeSearchNet update (where yield statements are removed and comprehensions enter at 20%), we trace native slot 1768 across both child seeds:

  • Parent slot and release reason: At step 0, parent slot 1768 had no comprehension role (AUPRC 0.243 vs. 0.225 prevalence) and had zero positive TopK activations across all 32,768 step-1k scoring tokens, so Eq. 4 releases slot 1768 in both seeds.
  • Child feature at slot 1768: Guided by pre-ReLU ordering (Eq. 6), both LineageSAE seeds converge at slot 1768 to a detector for the token "for" in Python comprehensions, such as "[Asset(o) for o in resp]" (12.27 and 11.94 in seeds 1 and 2) and "[(rps, len(list(rpl))) for rps, rpl in rps_groups]" (11.42 and 12.41; C-C steering cosine 0.838). Across 569 held-out files, slot 1768 fires on comprehension "for" tokens but never on ordinary for-loops:
Model at Slot 1768AUPRCComp. "for" (all / 50 shared)Loop "for"
Parent (0 / 1k)0.243 / 0.2343/188 / 1/72 (1k: 2/188)0/262
LineageSAE Sec. 2 (s1 / s2)0.822 / 0.829135/188, 136/188 (47/72, 48/72)0/262
Stage-Wise (s1 / s2)0.225 / 0.2250/188, 0/188 (0/72, 0/72)0/262
  • Comparison with Stage-Wise: Without slot release, slot 1768 stays dead in both Stage-Wise seeds (0/188 firings, AUPRC 0.225), along with 38.9% of the dictionary (C-C J 0.376 vs. 0.716). Instead of reusing dead slots, Stage-Wise adapts active parent slots (e.g., slot 1597 shifts from parent AUPRC 0.331 to 0.940/0.942), which also overwrites suppressed with/async-with features (35.1%-43.4% silent vs. 4.9%-6.9% in LineageSAE, Table 42).
Reviewer_b5zwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes a method to adapt an SAE to domain shifts and model updates while preserving feature correspondence. It uses a pretrained parent SAE on the original data and then trains a successor SAE using guidance from the parent SAE’s weights.

Critical strengths
  • The problem is well motivated. I can see why we need to adapt an SAE when tools already depend on its features.

  • The experiments cover both synthetic and real data.

Critical weaknesses

What can be fixed during rebuttal

  • Domain adaptation of an existing SAE has already been explored in SAE Boost [1], and I think this comparison is missing. It should also be discussed in the related work or introduction, with the additional contributions of LineageSAE made clear.

  • I think simple fine-tuning initialized from the parent SAE and SAE Boost [1] should be added to the main comparison tables. These baselines should be evaluated under the same domain-shift and model-update settings as LineageSAE.

  • The experiments with LMs cover only one domain shift, Wiki-to-Code, and I do not think this is enough. Adding additional domain shifts to other specialized datasets would be beneficial.

More fundamental concerns

Motivation

  • I can see some motivation for retraining the SAE when the underlying LM changes, since its representations may change as well. However, I am not convinced by the motivation for changing the dictionary within a limited capacity when only the data distribution changes. From my perspective, a new domain may bring new features to the SAE. In the Wiki-to-Code experiment, the underlying LM remains unchanged, so I do not see why one should replace existing Wiki features instead of keeping the parent SAE and learning additional Code features. I would like to see whether the released features are still useful on Wiki and how this compares with expanding the dictionary.

  • I cannot clearly see how changing feature directions would work in practice. Imagine one deployed an LM with some steering enabled, then retrained the SAE. Could the steering vector be reliably replaced with the corresponding one from the new SAE? I have the same question about a classifier trained on SAE features.

  • I strongly suggest reporting downstream metrics, for example, sparse probing on code for the parent and successor SAEs. I would particularly like to see whether a probe or steering intervention built using the parent SAE still works with the successor.

Method feasibility

  • Why would one use this method if simple fine-tuning initialized from the parent SAE (the "Parent copy + continue" baseline) performs very similarly in repeated retraining on fixed Wiki data, as shown in Table 25? I think this baseline should also be included in the domain-shift and model-update experiments to show when the additional complexity is necessary.

  • The frozen parent decoder performs better on continuity metrics in Table 25. This makes me question whether the reported metrics are enough to claim that one method is more useful than another.

[1] Koriagin et al. Teach Old SAEs New Domain Tricks with Boosting. COLM 2025.

Questions for authors

See Weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer b5zw (R2)

Thank you for the constructive and specific review. Below we address each point on SAE Boost, fixed-capacity adaptation, additional shift benchmarks, downstream steering and probing, and the static-Wiki baselines.

1. SAE Boost, fixed capacity versus expansion, and released Wiki features

We added a discussion of SAE Boost (Koriagin et al., COLM 2025) in Sections 1 and 4, added Stage-Wise (Parent copy + continue) to all main tables (Tables 1-4), and evaluated a fixed-width SAE Boost-style residual bank on Wiki-to-Code in Appendix H.2 (Table 38, Figure 6b):

Wiki-to-Code Variant (N = 4096, k = 16)P-C JC-C JP-C / C-C CosNMSECE Rec.
0-Shot Unchanged Parent--------43.04%
Frozen Inherited + Residual Bank0.14050.62370.7251 / 0.97340.0282380.00%
Stage-Wise (Parent copy + continue)0.08070.19280.1818 / 0.34530.0140095.16%
LineageSAE (adaptive handoff)0.10540.47660.2401 / 0.69050.0150694.25%

When the base LM is fixed and budgets allow growing width, freezing the parent SAE and appending residual atoms (SAE Boost) naturally keeps all source features. Fixed-width versioning targets the complementary setting where (i) the LM itself updates so frozen parent directions rotate out of alignment, (ii) multi-stage updates (Tables 40-42) would otherwise grow width linearly at every step, or (iii) deployment fixes memory and sparsity budgets and only needs active target-domain features in the current version (while the saved parent checkpoint retains historical source features).

On Wiki-to-Code (262k tokens, Appendix H.2), 4,073 / 4,096 (99.4%) parent features fire on Wiki, whereas on Code 1,433 / 4,096 (35.0%) never fire once and 538 (13.1%) have negative marginal utility. Strictly freezing the 2,125 inherited rows and training only the 1,971 released slots on parent residuals worsens NMSE by 87.43% (0.01506 to 0.02823) and lowers CE recovery from 94.25% to 80.00%, showing that surviving features also need gentle decoder handoff rather than hard freezing.

2. Additional domain and checkpoint shifts beyond Wiki-to-Code (Appendix H.3, Tables 40-43, Figure 6)

(1) 3-stage Pythia-1.4B Wiki updates (300, 600, 900 steps) with 10x suppression (128 features), exclusion (128 features), and unseen Wiki articles at step 900 (Tables 40-41, Figure 13): LineageSAE leads step-900 P-C / C-C J (0.4862 / 0.7005 vs. Stage-Wise 0.3968 / 0.5615; CE 93.22% vs. 92.24%) and reduced-exposure C-C J (0.5512 vs. 0.2945). (2) 4-stage CodeSearchNet Python AST shift (1k to 4k steps) tracking syntax suppression, restoration, and new comprehensions (Table 42, Figure 6c): LineageSAE leads C-C J at every stage (0.716 to 0.426 vs. 0.376 to 0.273). (3) Official Pythia-1.4B (100k to 143k) and Qwen2.5-0.5B (Base to Instruct) transitions (Table 43, Figure 6d), where unchanged-parent NMSE degrades 2.46x (0.0206 to 0.0507): LineageSAE leads C-C J (0.7571 and 0.515 vs. 0.5613 and 0.309) and active slots (4,084 and 3,705 vs. 3,030 and 1,732), while Stage-Wise retains higher P-C J and CE recovery (93.46% / 93.99% vs. 88.17% / 92.43%) on large checkpoint jumps.

3. Downstream steering and sparse probing on code (Section 3.2, Table 4, Figure 5; Table 42)

For any inherited slot, a deployed probe or steering hook keeps the same index i and swaps in the child activation and adapted decoder vector without retraining or remapping. On updated Pythia-1.4B across 63 inherited features (1,941 positions; Table 4, Figure 5, Table 37), native-index reuse in LineageSAE retains 0.9208 logit-shift cosine (1.55% silent), 76.3% / 77.6% exact 32-token continuations, and 70.8% AxBench directional agreement (0.141 MAE), ahead of Stage-Wise (0.8321, 7.01% silent, 61.5% / 62.3%, 56.2%) and matched Stable Pool (0.5809, 29.21% silent). On Python 1k-to-4k updates (Table 42, Figure 6c), fixed-index probes on 9 "with" / "async with" AST features retain 0.533 to 0.557 AUPRC, 4.86% to 6.94% silent rate, and 0.766 / 0.868 intervention cosine (vs. Stage-Wise 0.431 to 0.460, 35.07% to 43.40% silent, 0.392 / 0.269 cosine), while released slots learn new comprehension probes at 0.680 AUPRC (+0.266 over parent).

4. Static-Wiki initial Table 25 (revised Table 28): Parent copy and Frozen decoder

We agree: initial Table 25 (now Table 28) retrained on static Wiki data with an unchanged LM. With zero shift, freezing the parent decoder trivially gives cosine 1.0 without testing adaptation, and Parent copy + continue needs no slot release. Under real shifts (Tables 1-4, Table 38), freezing drops Code CE recovery to 43.04% or 80.00%, and Stage-Wise trails LineageSAE across C-C J and steering.

Reviewer_DLcQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper describes a method for doing 'version control' of features in SAEs. That is, keeping track of features between SAE training runs. This includes renumbering and semantic similarity. Consider the following:

Feature #1 detects Python function definitions, and you build a monitor around it. You retrain the SAE because you updated the model or the data changed.

  • In the new SAE The feature #1 appears at #2.
  • #1 changes or splits into several features.
  • It disappears, and #1 now detects something unrelated.

This lets you track that happening

Critical strengths

None

Critical weaknesses

The paper shows some lift over baselines: L2-regularized TopK, Ordered SAE (OSAE), Identifiable SAE (iSAE), archetypal constraints, joint training (SharedSAE), and feature pooling (Stable Pool). They did not compare against matching features after SAEs are trained

It may be useful to be able to match Their experiments show more consistent numbering and activation patterns, but nothing about not the more important kind of continuity across training.

I have never felt that this was an issue that could not be easily overcome by matching features based on activation correlations or through normal continued training of an SAE. I have not typically seen feature drift from training a single SAE, and activation correaltions seem to align SAE features from separate models well enough. It isn't clear to me that the paper addresses an important or impactful problem, and does not really deeply engage with problems the field is focused on right now.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Response to Reviewer DLcQ (R1)

Thank you for the candid review. Your three main questions are whether post-training feature matching already solves alignment, whether native numbering gives functional continuity across training, and whether ordinary continued training from a parent SAE is already enough. Below we address each point with new experiments in the revised paper.

1. Comparison against matching features after SAEs are trained

We agree that post-training feature matching is a natural baseline, and we now compare native-index LineageSAE directly against three post-training matching families across synthetic and LLM tasks.

Under held-out activation-correlation (Pearson-Hungarian) matching on SynthSAEBench and Toy superposition (Section 3.1, Table 2), post-hoc matching lifts independent L2-TopK from native J = 0.006 / 0.013 to matched J = 0.453 / 0.483, and the strongest matched baseline (MP-SAE) reaches 0.607 on SynthSAEBench. LineageSAE reaches native J = 0.870 / 0.853 directly at raw indices without matching (0.876 / 0.853 if matched). Even after 120k training steps on fixed Wiki data (Appendix H.2, Table 39), post-hoc activation matching on independent L2-TopK reaches at most matched J = 0.461 with 0.00% of pairs at J >= 0.8, whereas LineageSAE reaches native J = 0.860 with 83.36% of indices at J >= 0.8.

We also evaluate Semantic Optimal Transport and calibration decoder matching across all four tasks (Appendix F.6-F.7, Tables 35-36). On Wiki-to-Code and Pythia updates, Semantic OT yields mapped P-C / C-C J of 0.0200 / 0.3160 and 0.3545 / 0.3546, well below native LineageSAE (0.1054 / 0.4766 and 0.5804 / 0.8013). Independent overcomplete SAEs split and absorb features differently across runs, so post-hoc bijection matching neither recovers high support overlap nor separates surviving features from obsolete ones that should be retired.

2. Functional and causal continuity beyond consistent numbering

We agree that consistent numbering only matters if an index keeps its downstream causal effect and monitor behavior after retraining. We added direct intervention and probing benchmarks modeled on your Python-monitor example (Section 3.2, Table 4, Figure 5; Appendix H.1, Table 37, Figures 8-10; Appendix H.3, Table 42):

MethodSynth J (nat / match)120k J >= 0.8Pythia Logit CosSilent (%)32-tok ExactAxBench Dir.
LineageSAE (native)0.870 / 0.87683.36%0.92081.55%76.3% / 77.6%70.8%
Stage-Wise (native)0.464 / 0.47845.78%0.83217.01%61.5% / 62.3%56.2%
Stable Pool (old match)0.557 / 0.56516.63%0.580929.21%43.5% / 43.2%40.3%
L2-TopK (old match)0.006 / 0.4530.00%0.034694.51%----

Without post-hoc matching, all 11 independent baselines collapse to < 0.001 logit-shift cosine and > 99% silent interventions on updated Pythia-1.4B (Table 37; even with oracle updated-model matching, L2-TopK reaches only 0.6020). Across a 4-stage Python AST shift on CodeSearchNet (1k to 4k steps, Table 42, Figure 6c), keeping fixed-index sparse probes and interventions on "with" and "async with" features preserves 0.557 detection AUPRC, 6.94% silent rate, and 0.766 / 0.868 P-C / C-C intervention cosine in LineageSAE, vs. 0.460 AUPRC, 43.40% silent rate, and 0.392 / 0.269 cosine under continued fine-tuning.

3. Why normal continued training (Stage-Wise) is not enough

On a single static corpus with an unchanged model, continued training from the parent SAE (Parent copy + continue in Appendix Table 28) indeed shows minimal drift. However, evaluating continued training (Stage-Wise) across all shifted benchmarks (Tables 1-4, Figures 5-6, Appendix H.2-H.3) shows three clear limitations once the data or model changes.

First, without utility-based slot partitioning or parent-guided handoff, rare and temporarily suppressed features get overwritten: Stage-Wise produces 4.5x more silent causal interventions on updated Pythia (Table 4: 7.01% vs. 1.55%), 6.3x more silent AST interventions after temporary 10x syntax suppression (Table 42: 43.40% vs. 6.94%), and lower rare-survivor recovery on SynthSAEBench and Toy (Figure 3: 51.6% / 40.3% vs. 74.9% / 75.0%).

Second, drift accumulates across multi-stage updates: over 300-to-900-step Wiki updates (Tables 40-41, Figure 13), Stage-Wise C-C J falls to 0.5615 overall and 0.2945 on reduced-exposure features, vs. 0.7005 and 0.5512 for LineageSAE.

Third, when new features emerge, independent continued-training runs place them in uncoordinated slots, giving C-C J of only 0.1928 on Wiki-to-Code (Table 1, vs. 0.4766) and 0.464 on SynthSAEBench (Table 2, vs. 0.870).

Thank you for your responses and siginificant rebuttals. I think that the feedback is not enough to swing the decision of this paper and I will maintain my score

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response: Sincere Clarification of Initial Draft Discrepancy, Unified Method Evaluation, and New Empirical Evidence

Thank you to the Area Chair and Reviewers DLcQ, b5zw, and AfR2 for the detailed feedback. We have uploaded a revised manuscript (10-page main text, revisions in teal-blue) with unified main-body evaluations, new figures (Figures 5-6; Appendix Figures 8-13), and extended experiments in Appendix H (Tables 37-44).

1. Correction of the Table 1 and Table 3 inconsistency (Reviewer AfR2 W1, Q1)

Reviewer AfR2 caught a discrepancy between Section 2 and Appendix A.5-A.8 in our initial submission. Section 2 describes our native-index method: corrected marginal-utility release (Eqs. 3-4), parent-guided decoder handoff (Eq. 5), and pre-activation response ordering across all released addresses (Eq. 6), with zero post-training permutation. Appendix A.5-A.8 recorded an earlier prototype combining residual learning with post-hoc training-data assignment.

During submission preparation, Tables 2 and 4 used Section 2, whereas Tables 1, 3, 25, and 28 mistakenly reported runs from that Appendix A.5-A.8 prototype. That was our mistake, and we apologize for the confusion. All main-body tables (Tables 1-5) now use the single Section 2 native method (r = 0 on single-step shifts, zero post-hoc assignment). Appendix A.5-A.8 and Appendix F.2 (Tables 30-31) retain the residual prototype as a labeled ablation and confirm that both variants independently beat all 13 baselines across synthetic and LLM tasks.

2. Unified main-body results with Stage-Wise, matching, and unchanged parent (Tables 1-3)

Task and MethodP-C JC-C JC-C rhoP-C / C-C CosNMSECE Rec.
Code (Table 1): 0-Shot Parent----------43.04%
Code: Stable Pool0.08180.21670.28780.2330 / 0.43210.014295.25%
Code: Stage-Wise0.08070.19280.24760.1818 / 0.34530.014095.16%
Code: LineageSAE (Sec. 2)0.10540.47660.59420.2401 / 0.69050.015194.25%
Update (Table 3): Stable Pool0.39020.67460.79520.7565 / 0.95390.014795.57%
Update: Stage-Wise0.52400.68720.80040.8267 / 0.93950.016294.60%
Update: LineageSAE (Sec. 2)0.58040.80130.88410.8892 / 0.98030.015495.13%

On SynthSAEBench and Toy superposition (Table 2, Figure 3), LineageSAE reaches native C-C J of 0.870 and 0.853 without matching (0.876 and 0.853 if matched; rare recovery 74.9% and 75.0%), ahead of Stage-Wise (0.464 and 0.601; rare 51.6% and 40.3%), Stable Pool (0.5566 and 0.4229), and the best post-hoc matched baselines (MP-SAE 0.607 on SynthSAEBench; L2-TopK 0.453 and 0.483).

3. Summary of new downstream, capacity, and multi-stage evaluations

Downstream causal steering on updated Pythia-1.4B across 63 inherited features (1,941 positions; Section 3.2, Table 4, Figure 5; Appendix H.1, Table 37, Figures 8-10):

Method (Pythia-1.4B Update)Logit CosRel. Err.Silent (%)32-tok ExactAxBench MAE / Dir.
LineageSAE (native)0.92080.16061.55%76.3% / 77.6%0.141 / 70.8%
Stage-Wise (native)0.83210.27827.01%61.5% / 62.3%0.214 / 56.2%
Stable Pool (old match)0.58090.629829.21%43.5% / 43.2%0.210 / 40.3%
L2-TopK (old match)0.03460.989494.51%----

Fixed-width residual bank and 120k training (Appendix H.2, Tables 38-39, Figure 6a-b, Figures 11-12): At fixed width (N = 4096, k = 16) on Wiki-to-Code, freezing the 2,125 inherited parent rows and adapting only the 1,971 released slots as an SAE Boost-style residual bank worsens NMSE by 87.43% (0.01506 to 0.02823) and drops CE recovery from 94.25% to 80.00%. Over 120k Wiki steps (Table 39), LineageSAE retains 83.36% of indices at J >= 0.8 (mean J = 0.860) vs. 45.78% for Stage-Wise (0.765) and <= 0.461 matched J for L2-TopK.

Multi-stage shifts, Code AST probing, and official checkpoints (Appendix H.3, Tables 40-43, Figure 6c-d, Figure 13): Across 300-to-900-step Wiki updates and 4-stage Python AST shifts, LineageSAE preserves higher step-900 C-C J (0.7005 vs. 0.5615) and step-4k AST probe AUPRC / steering cosine (0.557 and 0.868 with 6.94% silent vs. 0.460 and 0.269 with 43.40% silent for Stage-Wise), while learning new comprehension probes in released slots (0.680 AUPRC). On official Pythia 100k-to-143k and Qwen2.5 Base-to-Instruct checkpoints (Table 43), LineageSAE leads C-C J (0.7571 and 0.515 vs. 0.5613 and 0.309), while Stage-Wise retains higher P-C J and CE recovery on large checkpoint jumps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.