acceptodds
Under review as a conference paper at ICLR 2027

Certified Cross-Dictionary Feature Matching via Co-Activation Geometry

Abstract

Feature matching across dictionaries is central to tracking concepts across layers, models, and fine-tunes, but common matchers return correspondences without a certificate. For non-negative codes evaluated on shared samples, the cross co-activation relation defines a filtered relation whose two feature-side Dowker complexes have identical persistent homology for every pair of codes, even when widths and ambient activation spaces differ. Restricting the induced maps to mutual matches yields a certified partial correspondence with zero round-trip error; its metric distortion is reported together with coverage. Across twelve synthetic width–sparsity configurations, every committed match is correct (precision ) and ; additional samples increase coverage with an approximately inverse-linear law in samples per feature. Released GPT-2 dictionaries occupy a different regime: max–min scores saturate with sample count, so coverage must be calibrated empirically rather than extrapolated from the synthetic law. At 8192 tokens, adjacent-layer coverage falls from at layers to a minimum of at , and agreement with correlation and decoder-cosine matchers falls with it from onward; the embedding transition is the one place where the relational matcher commits most and the baselines agree with it least. The method converts feature matching from an unqualified assignment into an auditable correspondence whose precision, coverage, and geometric distortion are explicit.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, feature matching, representation alignment, cross-layer analysis, mechanistic interpretability, co-activation geometry, Dowker duality, GPT-2

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_juE1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the feature matching problem between sparse autoencoder dictionaries of different layers or models. Based on non-negative activations on shared samples, it constructs a max-min common activation relationship. By leveraging Dowker duality, it establishes the equivalence of the continuous homology on both sides of this relationship. Through bidirectional mutual selection, it obtains partial correspondences with zero round-trip error. It jointly evaluates the matching results using coverage rate and geometric distortion. In terms of validation, the authors conducted synthetic transformation experiments such as permutation, scaling, noise addition, splitting, and merging. In twelve groups of precise permutation tasks with different widths and sparsity configurations, the accuracy of retained matches reached 100% and geometric distortion was zero! Further, using the public dictionaries of GPT-2 and Gemma-2-2B, cross-layer experiments were conducted and compared with methods such as activation correlation and decoder cosine to reveal the sample number saturation phenomenon in the real dictionary and the differences in layer-wise matching coverage as the model and depth change.

Critical strengths
  1. The method directly utilizes the co-activation relations on the shared samples, without requiring the two dictionaries to have the same width or share the activation coordinate space, nor does it need to retrain the joint dictionary.
  2. The bidirectional mutual selection allows the method to retain only the feature pairs that mutually support, and present the matching range through coverage.
Critical weaknesses
  1. There is no main figure, making it difficult to understand the author's intention, and the format is too messy, especially the "Contributions." and its contents are even divided into two pages on page 1.
  2. Can the Abstract contain a formula?
  3. I'm not entirely convinced that the experimental results are 100% accurate.
  4. The round-trip error on the mutually selected set is zero, which is guaranteed by the definition. Even smaller geometric distortion cannot rule out semantic errors, because the corresponding errors in the symmetrical structure may still maintain a distance.
  5. The 100% accuracy rate emphasized in the paper only appears in the precise permutation experiments. The noisy experiment in Table 1 has dropped to 82.6%, so it cannot be generalized as a universal correctness guarantee.
  6. On page 2, Eq. 2 can provide an effective upper bound, but the round-trip displacement condition used in the proof is overly restrictive and cannot lead to the conclusion of minimality.
  7. Theorem 2 mainly represents a direct application of the existing Dowker duality theory. Its conclusion is also valid for unrelated dictionaries, and it cannot be used to prove independently that two dictionaries have similar semantic structures.
Questions for authors

Please see "Critical Weaknesses".

3 comments
AuthorsOpenReview ↗

response for Reviewer juE1

We thank the reviewer for a careful reading. In particular, the point about the minimality of Eq. 2 was correct, and fixing it gave a stronger result.

Point 1 (main figure and Contributions). Figure 1 (new) shows the pipeline on a toy example: shared samples, the cross relation, the induced maps and the mutual set, the certificate and per-pair slack, and the permutation test that drops a mutual but chance-level pair. The Contributions are now a compact paragraph that does not break across pages.

Point 2 (formula in the abstract). Removed. The abstract is now formula-free.

Point 3 (precision for every transformation). The revision reports precision against ground truth for every transformation instead of a blanket statement, and the revised matcher is exact in all but one of them (Table 1). Precision is 1.000 under permutation, rescaling, noise σ=0.01\sigma=0.01–0.100.10, splitting and merging, and 0.988 under absorption. Between unrelated dictionaries, 1.4% of features are committed (9 of 640), within sampling error of the nominal α=1%\alpha=1\% (binomial tail 0.19). To make this directly verifiable, the supplementary material contains a script that generates benchmark A, runs the matcher and the test over the ten seeds, and prints Table 1 as it appears in the paper, in about two minutes on a CPU.

Point 4 (zero round-trip error, and wrong matches with small distortion). Both points are now explicit in the paper, and §4.1 turns the second into an experiment. On dictionaries whose co-activation geometry has a symmetry (features on a regular grid of the torus), mapping every feature to the partner of its translate is entirely wrong. This map has δmut\delta_{\mathrm{mut}} = 0.173, close to the true correspondence's 0.113, and its mean slack is indistinguishable from the truth's. It is rejected by the shared-sample mutual maximum, not by the geometry: the calibrated matcher recovers the truth with precision 1.000, while Gromov–Wasserstein, which sees only the two geometries, reaches 0.028. The paper now states what each part certifies. The certificate measures geometric consistency, and identity comes from the shared samples, through the mutual rule and the permutation test.

Point 5 (precision under noise). This was the most important experimental point, and it led to two changes in the matcher. Matching is now by mean co-activation (§3.1), because max–min scores tie exactly through saturation and are decided by a single sample, which is what failed under noise. A per-pair permutation test (§3.2) is added, valid at level α\alpha for each committed pair (Proposition 4). Together they give precision 1.000 and coverage 1.000 at every noise level from σ=0.01\sigma=0.01 to 0.100.10 (Table 1). The submitted rule had precision 0.826 and coverage 0.527 at σ=0.05\sigma=0.05.

Point 6 (the round-trip term of Eq. 2). The reviewer is right. The round-trip term should have coefficient 1/41/4 and be taken in pairwise form, 14max⁡a,a′[d1(a,νμa′)−d1(a,a′)]+\frac{1}{4}\max _ {a,a'} [d_1(a, \nu\mu a') - d_1(a,a')] _ {+}. The new Proposition 3 gives the exact least δ\delta, necessary and sufficient, and its proof needs no triangle inequality. That matters here, because the co-activation dissimilarity violates it in 42% of code-derived instances. The formula agrees with a direct search over δ\delta to within 5×10−135 \times 10^{-13} on 700 instances. The old bound was strictly loose in 99.5% of metric instances (Appendix A). On the mutual set both round-trip terms vanish, so δmut\delta _ {\mathrm{mut}} and the results that use it are unchanged.

Point 7 (the role of Theorem 2). The revision states the theorem's role precisely (§2). It is the filtered Dowker duality of Chowdhury and Mémoli applied to the cross relation. Its role is well-posedness: the comparison does not depend on which dictionary is the source. The evidence of correspondence comes from the mutual rule and the permutation test, which commit to 1.4% of features between unrelated dictionaries, and from the intervention tests on GPT-2 (§5.1). The paper's own theoretical results are the exact certificate (Proposition 3) and the per-pair validity of the test (Proposition 4).

We are glad to answer any further questions during the discussion.

To be honest, Figure 1 is not the main figure. It is recommended to refer to the top-level conference papers. I maintain my rating.

AuthorsOpenReview ↗

Redrawn Figure 1

Thank you for the follow-up. We have redrawn Figure 1 as an overview of the method, on page 2 of the updated PDF. From left to right it shows the two dictionaries evaluated on the same inputs, the three checks applied to every candidate pair, and the output. The checks are the mutual best partners under mean co-activation, the exact certificate that compares the two within-dictionary Vietoris–Rips complexes through the matching (Proposition 3), and the sample-permutation test (Proposition 4). The output is the committed set, with the partner, slack and p-value of each pair. One toy example runs through all panels and the caption follows it: a feature whose partner prefers another feature is not matched, and a mutual pair at chance level is dropped by the test.

We would be grateful to know whether any concern remains on the other points of the review: the exactness of the certificate (Proposition 3, checked in Appendix A and by a script in the supplementary material), precision of 1.000 under every transformation except absorption (0.988) in Table 1, which the supplementary material reproduces from scratch, the symmetric case in which small distortion does not imply correctness (Section 4.1), the intervention tests on GPT-2 (Section 5.1), and the role of Theorem 2. If these points are now resolved, we hope the assessment can reflect them.

We are glad to answer any further questions during the discussion.

Reviewer_hY5mOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses a simple question: given two sets of learned features, such as those from two layers of a large model, can we match features across the two sets, and how many have a clear correspondence? The paper proposes a method based on the maximum overlap between the samples that activate features in each set. What I find most interesting is that this method allows us to measure a “gap”. That is to say, we can use such maximum “similar” sample’s relative distance to give a exact similarity score. This would be a new way to measure the similarity of two models/layers.

Critical strengths

The intuition itself would be most interesting part by my side. The author did raise a good question with a good way.

Critical weaknesses
  1. The paper writing is very poor, I could not understand abstract and the introduction. The introduction is not well written, e.g., the introduction should give exact position of this paper, but we only get three paragraph with no coherent logic flow. In the main body of this paper, a lot of mathematical notations are not defined properly. For the experiments, I cannot fully understand what hypothesis that each experiments want to verify.
  2. The proof is difficult to follow. Although the authors highlight a connection between their theorem and Dowker complexes in the abstract and introduction, it is unclear how this connection informs either the main line of argument or the experiments.
Questions for authors

In L046–051, what does the “width–sparsity sweep” refer to? What does it mean for the similarity scores to saturate and for feature correspondences to become scarcer? Please explain how these observations relate to the paper’s main contributions.

  1. Gemma-2-2B appears in the contribution statement, whereas the preceding discussion focuses on GPT-2. Could the authors introduce both models together and explain how their results support the paper’s claims?
  2. The sets (J_1) and (J_2) do not appear to be explicitly defined before their first use around L074. Please define them when introducing the cross relation.
  3. What role does persistent homology play in Theorem 2? How does the equality of persistent homology between the two constructions help establish or evaluate feature correspondences? The practical implications of this result need further explanation.
  4. What exactly are the “mean-normalised Dowker metrics” mentioned in L091? Please provide their mathematical definition and specify the normalization procedure.
  5. What does (VR) denote in L094? Please introduce the Vietoris–Rips construction and explain the scale convention used in the proposition.
  6. In the caption of Table 1, (k) is introduced without a definition. Please define it and explain precisely how “recovery” is computed on the mutual-match set.
3 comments
AuthorsOpenReview ↗

response for Reviewer hY5m

We thank the reviewer for engaging with the core idea and for the precise list of presentation problems. We have rewritten the paper along the lines the reviewer suggests.

W1 (abstract, introduction, notation, hypotheses). The abstract is rewritten without formulas. It states the problem, the method, what is certified and the evidence. The introduction (§1) now follows one line. Matchers must say which feature corresponds to which and for which features that answer holds. Existing matchers (decoder cosine, correlation, crosscoders, SemOT) answer only the first. We construct a partial correspondence with three checks, test each check with ground truth and by intervention, and introduce the models used (GPT-2, Gemma-2-2B and Gemma-2-9B) together. A new Figure 1 shows the pipeline. §2 defines every symbol before use, and each experiment opens with its hypothesis (H1–H5) and ends with whether it holds.

W2 (the proof, and how the Dowker connection is used). The certificate (Proposition 3) is now proved directly on flag complexes. A vertex map is simplicial, and two maps are contiguous, iff a condition holds on every pair of vertices. Each term of the formula is the exact slack of one such condition, which also makes the formula necessary and sufficient. The connection to Dowker complexes is stated where it is used, in two places: the max–min co-activation is the filtration value of the Dowker complex (for a dictionary and for the cross relation), and the within-dictionary dissimilarity (Eq. 2) comes from the same construction inside one code. Appendix A compares the formula with the least δ\delta found by bisection on the defining conditions in 700 instances, and they agree within 5×10−135\times10^{-13}; check_certificate.py in the supplementary material repeats this comparison. For the role of Theorem 2, see Q4.

Q1 (L046–051: "width–sparsity sweep", "saturate", "scarcer"). The width–sparsity sweep is the 12 synthetic configurations, n∈{64,128,256}n\in\{64,128,256\} features and k∈{4,8,16,32}k\in\{4,8,16,32\} active features per sample. "Saturate" refers to K12(a,b)K_{12}(a,b) being a maximum over samples: adding samples can only raise it, and it approaches its cap for many pairs at once, so the best partner becomes less decisive. "Scarcer" means that fewer features have a mutual partner. These were observations about how the matcher behaves on real codes. The revision replaces them with explicit hypotheses and measured quantities (calibrated coverage, intervention tests), and the sentence has been removed from the introduction.

Q2 (GPT-2 and Gemma introduced together). This is done in §1. GPT-2 small (jbloom dictionaries) carries the intervention tests and the depth profile. Gemma-2-2B and Gemma-2-9B (Gemma Scope) carry the unequal-width and cross-model experiments (§5.3).

Q3 (J1J_1, J2J_2 at L074). They are now defined at the start of §2, as the feature index sets of the two dictionaries, before the cross relation is introduced.

Q4 (the role of persistent homology in Theorem 2). Theorem 2 holds for every pair of codes, unrelated ones included, so it is not evidence that two dictionaries correspond. We now say this explicitly. Its role is to make the comparison well posed. The persistent homology of the cross relation is the same whichever dictionary is taken as the source. So any discrepancy between the dictionaries must come from the induced maps and the two within-dictionary geometries, and that is exactly what the certificate measures. Evidence for a particular correspondence comes from the mutual rule and the permutation test (§3.1–3.2), which commit to 1.4% of features between unrelated dictionaries.

Q5 ("mean-normalised Dowker metrics", L091). Eq. 2 now gives the definition: d(a,a′)=(1−K(a,a′)) / meanc≠c′(1−K(c,c′))d(a,a')=(1-K(a,a'))\,/\,\mathrm{mean}_{c\ne c'}(1-K(c,c')) with d(a,a)=0d(a,a)=0, where K(a,a′)=max⁡imin⁡(Z[i,a],Z[i,a′])K(a,a')=\max_i\min(Z[i,a],Z[i,a']) is computed on the max-normalised code. It is renamed a dissimilarity, because it violates the triangle inequality in 42% of code-derived instances.

Q6 (VR at L094). §2 now defines VRt(J,d)\mathrm{VR}_t(J,d) as the flag (Vietoris–Rips) complex of subsets of diameter at most 2t2t, and VR(J,d)\mathrm{VR}(J,d) as the filtration over tt. This is the convention used in Proposition 3.

Q7 (kk and "recovery" in Table 1). kk is the number of active features per sample in the generator (k=6k=6; Appendix C). "Recovery" is replaced by precision: the fraction of committed pairs whose partner is a ground-truth partner. For a split, both halves count as correct; for a merge, the merged feature counts as correct for both sources. Coverage, δmut\delta_{\mathrm{mut}} and slack AUROC are defined in the caption.

We are glad to answer any further questions during the discussion.

Dear authors,

I didn't fully understand the paper when I reviewing this paper, and now I couldn't understand your rebuttal as well. Please stop using AI generated slops, it is wasting everyone's time.

In short, this rebuttal haven't address my concerns, and I'll maintain my score.

Best, Reviewer hY5m

AuthorsOpenReview ↗

Response regarding the revised manuscript

Reviewer hY5m,

Dismissing a paper and technical response with derogatory labels rather than engaging with specific arguments is not constructive peer review.

Your initial review raised presentation and notation questions (W1–W2, Q1–Q7) without identifying a mathematical error or experimental flaw. In the revised manuscript, we rewrote the exposition so that every step can be followed without prior background in algebraic topology:

  • Problem and method in plain terms (W1, Q1, Q2): Existing SAE matchers force every feature in one dictionary to pair with a feature in another, even when features split, merge, or have no counterpart. Our method feeds the same inputs into both models (e.g., across GPT-2 layers or between Gemma-2B and 9B) and commits to a pair only if the two features are mutual best co-activating partners and pass a token-shuffle permutation test against chance. Figure 1 (Page 2) traces this full pipeline step by step.

  • Self-contained notation (Q3, Q5, Q6, Q7): Section 2 defines every symbol from scratch before its first use—including the feature index sets, sample and feature counts, the normalised within-dictionary dissimilarity (Eq. 2), and the Vietoris-Rips neighborhood complex—and the abstract is now completely formula-free. Table 1 defines precision, coverage, and active-feature count in its caption.

  • Why Dowker complexes and Proposition 3 matter (W2, Q4): Within each dictionary, features that co-activate on the same tokens form a neighborhood structure (a Dowker / Vietoris-Rips complex). Section 2 clarifies that Theorem 2 simply guarantees symmetry: comparing Dictionary 1 to 2 yields the same global structure as comparing 2 to 1. Proposition 3 then gives the exact worst-case neighborhood distortion of the matched pairs in 9 self-contained lines, accompanied by runnable two-minute verification scripts in the supplementary zip.

If any specific definition, proof step, or experimental result in the revised PDF is incorrect or remains unclear, please point to it specifically and we will address it on technical grounds.

Reviewer_h9DTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper talks about a certification framework for the matching of dictionaries that come from different layers or models. The proposed certification comes includes finding best-partner matches and excluding non-mutual matches. In this setup, round trip is exact by definition. The paper moves on to describe the coverage (fraction of matched features) and delta_mut (changes in geometry affected by the matches). The matches are accurate on synthetic data but, since there is no ground truth for real dictionaries, coverage needs to be calibrated empirically. Therefore, the extent to which the certificate actually certifies correctness, is not very clear.

Critical strengths

It is an honest report of the findings. Even though the “certified” framing is too strong, the paper does not overstate. For example, highlighting the differences between gpt-2 and gemma2-2b, reveals that only the location of a minimum transfers, and extrapolations from the synthetic experiments are not faithful.

Critical weaknesses

The major weakness is the claim of certification. First, matched pairs may preserve each dictionary's internal geometry, but is the match semantically right? Does delta_mut or coverage predict recovery? In the absorption experiment, a matcher with lower distortion does not recover all the intended pairs, and the authors mention that distortion and accuracy are distinct axes. All these are important questions because the certificate is the main aspect of the study but it is not backed by correctness validation.

Questions for authors
  1. Many symbols are not defined, and make the comprehension of the text difficult. E.g. m, n1, n2, J are not defined. Also how are d1 and d2 in section 3 defined? Is it based on the same max-min kernel applied within one dictionary?
  2. Can you provide some evidence whether delta_mut or coverage are able to predict recovery on split, absorb, or noise?
  3. Please show a case where delta_mut is small but the matches are wrong? If this is not possible, explain why this is the case.
  4. How are arg-max ties broken? What fraction of coverage loss on real dictionaries is due to ties or saturation rather than genuine absence of a counterpart?
  5. How is “agreement” with the baselines defined (all features or mutual set only), and is it coupled to coverage by construction?
  6. Compare directly against the other methods mentioned in the related work section, e.g. the semantic optimal transport method and crosscoders.
4 comments
AuthorsOpenReview ↗

Response for Reviewer h9DT

(1/2) We thank the reviewer for pinpointing the central question: whether the certificate establishes correctness or only geometric consistency. The revision answers it directly. The certificate measures geometric consistency. Correctness is established by two further checks that we have added, and is then tested with ground truth and by intervention. The paper now states what each check does and does not establish (§3, §4.1).

Main weakness (whether a committed match is right, and how this is validated). A pair is committed only if it is a mutual best match under the shared-sample relation (under mean co-activation, §3.1) and passes a sample-permutation test (§3.2). The test is valid at level α\alpha for every committed pair (Proposition 4). With ground truth, calibrated matches have precision 1.000 at every noise level from 0.01 to 0.10, and false commitments between unrelated dictionaries fall from 45% to 1.4% (Table 1). Without ground truth, on GPT-2, we ablate aa and rank its partner among all 24,576 next-layer features. Committed partners rank first in 79% of cases, against 31%–43% for assignment matchers (Table 3). Replacing bb by an affine fit of aa recovers 0.840 of bb's ablation effect on the next-token distribution, against 0.300–0.678 for the assignment matchers (substitution recovery, Table 3).

Q1 (definitions of the symbols and of d1,d2d_1,d_2). §2 now defines mm, n1n_1, n2n_2, J1J_1, J2J_2, the codes, K12K_{12} and its mean counterpart before use. d1d_1 and d2d_2 are, as the reviewer guessed, the same max–min co-activation applied within one dictionary, turned into a dissimilarity and normalised to mean one (Eq. 2). We also note that it is not a metric: it violates the triangle inequality in 42% of code-derived instances. This is why the certificate is now proved without the triangle inequality.

Q2 (what δmut\delta_{\mathrm{mut}} and the slack predict). At the set level, δmut\delta_{\mathrm{mut}} on the calibrated set grows monotonically with the noise level (Table 1: 0.000 on the exact permutation, 1.065 at σ=0.10\sigma=0.10), while precision stays at 1.000. δmut\delta_{\mathrm{mut}} therefore measures how far the two geometries disagree, and identity is established by the checks above; the paper now says so. At the pair level, the per-pair slack s(a)s(a), each pair's share of the distortion (Eq. 4), ranks correct pairs before incorrect ones with AUROC 0.87–1.00 across noise, splitting and absorption (Table 1). Ordering pairs by slack gives the precision–coverage curves in the Appendix.

Continued in part 2/2 below.

AuthorsOpenReview ↗

Response for Reviewer h9DT (2/2)

(2/2) Q3 (a wrong correspondence with small δmut\delta_{\mathrm{mut}}). Such a case exists, and §4.1 constructs it. We place 64 features on a regular 8×88\times 8 grid of the torus, so every grid translation is an automorphism of the co-activation geometry. Mapping each feature to the partner of its translate is wrong for every feature. With noise σ=0.05\sigma=0.05, this map has δmut\delta_{\mathrm{mut}} = 0.173, against 0.113 for the true correspondence, and its mean slack (0.028) cannot be distinguished from the truth's (0.027). The geometry cannot tell them apart because a distortion certificate is invariant under automorphisms of the geometry, so it can only certify a correspondence up to such symmetries. Neighbouring grid features also genuinely co-activate, so 64.5% of the translated pairs pass the permutation test. What rejects the translate is the mutual maximum on the shared samples. A feature co-activates with its own partner more than with any translate, and the calibrated matcher recovers the true correspondence (coverage 1.000, precision 1.000). Gromov–Wasserstein matching, which sees only d1d_1 and d2d_2, reaches precision 0.028.

Q4 (tie-breaking, and the coverage lost to ties and saturation). Ties are exact in max–min scores: K12(a,b)K_{12}(a,b) is capped at aa's own maximum. The submitted version broke them by feature index, which made results depend on feature order; on an exact permutation it left δcross\delta_{\mathrm{cross}} = 1.094 where the correct value is 0.000. On GPT-2, 31.7% of rows and 34.1% of columns of K12K_{12} are tied. Breaking ties by mean co-activation changes max–min coverage by at most 1.6 percentage points, so ties themselves account for little of the loss. Saturation is the larger effect, since a single sample decides the max–min score. The revision therefore matches by mean co-activation (§3.1), which leaves 0.2% of rows tied, and max–min breaks those ties. §5.4 locates the remaining loss: most unmatched features share their best partner with another feature (many-to-one convergence onto dense features, 51.8% of all features), and only 8.8% have no partner beyond chance.

Q5 (definition of agreement). It is taken over all n1n_1 features: the fraction for which our map μ(a)\mu(a) and the baseline's best partner coincide. It is not coupled to coverage by construction. The Table 4 caption now says this. Because agreement alone is weak evidence (as C7pZ also notes), we add intervention tests and head-to-head comparisons on features committed by both matchers (§5.1).

Q6 (SemOT and crosscoders). Both are added. SemOT follows its released code (Tables 2–3), and we trained TopK crosscoders on three GPT-2 layer pairs (Table 3). Mutual correlation and kernel ablations are included as well.

We are glad to answer any further questions during the discussion.

Thank you for your reply. Many of my concerns have been addressed and I will raise my score.

Thank you very much for taking the time to read our revision and for raising your score. Since you noted that many of your concerns were addressed, please let us know if any minor question remains and we would be happy to clarify it before the discussion period closes.

Reviewer_C7pZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors proposed a way to match features extracted from sparse auto encoders dictionary, it finds the sample where both features agree and have high correlation. It works on mutual matches and picks where feature from both dictionaries agree. It reports coverage and certificate for matched features. Authors show experiments on synthetic features, GPT 2 and Gemma 2-2B.

Critical strengths
  • The idea is interesting and understanding matches that are missed is more important in large models where understanding small information is important.
  • The method does not need a shared activation space or equal dictionary sizes, only shared input samples.
  • The method is simple and cheap.
Critical weaknesses
  • Since the perfect recovery only permutation change can get, the authors should tone down the abstract claim that every committed match is correct.
  • There is no ground truth for GPT-2 or Gemma - no proper comparison for top-activating examples, no output check when matches are changes, no causal test as in what if partner changes when the features are ablated.
  • Since matchers are different and they point at different things, low agreement at depth layers does not specifically says anything for same set of features.
  • Baseline are very limited: crosscoder and semantic optimal transport which author recognizes as closes work is not used as baseline.
  • Mutual neighbours under correlation is also missing.
  • Experiments should be more detailed to validate the claims: only two models and one SAE family is used.
  • The paper motivates the method by different widths, different models, and fine-tunes, but every real experiment matches equal-size dictionaries between layers of one model.
  • coverage on real models is very less, claiming most features are unmatched - if this is the case, what next? no downstream cases and applications are discussed and explored,
  • The text says mean distortion is 0.001–0.004, but Table 4 shows values up to 0.0097.
Questions for authors
  • What does the unmatched features actually represent? need more analysis on that,
  • a coverage precision curve can help understand how much change is anticipated with adding non-matched features.
  • If more features are added, does mutual matching score changes as new samples are there?
  • ll experiments use equal-width dictionaries. How does the method behave when matching, for example, a 16k SAE to a 65k SAE on the same layer?
  • what happens when features across languages, modalities are matched? One claimed point is not needing a shared activation space. Have you tried a vision–language setting, for example image–caption pairs as shared samples, or image-token vs. text-token features inside a VLM?
  • Why max–min specifically ? an ablation study can help understand.
4 comments
AuthorsOpenReview ↗

Response for Reviewer C7pZ (1/3)

(1/3) We thank the reviewer for a detailed review. Each point led to a new experiment or a correction. Section numbers refer to the revised PDF.

W1 (precision for every transformation). The revision replaces the blanket statement with precision against ground truth for every transformation, and it changes the matcher so that the statement now holds well beyond exact permutations. The submitted rule had precision 0.826 at σ=0.05\sigma=0.05. The revised matcher matches by mean co-activation rather than max–min (§3.1, prompted by your Q6) and adds a per-pair permutation test (§3.2). With these, precision is 1.000 and coverage 1.000 at every noise level from σ=0.01\sigma=0.01 to 0.100.10, and precision is 1.000 for every transformation in Table 1 except absorption (0.988).

W2 (top-activating comparison, output check and causal test). All three are added (§5.1, Table 3), each for every matcher on the same features. The causal test ablates aa at layer LL, runs the block, and ranks its partner bb among all 24,576 features of layer L+1L+1 by the change in activation. The output check replaces bb by an affine fit of aa and measures the fraction of bb's ablation effect on the next-token distribution (KL) that is recovered, on a fixed set of 256 target features shared by all matchers. The top-activating comparison is the Jaccard overlap of held-out top-50 tokens, together with held-out R2R^2. Committed pairs have causal rank 1 in 79% of cases. Assignments from correlation, decoder cosine, SemOT and a crosscoder reach 31%–43%.

W3 (evidence on the same features at depth). The revision adds intervention tests at every layer pair (hit@1 column of Table 4) and head-to-head comparisons on the same features. Against mutual correlation, on features committed by both, the two give the same partner in 98.2% of cases. Where they differ (79 features), our partner has the better causal rank in 39 cases and correlation's in 36. Against mutual decoder cosine, agreement is 95.4%, and the better causal rank is ours in 115 cases against 68.

W4–W5 (crosscoder, SemOT and mutual-correlation baselines). All three are added, on synthetic data (Table 2) and on GPT-2 (Table 3). SemOT is run as in its released code (top-500 support, 25-candidate shortlist, geomloss Sinkhorn with p=2, blur 0.01, scaling 0.9, score 12W22/s\tfrac{1}{2}W_2^2/s). Its density filter removes every feature once noise is added, so we also report it without the filter. The crosscoders are TopK acausal crosscoders trained on GPT-2 layers 1→2, 5→6 and 9→10 (16,384 latents, k=32, 65M tokens). Mutual correlation is run with and without our permutation test.

On synthetic codes (Table 2), assignment baselines fail once the dictionaries differ: correlation argmax has precision 0.778 under absorption, SemOT without its filter 0.691 under noise, and both commit to every feature of an unrelated dictionary. Our permutation test transfers to correlation and removes its false commitments on unrelated dictionaries (0.38 → 0.01). On GPT-2 (Table 3), calibrated mutual correlation, the strongest baseline, commits more pairs than our matcher (596 vs 403 per layer pair) at lower hit@1 (0.714 vs 0.790); the pairs it adds beyond ours have hit@1 0.563. Where both commit, the two choose the same partner for 98.2% of features. Assignments from SemOT and the crosscoder reach hit@1 0.411 and 0.309.

Continued in part 2/3 below.

AuthorsOpenReview ↗

Response for Reviewer C7pZ (2/3)

(2/3) W6–W7 (other widths, other models, other SAE families). §5.3 adds two settings. For different widths we match Gemma Scope 2B dictionaries of width 2142^{14} against 2162^{16} on the same layer (layers 6, 12, 18); the results are in Q4 below. For different models we match Gemma-2-2B layers 6/13/19 against Gemma-2-9B layers 10/21/31, which share no activation space. The calibrated correspondence covers 0.050–0.243 of the 2B features. Committed pairs have held-out R2R^2 0.739–0.943, against 0.212–0.397 for correlation assignments. Calibrated mutual correlation commits 631–700 pairs, with held-out R2R^2 0.413–0.561. The real experiments now span three models (GPT-2, Gemma-2-2B, Gemma-2-9B) and two released SAE families (jbloom ReLU for GPT-2, Gemma Scope JumpReLU), plus the TopK crosscoders we trained.

W8 (what the unmatched features are, and what the correspondence is used for). §5.4 characterises the unmatched features. Each is classified as mutual but at chance level (0.1% of features), sharing its best partner with another feature (51.8%; many-to-one, the signature of merging or absorption), or without any partner beyond chance (8.8%). The shared class grows with depth from 12.7% to 85.1%, so the coverage lost in deep layers is many-to-one convergence, not missing counterparts. Features without a partner seldom have a decoder twin (cosine > 0.7 for 8%, against 82% for committed ones). The calibrated correspondence also has two direct uses, both measured. Committed partners can stand in for each other in the forward pass (substitution, Table 3). Committed matches also compose across consecutive layers into feature lineages (new Figure 2b): 40.0% of layer-0 features have a committed chain to layer 2 and 10.5% to layer 6, against 28.2% and 1.4% if commitment at each layer were independent, so a feature with a counterpart tends to keep one.

W9 (the range 0.001–0.004). The range referred to the token-budget sweep in the preceding paragraph, not to Table 4. Table 4's layer average, 0.0061, was quoted in the same sentence. We have rewritten the passage so the two cannot be confused.

Q1 (unmatched features). See W8 above and §5.4. For each class we report its share, its activation density, and whether it has a high-cosine decoder partner (cos > 0.7). The last indicates a geometric twin that does not co-activate.

Q2 (coverage–precision curve). On synthetic data the curve orders mutual pairs by slack (Appendix). The slack AUROC column of Table 1 summarises it. On GPT-2 the curve orders pairs by pp-value and slack and plots causal hit@1 against the fraction of the committed set admitted (Appendix). Hit@1 is nearly constant across the committed set, so the pairs admitted last are about as reliable as the first.

Continued in part 3/3 below.

AuthorsOpenReview ↗

Response for Reviewer C7pZ (3/3)

(3/3) Q3 (adding features or samples). Matches change very little. We reran three layer pairs with 512, 1024 and 2048 features at 8192 tokens, and with 2048 to 16,384 tokens at 1024 features (Appendix, Table 8). When features are added (1024 → 2048), the fraction of committed features that are committed again is at least 0.996, and the fraction keeping the same partner is 0.973–0.996. When tokens are added, the fraction of features committed in both runs that keep the same partner is 0.926–0.981. With fewer tokens the mutual set contains more chance-level pairs, which the test rejects; as tokens are added the mutual and calibrated sets converge (Table 8).

Q4 (16k against 65k on the same layer, §5.3 and Table 5). At Gemma-2-2B layers 6, 12 and 18 we match the 1024 most active features of the 16k dictionary against the 4096 most active of the 65k one. The calibrated correspondence covers 0.119–0.310 of the narrow features. Because both dictionaries read the same residual stream, decoder cosine is an independent check: 98%–100% of committed pairs have cosine above 0.7, against 73%–84% for correlation assignments and 68%–79% for SemOT. Held-out R2R^2 is 0.869–0.934. Calibrated mutual correlation, the strongest GPT-2 baseline, commits 934–966 pairs, of which 79%–88% have cosine above 0.7. Coverage is lower than between adjacent layers because a narrow feature is often split across several wide features, and mutuality keeps at most one of them.

Q5 (languages and modalities). The procedure needs only shared samples, so it applies to a vision–language setting with image–caption pairs, or image and text token positions of one VLM, as the samples; §7 states this scope. The 2B↔9B experiment (§5.3) already covers the property that setting needs: two representations with no shared activation space, matched through shared samples alone.

Q6 (choice of kernel). This question changed the method. Under the same mutual rule and test we compared max–min, mean co-activation (mean–min), max–product and correlation (Tables 2–3, Appendix). Mean co-activation is the better matching score. Max–min ties exactly through saturation (31.7% of GPT-2 rows), and a single sample decides it. On all GPT-2 layer pairs, max–min commits 284 pairs per layer pair at causal hit@1 0.721, against 403 at 0.790 for mean co-activation; where the two disagree, mean co-activation's partner has the better causal rank in 95 of 126 cases. On synthetic codes under noise σ=0.05\sigma=0.05, max–min keeps coverage 0.575 against 1.000. Max–min keeps two roles: it breaks the rare ties of mean co-activation, and within each dictionary it is the filtration value of the Dowker complex, which defines the geometry on which the certificate is computed. The kernel was chosen on three layer pairs (1→2, 5→6, 9→10). On the 8 others, mean co-activation has the higher hit@1 in 8 and commits more pairs in 7.

We are glad to answer any further questions during the discussion.

AuthorsOpenReview ↗

Brief update on the revised PDF

Dear Reviewer C7pZ,

As the discussion period wraps up soon, we wanted to briefly note that yesterday we also updated Figure 1 (Page 2) into a full pipeline overview diagram to make the method easier to follow at a glance. Your initial review directly shaped our revision, especially Q6 (which led us to adopt mean co-activation as the matching score) alongside the new GPT-2 causal ablations, SemOT and TopK crosscoder baselines, and Gemma 16k-vs-65k and 2B-vs-9B experiments. Whenever you have a moment, we would really appreciate knowing whether the revised manuscript addresses your concerns or if any questions remain.

Comments on the paper

2 comments
AuthorsOpenReview ↗

summary of the revision (1/2)

(1/2) We thank all four reviewers. Read together, the reviews raise one central question: does the certificate show that a match is correct, or only that it is geometrically consistent? There are also requests for stronger baselines, for settings beyond equal-width layers of one model, and for a much clearer presentation. The revised PDF addresses each request. The supplementary material now contains code for the matcher, the certificate and the test, with scripts that reproduce Table 1 and the exactness check of Proposition 3 in about two minutes on a CPU.

What is certified, and how correctness is now tested (Sections 3–4, new Figure 1). The method commits a pair only after three checks, and the paper now states what each one establishes.

The certificate is now exact. Following juE1, the total certificate (old Eq. 2) is replaced by its exact form, with round-trip terms of coefficient ¼ in pairwise form. It is necessary and sufficient for the maps to interleave the two Vietoris–Rips filtrations, with no triangle inequality assumed. This matters because the co-activation dissimilarity is not a metric: it violates the triangle inequality in 42% of code-derived instances. The formula matches a direct search on 700 instances to within 5×10−135\times10^{-13}. δmut\delta_{\mathrm{mut}} on the mutual set is unchanged.

The calibration test is new. Each mutual pair is tested against a sample-permutation null, with the null taken as a maximum over all candidate partners. The test is valid at level α\alpha for every committed pair, whatever rule selected the partner (Proposition 4). Between unrelated dictionaries, false commitments drop from 45% to 1.4%.

The matching score is new as well. Max–min scores tie exactly through saturation: 31.7% of rows are tied on GPT-2, and on an exact permutation first-index tie-breaking leaves δcross\delta_{\mathrm{cross}} = 1.094 where the correct value is 0. Following C7pZ's question about max–min, we ablated the kernel. Mean co-activation, which was already our tie-break and test statistic, gives better matches. It is now the matching score (§3.1). Max–min remains the within-dictionary Dowker geometry on which the certificate is computed, and it breaks the rare remaining ties.

With ground truth (synthetic), calibrated matches have precision 1.000 and coverage 1.000 at every noise level from σ=0.01\sigma=0.01 to 0.100.10. The submitted rule had precision 0.826 and coverage 0.527 at σ=0.05\sigma=0.05. The per-pair slack separates correct from incorrect mutual pairs with AUROC 0.87–1.00 (Table 1). We also construct the case h9DT and juE1 asked about, a wrong correspondence with small distortion, and show what rejects it (Section 4.1).

Continued in part 2/2 below.

AuthorsOpenReview ↗

summary of the revision (2/2)

(2/2) Correctness without ground truth (new Section 5.1). For each committed GPT-2 pair (a,b)(a,b) we report four tests. The causal rank ablates aa at layer LL, runs the block, and ranks bb among all 24,576 features of layer L+1L+1 by how much it changes. Substitution recovery replaces bb by an affine fit of aa and measures how much of bb's ablation effect on the next-token distribution (KL) is recovered, on a fixed target set shared by all matchers. We also report held-out R2R^2 and top-50 token overlap.

Our committed pairs have hit@1 (causal rank 1) of 0.790 and hit@10 of 0.916. Matchers that assign every feature reach hit@1 0.31–0.43, and their mutual variants 0.686–0.709. Calibrated mutual correlation, the strongest baseline, commits 596 pairs per layer pair at hit@1 0.714; the 2199 pairs it adds beyond ours over all 11 layer pairs have hit@1 0.563. Where both commit, the two choose the same partner for 98.2% of features.

Baselines (new Tables 2 and 3). All baselines are run on the same samples and features, both as an assignment and, where defined, restricted to mutual pairs. They are SemOT, run as in its released code (top-500 support, 25-candidate shortlist, geomloss Sinkhorn with p=2, blur 0.01, scaling 0.9), together with a variant without its density filter; TopK crosscoders trained on GPT-2 layer pairs 1→2, 5→6 and 9→10; mutual nearest neighbours under correlation, with and without our test; decoder cosine, as an assignment and mutual; and the max–min and max–product kernels in place of mean co-activation (§3.1). On all GPT-2 layer pairs, max–min commits 284 pairs per layer pair at hit@1 0.721, against 403 at 0.790 for mean co-activation. Where they disagree, mean co-activation has the better causal rank in 95 of 126 cases.

Settings beyond equal-width layers of one model (new Section 5.3). For different widths we match Gemma Scope 2B dictionaries of width 2142^{14} against 2162^{16} on the same layer (layers 6, 12, 18), with decoder cosine as an external check. Of the committed pairs, 98%–100% have cosine above 0.7, against 73%–84% for correlation assignments. Calibrated mutual correlation commits 934–966 pairs, of which 79%–88% have cosine above 0.7. For different models we match Gemma-2-2B layers 6/13/19 against Gemma-2-9B layers 10/21/31, which share no activation space. The calibrated correspondence covers 0.050–0.243 of the 2B features. Committed pairs have held-out R2R^2 0.739–0.943, against 0.212–0.397 for correlation assignments. Calibrated mutual correlation commits 631–700 pairs, with held-out R2R^2 0.413–0.561. The real experiments now cover two SAE families (ReLU for GPT-2, JumpReLU for Gemma Scope) plus TopK crosscoders.

Unmatched features (new Section 5.4). Every unmatched feature is classified (shares of all features; 39.4% are committed): mutual but at chance level (0.1%); sharing its best partner with another feature, a many-to-one pattern like merging or absorption (51.8%); or without any partner beyond chance (8.8%). Ties account for almost none of the loss. Committed matches chain across layers far more often than independent layers would give (10.5% of layer-0 features reach layer 6, against 1.4%), and they are stable when samples or features are added.

Presentation. Two new main-text figures show the pipeline on a toy example (Figure 1) and the fate of every GPT-2 feature across depth (Figure 2). The abstract is rewritten without formulas, and the title is unchanged. The introduction now states the problem, what existing matchers lack, what we provide and how each claim is tested, and it introduces GPT-2 and Gemma together. Every symbol is defined before use (m,n1,n2,J1,J2,d1,d2m, n_1, n_2, J_1, J_2, d_1, d_2, VR and its scale convention, coverage, precision), and each experiment starts with a stated hypothesis. The Contributions are one compact paragraph, and "Dowker metric" is renamed "dissimilarity", with its definition.

The agreement column of Table 4 is now computed from the certified map, which is built on per-code normalised codes like every other quantity in the paper. Decoder cosine agrees least at the embedding transition (0.386, against at least 0.451 elsewhere).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.