acceptodds
Under review as a conference paper at ICLR 2027

Sparse-Autoencoder Attribution Graphs Are Decided at the Margin

45 views

Abstract

Attribution graphs over sparse-autoencoder features predict an ablation’s effect, and are compared across seeds. Under a TopK rule a feature is active where its pre-activation is positive and among the largest, and its signed margin, positive on an active feature, measures the move that changes the membership. On five public TopK pairs of three Pythia models, zeroing one feature moves a downstream feature across the cutoff in 0.74 of cases or more, where the crossing features carry most of the response’s energy. A write scales a feature’s code by its dose. Over three doses the graph held on the active set has a median relative error of 0.247 or more on the code change. The margin graph adds the write’s first-order change to every pre-activation and re-ranks, at the same one Jacobian-vector product per live feature, and errs by at most 0.033. At the median a crossing coordinate’s code moves several times as far as its pre-activation, and the crossing features’ decoded change has 1.24 times the model’s change in norm or more. On one testbed of eight runs, sparse nonnegative maps writing one dictionary’s atoms over another’s explain 0.85 of the energy of the operator between two layers’ codes on the source’s block of live upstream and downstream features, and 0.47 on graphs masked by their margins’ signs.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, sparse autoencoders, TopK, attribution graphs, activation margins, feature interventions, support changes, residual-stream propagation, cross-seed operator transport

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_pb63OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper explores the reliability of sparse autoencoder attribution graphs under feature interventions. It shows that existing fixed-support graphs often fail to account for changes in active features after interventions. To address this, the paper proposes Margin Graph, which combines first-order approximations with TopK re-ranking to predict downstream feature changes.

Critical strengths

None

Critical weaknesses
  1. The presentation of this work is poor. The paper contains numerous awkward and unclear sentences. The overall organization also lacks a clear focus, making it difficult to identify the main contributions. In particular, the theorem statements and figure captions are unnecessarily lengthy and difficult to follow. The paper needs substantial improvement in writing, structure, and clarity. The AI Use Statement of this work also indicates extensive use of generative AI tools throughout the research and writing process.

  2. The technical novelty is limited. The proposed Margin Graph essentially applies a first-order approximation followed by TopK re-ranking, which is closely related to a single-step gradient approximation (e.g., IG with k=1). Moreover, it does not outperform the compared re-ranked baselines, including IG, in prediction accuracy. The paper therefore does not clearly establish a substantial methodological advance over existing techniques.

Questions for authors

No specific questions. The main concerns relate to the limited technical novelty and very poor presentation, which are unlikely to be resolved through clarification alone.

4 comments
AuthorsOpenReview ↗

Comment to Reviewer pb63

Thank you for your time on our paper and for your candid review. We have revised the paper for clarity and flow.

1. Presentation, sentences and focus. Now, the main text follows one claim, that an attribution graph between TopK sparse autoencoders (SAEs) is decided at the margin. The introduction closes on four numbered contributions, and Figure 1 draws the method above one measurement for each contribution.

2. Theorem statements and captions. Each theorem in the main text now states its hypotheses and its conclusion within five lines, and a paragraph after it says what it means. Each caption now opens on the finding its exhibit shows and stays within sixty words, or within a hundred where its symbols or counts need definitions. The theorem statements and the captions describe the objects and the comparisons they present.

3. The statement on the use of language models. The statement is our honest disclosure under the venue's policy, and we respectfully suggest that it might not be counted as a critical weakness.

4. Technical novelty and integrated gradients at one step. The margin graph does apply a first-order approximation and then rerank. Section 4 now compares it with the reranked corrections and separates clean derivatives that can be reused from work that is repeated for each intervention. We measured integrated gradients at five and at ten JVPs and not at a single step, so the page reports no number for one step. The margin graph takes its JVPs at the clean input, so one set of JVPs at a position serves every intervention at that position, at every scaling factor and on every subset. A path average must be computed again for each intervention.

5. Accuracy against the reranked corrections. On our SAEs, reranked integrated gradients err by 8.8 × 10⁻⁶ at five JVPs per intervention and the reranked softmax correction errs by 0.013, against 0.017 for the margin graph at one JVP. The reranked softmax correction errs less than the margin graph on all five public pairs. Section 4 reports this as a finding and prints each method's JVP count beside its error. On the clean active set, every correction errs by 0.777 to 0.783 against 0.786 for the standard graph, so reranking removes most of the error and the first-order correction removes little of it.

6. The methodological advance. The advance the page claims is the measurement around a known form. On five public TopK pairs, zeroing one feature moves a downstream feature across the cutoff in 0.74 to 0.89 of cases. Theorem 1 proves the margin graph exact on an affine segment. Section 4 finds that reranking removes most of the error, and the margin graph reranks at the JVPs the standard graph already takes.

7. Concerns beyond clarification. The general response states the page's claim, its behavioral result and its measured costs, and it names the sections that establish each of them.

AuthorsOpenReview ↗

Follow-up Comment to Reviewer pb63

We further revised the paper again to make it easier to read. The changes are listed below:

  • The abstract and introduction now explain features, coefficients and selection before they introduce the margin graph.
  • Section 2 gives the definitions, and Table 2 collects the symbols in one place. Where the words preactivation, active, inactive and code named measured quantities, we replaced them with encoder score, selected, unselected and coefficient vector, respectively.
  • The introduction now lists six findings.
  • Figure 1 now illustrates selection before and after an intervention and marks the cutoff and one feature's margin. It no longer contains the measurements for the contributions described in our first response.
  • Plain explanations now come before the main theorems, and the captions define their quantities and populations.
  • Our earlier statement that every theorem and caption met the short length limits no longer holds for this version. Some statements and captions are now longer, so that their objects and conditions can be read alongside them. The mathematical conditions and conclusions are unchanged.
  • Each results section explains the purpose of its comparisons.
  • Section 4 still distinguishes selecting features again from improving the derivative approximation, and its comparison keeps the stronger reranked baselines. Figure 3 now sits in Appendix B.
  • Section 7 states the relation to attribution patching and integrated gradients, and Section 5 keeps the adverse result for the measured model output.

No experiment or numerical result has changed since last version, and the experiments still use the same models, SAE pairs and task.

Thank you again for your time and consideration.

Thanks for the authors' response. The writing is better than before, but the presentation and organization still need considerable improvement. In addition, the comparison with baselines also does not show a clear advantage of the proposed method. Therefore, I will maintain my score.

Thank you for reading the revision. We always agree that reranked integrated gradients achieves lower coefficient error than the margin graph, and we also agree with your observation about one-step approximations: evaluating the derivative at the clean endpoint and then applying TopK gives our predictor. Section 7 acknowledges the prior work behind this order of operations.

The comparison we would most like to draw your attention to is within the softmax baseline itself. When one feature is zeroed on the public Pythia-70M pair at layers 3–4, the median relative coefficient errors are:

Score approximationKeep the original selectionSelect again from all predicted scores
First order0.7390.121
Softmax correction0.7310.072

These four cells use the same interventions (Table 15, p. 33). Improving the score approximation while keeping the original selection changes the error from 0.739 to 0.731. With that same softmax correction, selecting again changes it from 0.731 to 0.072. The stronger baseline is therefore particularly informative: it shows why improving the score approximation alone leaves much of the coefficient error in place. Across the five public TopK pairs, zeroing one feature changes the downstream selection in 74–89% of cases (§3).

We appreciate that you may judge this empirical finding insufficiently significant for ICLR. Could you clarify whether that is your remaining concern, or whether you see a particular problem with the evidence for the finding? If a passage in the revised paper makes that evidence difficult to assess, a specific example would help us address it.

Thank you again for your time.

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

Attribution methods for SAEs construct a graph relating SAE features from one layer to SAE features in another layer. This work measures how often a change in an upstream feature turns a downstream feature off, invalidating the constructed graph.

Critical strengths

While I don't think the ideas are well presented here, I do think how to construct intervention invariant attribution graphs is an interesting and worthy idea.

Critical weaknesses
  1. I don't feel this paper is well written at all. Terms are used without definition, for example "field's devices", "read in one gauge", and most importantly the "margin graph" itself which is never defined formally and which an entire section rests on. Generally there is a lack of explanation as to why experiments are being run and what the results mean, for instance, in table 2 the "FVU up" and "FVU down" is listed in each row but these are never referred to in the main text, and the captions mentioning of them doesn't give the reader any insight into what they should understand from these numbers. There are examples like this throughout the paper.

As it stands I had real difficulty understanding the methods and contributions of this paper.

Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Comment to Reviewer HNoi

Thank you for taking the time to review our paper and for your candid comments. We have revised the paper for clarity and flow.

1. On the summary. An intervention can also switch a downstream feature on. On the five public TopK pairs, across the zero ablations that change the active set, the entering features alone account for a mean of 0.33 to 0.39 of the squared norm of the code change. Figure 1 now draws one feature entering the active set and one leaving it.

2. The writing. The paper follows one claim, that an attribution graph between TopK sparse autoencoders (SAEs) is decided at the margin. Figure 1 now draws the method above one measurement for each contribution, and the introduction closes on four numbered contributions, each naming its section.

3. Terms used without definition. Each term the review names is now defined or replaced. In Section 2, Display 1 now defines the margin of a feature as its signed distance to the cutoff, and Display 2 defines the prediction of the margin graph beside that of the standard graph. Section 4 now names and defines the two corrections of attribution patching where they enter. As Section 2 states, every SAE is measured in the units of the residual stream.

4. Why each experiment runs and what its results mean. Each results section now opens on its finding. Table 1 now reports the fraction of variance unexplained at each layer, and Section 2 now says what its two columns show. We measure the active set changes on SAEs that reconstruct their layers well, with fractions of 0.063 to 0.165 on the public TopK pairs. The fractions of our SAEs, 0.109 and 0.155, lie close to those of the running example at the same layers.

5. Examples throughout the paper. The main text now names every column of every table and says what each column shows. Each results section now states each finding under a short heading.

6. The methods and the contributions. Figure 1 now draws the method on six illustrative preactivations, and Displays 1 and 2 state it in two lines. The introduction now lists the four contributions, each with its section and its number.

I thank the authors for taking the time to respond to my concerns. While the presentation has improved, it hasn't improved enough to overcome all of my concerns and as such I maintain my recommendation.

Thank you for reading our response and the revised paper. We would like to resolve the concerns that remain. Could you kindly tell us, by section or by figure, which parts of the presentation are still unclear? We are revising the paper again and will address each part you name.

AuthorsOpenReview ↗

Follow-up Comment to Reviewer HNoi

We revised the paper further to make it easier to read. The changes are listed below.

  • Abstract and Introduction: The Abstract now defines a feature, its coefficient, TopK selection and the margin in plain words before it states the finding, and it closes on what selecting features again improves and what it does not. Before any formula, the Introduction now explains why an attribution graph that keeps the original selection misses a feature whose score crosses the cutoff.

  • Purpose of each experiment: Every results section now opens on its finding and then states what the experiment tests and why. Section 5, for example, asks whether more accurate coefficients also predict the model's behaviour, and it finds that they do not.

  • Table 1: Section 2 now states that the FVU up and FVU down columns measure the upstream and the downstream SAE respectively, and that our SAEs reconstruct their layers about as closely as the public pair does at the same layers. It also distinguishes this share of variance from the reconstruction residual that Section 5 measures.

  • Other unclear passages, we made four changes:

    • A sentence that gives the idea in words now precedes each theorem and each main definition.
    • Table 2 now collects the symbols.
    • Where the paper wrote "preactivation", "active", "inactive" and "code", we now write "encoder score", "selected", "unselected" and "coefficient vector".
    • Each caption now opens on what its exhibit shows and defines the terms it uses, so that the exhibit reads on its own.
  • Methods and contributions: The Introduction now lists the six findings in the order of the sections and names the section that reports each one. Section 2 defines the SAEs, the margin, the two graphs and their error before any result uses them.

Thank you again for your time and consideration.

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

The paper studies attribution graphs for TopK sparse autoencoders (SAEs). It shows that standard fixed-support attribution graphs often fail because interventions change which features are active.

The authors propose a margin-based attribution graph that predicts these changes by recomputing TopK activation after a first-order approximation. Experiments show substantially lower prediction errors without additional Jacobian computations.

The paper also investigates discrepancies between decoded feature changes and actual model activations, as well as the consistency of attribution graphs across SAE training seeds. Experiments are mainly conducted on small Pythia models.

Critical strengths
  1. Clear motivation and effective solution. The paper demonstrates that changes in active SAE features are common and can cause large errors in fixed-support attribution graphs. The proposed margin graph significantly improves prediction accuracy at similar computational cost.

  2. Careful empirical analysis. The paper provides theoretical justification, extensive experiments, and useful controls. The finding that decoded feature changes can differ substantially from actual model activation changes is also valuable for interpreting attribution graphs.

Critical weaknesses
  1. The main approach of linearizing before a nonlinearity and recomputing its activation is related to existing methods such as AtP*. Experiments focus on small Pythia models and TopK SAEs, without evaluating larger models or other SAE architectures. More importantly, the paper does not demonstrate whether the improved predictions lead to better circuit discovery or more faithful explanations. Results for multi-feature interventions are also mixed.

  2. The paper introduces many specialized terms and combines several loosely related analyses, making it difficult to follow. The cross-seed results are based on a limited testbed, and some experimental findings are difficult to interpret. A clearer focus on the main method and its practical implications would strengthen the paper.

Questions for authors
  1. Can the authors demonstrate that the margin graph improves downstream circuit-level outcomes, such as circuit faithfulness, edge selection, or predictions of logit changes?

  2. How well does the method generalize to larger models and other SAE architectures, such as JumpReLU SAEs or transcoders?

  3. Why are the main results averaged across different ablation strengths? Could the authors emphasize full-ablation results, where support changes are more significant?

  4. The margin graph performs worse for some multi-feature interventions. How should practitioners handle these cases, and what is the additional computational cost?

3 comments
AuthorsOpenReview ↗

Comment to Reviewer iChi 1/2

Thank you for taking the time to review our paper and for these questions. We have revised the paper for clarity and flow.

1. Relation to AtP*. The margin graph is indeed related to AtP*, it takes the factored form that the corrections to attribution patching use for a neuron's nonlinearity and applies it to the TopK readout of a sparse autoencoder (SAE), whose nonlinearity is a sort. Section 4 and the related work state this. Our contribution is the measurement around that form. On five public TopK pairs, zeroing one feature moves a downstream feature across the cutoff in 74 to 89 percent of cases. Theorem 1 proves that the margin graph is exact on an affine segment, and Section 4 finds that reranking removes most of the error while the first-order correction removes little of it.

2. Small Pythia models and TopK SAEs. We study two Pythia models under the TopK rule, with one ReLU release as a contrast. The one public JumpReLU release with SAEs at adjacent residual layers is on a model behind an access agreement. Section 2 and Table 1 name every pair we measured, and Section 2 gives the margin and the margin graph's prediction for a JumpReLU code, with each feature's threshold in place of the cutoff.

3. Circuit discovery and faithful explanations. We measure the downstream code, and we report the test on logit changes in Appendix F and Section 5. At the logits, the standard graph's code is more faithful once it is patched into the model. Whether better predictions of the code select more faithful circuits is the open question we state in the conclusion.

4. Mixed results for interventions on several features. These results were indeed mixed previously, and Table 2 gives them in the main text. On the running example with sixteen features, the margin graph errs by 0.670 against 0.631 for the standard graph, and Section 4 states this under its own heading. At five Jacobian-vector products (JVPs) per subset, reranked integrated gradients err by at most 0.0043 on our SAEs and at most 0.124 on the running example, at every size.

5. Specialized terms and loosely related analyses. Display 1 defines the margin, and Display 2 sets the standard graph's prediction beside the margin graph's. The main text follows the change in the active set through code prediction, decoding and transfer, and the further analyses stand in the appendices.

6. The limited testbed and findings that are hard to interpret. We measure transfer across seeds on our SAEs alone, because no public release carries a second seed, as Section 2 says. Table 4's caption gives that one configuration and how its intervals are formed. Each results section opens on its finding and states each further finding under a short heading.

7. Focus on the method and its practical use. The introduction closes on four numbered contributions. Section 4 separates the shared JVPs from the construction and query costs that Appendix G measures. The conclusion states that keeping every row and reranking improves code predictions, and that exact recomputation is faster at the workloads we measured.

AuthorsOpenReview ↗

Comment to Reviewer iChi 2/2

(Continue)

8. Question 1, predictions of logit changes. We report this test in Appendix F and in one paragraph of Section 5. On subject-verb agreement with Pythia-70M and the SAEs at layers 3 and 4, we patch each graph's predicted downstream code into the model at layer 4, together with the reconstruction residual from the clean pass. We then compare the resulting change in the logit difference between the two verb forms with the intervention's own change, under criteria we fixed before we saw the outcome. Over 4,893 interventions, the margin graph's code errs by 0.285, which is 95 percent more than the standard graph's 0.146. The 90 percent interval for that ratio runs from 1.84 to 2.07, and the margin graph's code also errs more than a prediction of no change, which errs by 0.237. At the logits, then, the standard graph's code is the more faithful one, which reverses the order we find on the code. The margin graph names the features that enter or leave the active set, and their decoded change overshoots the model's own change. Inside the model, the change in the reconstruction residual cancels that overshoot, but a patched code leaves out the residual's change. Section 5 reports this result plainly as a negative result.

9. Question 2, larger models, JumpReLU SAEs and transcoders. The margin is defined for a JumpReLU code with each feature's threshold in place of the cutoff, so the margin graph applies there, and its rate of active-set changes remains to be measured. The one public JumpReLU release with SAEs at adjacent residual layers is on a model behind an access agreement, and we measured no transcoder. Section 2 gives the margin and the prediction for a JumpReLU code.

10. Question 3, results averaged over scaling factors. We fixed the pooled median over three scaling factors as the summary in advance, before we read any predictor's error, and Section 2 says so. Table 2 gives each factor separately in the main text, and the zero ablation leads the abstract and the introduction. At the zero ablation, the standard graph errs by 0.739 to 0.803 on the five public TopK pairs, and the margin graph errs by 0.032 to 0.128.

11. Question 4, what a practitioner does with many features and at what cost. At a given position, the margin graph's cached JVPs serve every single-feature intervention. Its complete workload at the evaluation counts costs more than exact recomputation on both machines we measured. For many features at once, reranked integrated gradients at five JVPs per subset err by at most 0.0043 on our SAEs and at most 0.124 on the running example, at every size up to sixteen. They require five JVPs for each subset, while the margin graph reuses one set of JVPs per position. Appendix G gives the measured total workloads beside their code errors.

Thank you again for your time and effort!

AuthorsOpenReview ↗

Follow-up Comment to Reviewer iChi

We have revised the paper further to make it easier to read.

We made these changes to the presentation:

  • The Abstract now begins with the setting and introduces features, coefficients and selection before it gives the zero ablation comparison.
  • Section 2 defines the margin and both graph predictions, and Table 2 collects the symbols.
  • We use the terms encoder score, selected and coefficient vector throughout.
  • The Introduction now lists six findings, which replace the four contributions described in our first response.
  • Each results section states what its comparison tests and why that comparison follows from the preceding result.

The comparisons cited in our first response now appear in new locations:

  • Table 3 gives the errors at each ablation strength, including zero ablation.
  • Table 24 of Appendix E gives the subset results, and Section 4 interprets them.
  • Table 16 of Appendix C gives the transfer comparison that we previously cited as Table 4.
  • Figure 3 now appears in Appendix B.
  • Section 6 now defines the graph reconstruction score. It also separates reconstruction on the source's selected features from the further loss when the target is masked as well.

No experiment or numerical result has changed since the last version. The experiments still use the same models, SAE pairs and task, and the paper retains these results.

  • Section 4 reports the reversal for larger subsets, the additional derivatives needed by integrated gradients that select again, and the measured cost of exact recomputation.
  • Section 5 and Appendix F retain the adverse result for the measured model output.
  • Section 7 identifies the relation to attribution patching.

Thank you again for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary 1/2

1. The claim and its evidence

We study how changes in feature selection affect attribution graphs between two TopK sparse autoencoders (SAEs). The standard graph keeps the features selected before an intervention. The margin graph uses the same first-order prediction of the encoder score changes, keeps every downstream score, and applies TopK again. Section 2 defines a feature’s margin in Equation 1 and gives both graph predictions side by side in Equation 2.

On held-out Pile positions, zeroing one upstream feature changes the downstream selection in 74% to 89% of cases across five public TopK SAE pairs. On the public Pythia-70M pair at layers 3 and 4, the median relative coefficient error falls from 0.739 for the standard graph to 0.121 for the margin graph. Table 3 reports each intervention strength separately.

Table 15 shows why selection matters even with a stronger score predictor. On the same public pair and the same zero ablations, the softmax correction gives a median relative coefficient error of 0.731 when it keeps the original selection and 0.072 when it selects again. Figure 3 in Appendix B gives the corresponding comparisons for integrated gradients. Sections 4 and 7 explain the relation to earlier attribution methods.

For interventions on several features, Section 4 reports where the ordering reverses. When all 16 selected features of the public pair are zeroed together, the margin graph errs by 0.670 and the standard graph by 0.631. Table 24 in Appendix E gives the results for subsets of 2, 4, 8 and 16 features, including integrated gradients at five and ten derivative evaluations per subset.

2. The behavioural result

Section 5 and Appendix F test model output on a separate set of 200 subject-verb agreement prompts. We patch each graph’s predicted downstream coefficients into Pythia-70M at layer 4, retain the reconstruction residual from before the intervention, and measure the change in the logit difference between the correct and incorrect verb forms. Table 27 reports the mean absolute error in that change, in logit units.

Agreement interventionsStandard graphMargin graphExact coefficients with the earlier residual
All 9,6000.08430.17830.1689
4,893 retained by the development-set threshold0.14600.28460.2694

The coefficient comparison in Table 3 uses Pile positions. The output comparison in Table 27 uses agreement prompts. On those prompts, even exact downstream coefficients give a larger mean logit error than the standard graph when the patch retains the earlier residual. Table 25 gives the patch checks, and Table 26 reports the retained results by intervention strength.

3. Scope

We evaluate five public TopK SAE pairs on Pythia-70M and Pythia-160M, including a deduplicated 70M model. We also evaluate one public ReLU pair and eight TopK pairs that we trained on Pythia-70M. Table 1 identifies the model, layers and selection rule for each pair.

The agreement test covers the public Pythia-70M pair at layers 3 and 4. Whether more accurate coefficient predictions help select more faithful circuits remains open. Section 2 gives the margin and prediction rule for JumpReLU. The adjacent-layer JumpReLU release we considered requires an access agreement. Empirical results for JumpReLU SAEs, transcoders and larger models remain open.

4. Construction and query cost

Section 4 and Appendix G report the timing comparisons. The standard and margin graphs reuse the same Jacobian-vector products across intervention strengths and feature subsets at a position. The margin graph stores more downstream scores and applies TopK at query time.

Table 31 separates construction from query time on the Apple M4. Relative to the standard graph, margin graph construction took 1.23 times as long on the public pair and 1.14 times as long on our trained pair. Amortized query time was 312 and 118 times as long, respectively. Table 28 reports the complete workload separately. At 48 interventions per position on the public pair, the margin graph took 10.3 ms on the NVIDIA A100, compared with 2.93 ms for exact recomputation.

AuthorsOpenReview ↗

Rebuttal Summary 2/2

5. Revisions since our first responses

The abstract and introduction now explain features, coefficients and selection before stating the main finding. The introduction lists six findings and names the section that reports each one. Figure 1 illustrates one feature entering and another leaving the selection. The measurements appear in the results sections.

Section 2 defines the two graphs. It also explains what the upstream and downstream fractions of variance unexplained in Table 1 tell us about SAE reconstruction. Table 2 collects the notation. Each results section states the question its comparison addresses, and a plain explanation precedes each main theorem. Some theorem statements and captions now exceed the length targets in our first response so that readers can find the conditions beside the claims.

The final revision uses the same experiments, models, SAE pairs and tasks as the preceding version. The numerical results and the mathematical conditions and conclusions are the same.

6. Terms and exhibit locations

The revised paper uses encoder score for preactivation, selected and unselected for active and inactive, and coefficient vector for code when it denotes the SAE representation. Section 6 defines the graph reconstruction score used to measure transfer. The data order pair refers to two runs that share an initialization and see the training data in different orders.

Material cited in our first responsesLocation in the final revision
Illustration of selection before and after an interventionFigure 1
Coefficient errors at each intervention strength, including full ablationTable 3
Integrated-gradient and softmax correctionsFigure 3 and Table 15, Appendix B
Decoded changes and predictions across two blocksFigures 4 and 5, Appendix B
Transfer across seeds and recovery after adding unselected featuresTable 16 and Figure 6, Appendix C
Interventions on feature subsetsSection 4 and Table 24, Appendix E
Agreement test, patch checks and complete resultsSection 5 and Tables 25 to 27, Appendix F
Complete workload and separate construction and query timesSection 4 and Tables 28 to 31, Appendix G

We thank the reviewers and area chairs for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.