acceptodds
Under review as a conference paper at ICLR 2027

Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits

15 views

Abstract

Mechanistic interpretability seeks to explain transformer behavior through circuits: sets of internal components that causally support a behavior. However, self-repair creates a blind spot: ablating a primary component can activate a dormant backup, so a circuit that explains behavior in the intact model can become incomplete under the intervention used to test it. We formulate this gap as conditional circuit completion: given a primary set, identify components that become causally important after its removal. We introduce conditional co-ablation (CoAx), which ranks candidates by growth in ablation effect after primary-set removal. We show that a perfectly dormant backup can be indistinguishable from an irrelevant component to per-unit intact-state scores, whereas its conditional effect change exactly aggregates all interaction orders linking it to the removed set. On GPT-2-small's Indirect Object Identification (IOI) circuit, CoAx recovers the documented backup heads at 0.941 ROC-AUC, versus 0.665 for the strongest intact-state attribution baseline and 0.758 for the matched conditional-energy control. Recovery drops to 0.40 ± 0.13 AUC for alternative component sets matched in behavioral effect, output displacement, and depth, showing that recovery is specific to the removed circuit. Beyond recovery, the CoAx-selected heads are causally load-bearing: freezing them after primary removal sharply reduces the IOI margin, while adding them to the incomplete circuit reduces incompleteness from 0.47 to 0.24. More broadly, conditional growth aligns with intervention-derived repair in 11/12 held-out instances across 4 mechanism clusters, and CoAx completions outperform matched random completions on all 8 non-GPT-2 models spanning 6 architecture families. Therefore, causal explanations of self-repairing transformers must account for backup circuitry when primary components fail.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Transformer Circuits, Self-Repair, Circuit Discovery, Circuit Completeness, Causal Interventions, Backup Circuits, Conditional Ablation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_RNmdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a method to identify self-repair components of transformer circuits. Self-repair is a well-known phenomenon (e.g. “the Hydra effect”) in LLMs where ablating one component changes the loss little as another component takes over. Specifically, this backup component does normally not contribute to the loss, but starts contributing when the primary component is removed. The method introduced in this paper (“COAX”) identifies backup components as those whose importance (measured as ablation effect) increases when the primary circuit is ablated.

Critical strengths

Backup components are an important mechanism in LLMs that should be explained by interpretability; this paper helps identify these components

The method is cheap and simple to reproduce.

The evaluation is rigorous; I specifically appreciated the freezing LayerNorm check (section 3.2)

Critical weaknesses

Major: It seems to me that the method can only deal with a particular class of self-repair mechanisms, but fails for some other causal dependencies. Consider for example (1): A circuit A → B → Output where there exist two redundant backup components B’ and B’’. I believe COAX would not identify these, right? (2): The same circuit A → B → Output with a single backup head B’ that still relies on A. The method tests whether ablating B’ on the model sans A and B has an effect, but because B’ relies on A it no longer has any effect. If these examples are correct, then the method seems quite limited to certain cases to me. (If it depends on the choice of S: how would practitioners use how to choose S?)

Major: It is not clear that COAX provides novelty and significant improvement over past work (Wang et al., Rushing & Nanda). The metrics in section 3 and Appendix C don’t show a big advantage for COAX.

Minor: The paper focuses on 2023-era component circuits where components are large parts of the network like attention heads. More recent work in interpretability focuses on circuits made up of neurons, or sparse autoencoder or crosscoder features. It’s not clear whether COAX scales to such circuits with many components.

Minor: Using AUROC as a metric isn’t very sensitive to false positives because the prevalence of backup heads is low (8 out of 141), precision (which is closer to 50%) may be the more honest metric.

Questions for authors

Major: Does COAX scale to circuits made up of thousands of neurons or sparse dictionary features?

Minor: It’s not clear to me why the Fisher measure is preferred over the KL divergence; KL divergence and CE loss are more commonly used in circuit analysis.

7 comments
AuthorsOpenReview ↗

Response to RNmd (1/4): Backup Structures and Fault Set Selection

Thank you for your incisive questions and for recognizing the importance of backup mechanisms, the simplicity and reproducibility of the method, and the rigor of the evaluation. We address backup structures and fault set selection Part 1, what CoAx adds to prior work Part 2, finer-grained validation and scale Part 3, backup retrieval precision Part 4, and choice of metric Part 5.

Part 1. Backup Structures (W1)

The two counterexamples change different parts of the experiment. If backups are interchangeable, change what you ablate to test a candidate Part 1.1. If a backup depends on A, choose a primary fault S that leaves the shared input in place Part 1.2. We then give a procedure for choosing S Part 1.3.

Causal structureFault set SUnit of analysisWhat reveals the backupValidation
Fully redundant backups{B}; add faults when sampling contextsBackup group or individual unit within a fault contextRemove substitute paths togetherControlled blind recovery
Shared input A{B}Individual unit; validate on a fixed backup setPreserve input from AAblation and replay in IOI

Part 1.1. Redundant Backups

Takeaway. Fully interchangeable backups require a group query or a fault that disables their substitutes.

The single-unit score misses redundant backup components that fully substitute for one another. If B′ and B″ are interchangeable, removing either leaves the computation intact, so its score is zero.

① Ablate pairs to test pairwise redundancy. In a controlled experiment, we ablated all 2,016 pairs among 64 candidates. Only the pair containing both backups showed an increase in effect after the primary fault.

② Sample additional faults when there are more backups. Other backups can hide a candidate's effect. Within each sampled fault context, toggle the candidate while keeping the other heads fixed. This reveals its contribution when all its substitutes are masked. In 200 blind trials, we sampled 64 additional fault contexts per trial, masking each candidate with probability .75. Retrieval AUC rose from .50 to 1.00 for sets of 2, 4, and 8 interchangeable backups.

Part 1.2. Shared Upstream Inputs

Takeaway. To explain compensation for B's failure, retain the shared input A that the backup needs.

If the backup still relies on A, set S = {B} and retain A to explain B's failure. Removing A as well cuts off an input to the backup.

In IOI, B is the primary name-movers and A the upstream S-inhibition heads. We measure the IOI-margin loss from ablating the same eight documented backups in five batches of 512 new inputs.

Condition before ablating the backupsMargin loss from ablating the backups
B removed; A intact1.020
A and B removed0.206
B removed; A's intact activations replayed1.020

Removing A cuts the backups' effect by about 80%; replaying A restores it in all five batches. We also check replay across 28 fault and upstream-role combinations.

In a separate test, we removed only B and kept all eight backups in the computation. Freezing their activations at intact values reduced the logit margin by .557. The freeze test establishes that the backups help after B fails. The replay test shows why A belongs outside S for that question: the same backups depend on its input.

Part 1.3. Fault Set Selection

Takeaway. First specify which components fail, then test how the remaining components compensate.

To answer how to choose S, start with the failure you want to explain. If the relevant primary components are documented, remove them. Otherwise, find components whose removal impairs the target behavior on a detection split. In either case, leave shared inputs intact if the question is how the circuit responds to a downstream failure. Then fix S, rank candidates on calibration inputs, and freeze their responses on held-out inputs. Section 2.1 explains what S specifies. Appendix A.5 gives the protocol, and Appendix E.2 shows the split in practice for the induction study.

In IOI, choosing the right primary heads matters more than listing them all. In the original Table 21, taking the top one, two, or three known primary heads by ablation energy gives backup retrieval AUCs of .94, .97, and .94. A separate fault-set control in Appendix C.2 gives an AUC of about .40 for incorrect fault sets matched on behavioral damage, Fisher displacement, and layer position.

AuthorsOpenReview ↗

Response to RNmd (2/4): What CoAx Adds

Part 2. What CoAx Adds (W2)

CoAx finds heads whose effects grow under a specified fault, then tests whether their responses help and whether those heads recover what the circuit misses. On GPT-Neo year comparison, freezing the CoAx set causes a larger accuracy drop than freezing the two-state AtP* set Part 2.1. We separate the gains from comparing both states and measuring finite effects, then test whether the verified heads complete the circuit Part 2.2.

Part 2.1. Repair Gains

Takeaway. In year comparison, the CoAx set supplies stronger compensation than the two-state AtP* set.

In GPT-Neo-1.3B year comparison (adapted from Hanna et al., 2023), the fault alone leaves 78.6% accuracy. With the primary head still removed, freezing CoAx's selection lowers accuracy to 0.8%.

Lower accuracy after freezing means the blocked response helped more. To check whether the heads already mattered, we measure I by ablating the set in the intact model and R by blocking only its response after the fault. Both measure losses in later-minus-non-later year probability mass on the same held-out inputs. The table reports three-run means.

SelectionIntact task contribution IFreeze effect RAccuracy after freezing
CoAx signed0.9782+0.7920.8%
Two-state AtP*0.9207+0.31629.4%
Intact0.9771-0.28594.1%
Conditional0.9771+0.53112.4%
CoAx ratio−0.0041−0.29694.7%

① Comparison with two-state AtP*. The CoAx set has a higher R in all three paired runs. Mean R is 2.5 times that of AtP*, with ratios of 2.4–2.7 across three bins covering all tested year thresholds Reviewer d7Tf, Part 1.1.

② Similar intact contributions, opposite repair effects. CoAx and Intact have nearly identical I. After the fault, blocking the CoAx response harms year prediction, while blocking the Intact response improves it. Heads that are equally important in the intact model can therefore differ as backups.

Part 2.2. Measurement and Validation

Takeaway. The two-state comparison isolates newly important heads; freezing and completion establish what those heads do.

Wang et al. (2023) identified IOI backups, and Rushing & Nanda (2024) explained how effects such as LayerNorm rescaling contribute to repair. CoAx starts from a specified fault. It measures which components gain influence, then tests whether those components help and account for a missing part of the circuit's response.

① Why subtract the intact effect? Scoring only the faulted model conflates heads that were already important with backups that become important after the fault. We keep the finite ablations and Fisher metric fixed. The only change is subtracting each head's intact effect. In IOI, this raises backup retrieval AUC from .758 to .941. Freezing under the same fault gives R = .027 for the faulted-state ranking and .899 for CoAx (submitted Figure 3(c), revised Figure 4(c)). Subtracting the intact effect changes which heads are selected, and their responses provide more verified repair.

Table 7 of the submission tests a separate question: does CoAx still outperform intact-only ranking when we replace zero ablation with other interventions? Across the four replacement schemes, CoAx achieves .941–.974 retrieval AUC, versus .603–.753 for intact-only ablation ranking.

② Two-state AtP* also compares both states. What does direct measurement add? For the metrics in section 3 and Appendix C, CoAx and two-state AtP* do achieve similar AUC on known IOI backups (.941 vs. .938).

In Pythia-410M induction, we hold the task score fixed. Among downstream heads, the two methods select the same heads where their estimates agree. Where they disagree, direct ablation raises mean single-head freeze loss by .023–.130 across five runs. We see no gain over the full pool. Thus calibration disagreement helps target direct measurement among heads that can respond. See Reviewer pi2G, Part 2.1 for all five pairs.

③ What can the heads verified by freezing explain? In the original IOI experiment, adding the CoAx set tested by freezing reduces the fault-response gap D from .474 to .236. In Pythia-1.4B induction, freezing the CoAx set lowers the task score in all five runs. Adding those same heads to each run's base circuit reduces mean D from .1512 to .0718. Together, the interventions show what the retrieval AUC cannot: the selected response helps the full model and accounts for part of what the base circuit misses.

The historical IOI direct-logit score illustrates the distinction. It ranks the known backups slightly better (.971 AUC), using the IO−S direction, but its top eight give a poorer match to the documented knockout (margin distance 1.33 vs. .99; Appendix C.1.5).

AuthorsOpenReview ↗

Response to RNmd (3/4): Neurons and Features

Part 3. Neurons and Features (W3 / Q1)

We first verify neuron-level repair in GPT-2 and Pythia Part 3.1, then report scans of over ten thousand candidates and show how cached residuals let us score a sparse dictionary at the final layer Part 3.2.

Part 3.1. Neuron-Level Repair

Takeaway. CoAx identifies beneficial neuron responses in both GPT-2 and Pythia.

For GPT-2 IOI at L10 and natural text at L11, freezing the neurons selected by CoAx causes a larger mean task loss than freezing the neurons selected by two-state AtP*. In Pythia, freezing the CoAx neurons hurts performance in all five runs. We scan 3,072 candidates in each GPT-2 setting and evaluate each setting three times. For Pythia-410M, we scan 4,096 candidates and evaluate five times. The table reports mean freeze effects ± SD. Positive values mean the blocked response helped; compare methods within each task.

MethodGPT-2 IOI / L10GPT-2 natural text / L11Pythia-410M natural text
CoAx signed+0.0872±0.013+0.0493±0.011+0.0124±0.0048
CoAx ratio+0.0617±0.0066+0.00233±0.0092+0.00072±0.0016
Two-state AtP*+0.0618±0.011+0.0166±0.013+0.00577±0.008
Conditional+0.0862±0.0083-0.00674±0.032+0.012±0.0084
Intact+0.0588±0.022-0.0588±0.023+0.0128±0.0081

For these two GPT-2 queries, freezing the CoAx sets hurts more than freezing controls matched on layer and intact ablation energy. The full layer comparison in Appendix E.5 shows that the strongest responses occur at different layers for the two tasks; for IOI at L11, the signed set's mean R is negative. In Pythia, signed and Intact have similar means, but CoAx exceeds two-state AtP* and its response remains beneficial with final normalization gain fixed. These controls distinguish helpful neuron responses from score increases alone.

We can also verify a single neuron. On 61 new articles, clamping GPT-2 neuron 2362 in L11 to its intact activation lowers the mean log probability of the correct next token by .0187. Its direct contribution to the correct token rises from .0143 to .0313 after the fault, consistent with the direction of the freezing result.

Part 3.2. Scale and Cost

Takeaway. Exact scans cover over ten thousand neurons and thousands of active SAE features.

For thousands of neurons, we benchmark up to 12,288 candidates. With U candidates, scoring needs one ablation per candidate in each state plus two reference forward passes: 2U+2 states in all. There is no need to enumerate candidate pairs as U grows. All runs below use 96 calibration inputs. Neuron times include scoring, baselines, and validation; SAE times include scoring and validation.

Model / task / unitCandidate poolEnd-to-end time
GPT-2 / years / neuronsL9: 3,07251.5 min
GPT-2 / years / neuronsL9–10: 6,144110.2 min
GPT-2 / years / neuronsL8–11: 12,288152.7 min
GPT-2 / final-layer SAE24,576 dictionary entries; 6,374–6,622 active287.5–294.8 s

For sparse dictionary features, we score the 6.4–6.6k features active on calibration inputs in a 24,576-entry final-layer SAE. We can compute these scores exactly from cached residuals by recomputing only the output. The same two-state comparison applies to individually ablatable crosscoder features, but an intervention at an intermediate layer still requires running the layers that follow it.

Reviewer_pi2GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper deals with conditional circuit completion, which finds components whose causal importance increases when a given primary set is removed. The proposed method, COAX, involves performing actual ablations of each candidate component both before and after the removal of the primary set. It then measures the magnitude of these ablation effects within the same Fisher geometry of the clean output distribution and ranks the candidates based on the observed increase in importance. Theoretically, it formulates the informational limitations of a limited type of intact-model attribution score and links the change in the ablation effect vector to the interaction sum with the removed primary set. Experimentally, the study evaluates the performance in terms of backup recovery, self-repair, and circuit completion.

Critical strengths
  1. It raises a meaningful issue regarding the scope of validity of causal explanation. It is important to note the perspective that circuit explanations must be specified under which intervention conditions they are valid. By defining the failure conditions under analysis, the authors demonstrate that a circuit which effectively explains behavior under normal operating conditions may prove incomplete when specific components are removed. This point is relevant to how mechanistic interpretability methods are validated.
  2. The proposed method is defined well. Given a primary set, COAX can be executed via forward evaluation without task labels or gradients. Since both types of ablation effects are measured within the same output geometry, the comparisons remain consistent.
  3. It also explores the applicability of the approach to various mechanisms and models.
Critical weaknesses
  1. The contribution needs more explanation to compare with existing research. Previous IOI studies have already discovered backups after removing the primary name-mover heads, Hydra analyzed compensation, and subsequent circuit completeness studies addressed missing redundant paths. Consequently, the novelty of this paper is highly dependent on the proposed scoring rule and a corresponding verification framework. The approach is notable for relying solely on forward evaluations and requiring no task labels. However, two-state AtP*, which compares attribution before and after primary set removal, and the amplification-ratio variant, which uses the relative increase in the ablation effect, also showed competitive results. It therefore remains unclear whether COAX’s unique additional analytical capabilities have been sufficiently demonstrated or its practical trade-offs against these strongest alternatives adequately characterized.

  2. An increase in causal influence does not, by itself, demonstrate meaningful "self-repair." Since COAX measures the increase in the magnitude of an ablation effect, a positive score does not necessarily imply that the component in question restores the target behavior. The IOI freezing and retained-circuit tests, together with the held-out cross-mechanism repair analysis, provide meaningful evidence to bridge this gap. However, evaluations across multiple models measure the performance degradation when the primary set and selected components are ablated together. Such degradation can occur with components that were already important in the intact model, so this cross-model endpoint alone does not establish repair-specific contributions or retained-circuit completeness. This represents a significant gap between the empirical quantities measured and the broader cross-model conclusions drawn in the paper.

  3. The cross-mechanism analysis shows positive but modest agreement with causal repair. The current rationale supports localization procedures useful under specific conditions more convincingly than a general explanation of how causal responsibility is reconstructed after failure.

  4. Text overlaps with graphical elements in Figures 1 and 7. Figure 1 is particularly overcrowded, with too much information packed into a limited space.

Questions for authors
  1. What additional analytical capabilities, mechanistic insights, or practical benefits does COAX offer over primary-removal analysis, two-state AtP*, and amplification-ratio baselines? Under comparable experimental conditions, when does directly measuring ablation effects meaningfully improve the analysis relative to two-state AtP*?

  2. In the non-GPT-2 models studied, can you show that the selected components contribute specifically to repair, beyond their general contribution to task performance? Could the activation-freezing or retained-circuit tests already used for IOI be applied to the cross-model experiments?

  3. How do COAX’s repair predictions compare with those of two-state AtP* and the amplification-ratio variant on held-out inputs across the evaluated mechanisms? Can you provide additional qualitative examples of both successes and failures to clarify when an increase in the magnitude of an ablation effect reflects beneficial compensation?

10 comments
AuthorsOpenReview ↗

Response to pi2G (1/5): Repair Beyond Intact Contributions

Thank you for asking under which faults a circuit explanation remains valid and for recognizing the value of a common output geometry and validation across models and mechanisms. We address whether the responses help repair in new models Part 1, when direct ablation is worthwhile Part 2, held-out prediction and mechanism examples Part 3, and figure revisions Part 4.

Part 1. Repair and Completion (W2 / Q2)

The original joint-ablation test measures total contribution, including effects already present before the fault. In models beyond GPT-2, we freeze the selected heads' responses to test whether their changes help Part 1.1. We compare sets with similar intact contributions to check whether CoAx simply picks heads that already mattered Part 1.2. We then add the same verified heads to the base circuit. In Pythia-1.4B, this reduces the fault-response gap from .1512 to .0718 Part 1.3.

Part 1.1. Activation Freezing

Takeaway. CoAx finds beneficial responses to faults in non-GPT-2 models.

To test repair-specific contributions, we use activation-freezing to keep the components in the computation and block only their activation changes after the fault. We generate our own subject–verb agreement pairs (Marvin & Linzen, 2018) and adapt the year-span prompts of Hanna et al. (2023), where a valid end year must exceed the start year.

In GPT-Neo-1.3B year comparison, the model ranks a valid year first on 78.6% of inputs after the primary head is removed. We keep that head ablated and clamp each method's selected heads to their intact activations. Accuracy falls to 0.8% when we freeze CoAx's selection, compared with 29.4% for two-state AtP* and 12.4% for Conditional. In the same model, freezing CoAx's agreement selection also lowers is/are accuracy from 62.3% to 51.3%.

Across five calibration-qualified queries in four non-GPT-2 models, R and the gains over controls matched on layer and intact ablation energy are positive in all 19 held-out runs. All 19 freeze effects remain positive with final normalization gain fixed, so changes in that gain alone cannot explain the loss.

Part 1.2. Matched Intact Contributions

Takeaway. CoAx's selection contributes more to repair even when intact task contributions are similar.

① Similar intact importance, opposite responses after the fault. To distinguish repair from general contribution to task performance in GPT-Neo year comparison, we measure I, the loss from ablating the whole set in the intact model, and R, the loss from blocking the whole set's response after the fault. Both use the gap in probability mass between later and non-later two-digit endings.

SelectionIntact task contribution IFreeze effect R
CoAx signed0.9782+0.792
Two-state AtP*0.9207+0.316
Intact0.9771-0.285
Conditional0.9771+0.531

Ablating the CoAx and Intact sets in the intact model causes nearly the same loss. After the fault, freezing the CoAx set lowers the score by .792, while freezing the Intact set raises it by .285. An intact-only loss would treat these sets alike even though one helps and the other interferes under the same fault.

② Matched controls across three architectures. In a separate repeated-token induction test, we construct 50 controls for each CoAx set on calibration inputs with similar intact task contributions and layer composition. We compare freeze effects on 512 held-out inputs. Each row reports mean ± SD over five runs.

ModelRuns × test inputsΔR over intact-contribution-matched controlsRuns with positive ΔR
Pythia-410M5×512+0.084±0.0295/5
Gemma-2-2B5×512+0.297±0.2365/5
Llama-3.1-8B5×512+0.131±0.0725/5

The paired mean R is higher for CoAx in all 15 runs. Because the intact effects of CoAx's set and the controls still differ by 0.1%–10.5% on held-out inputs, we rescale the controls' R by the ratio of intact effects on calibration inputs. Mean R gains remain +.077, +.290, and +.131 for Pythia, Gemma, and Llama. The gain therefore survives an adjustment for residual intact differences.

AuthorsOpenReview ↗

Response to pi2G (2/5): Retained-Circuit Completion (W2 / Q2)

Part 1.3. Retained-Circuit Completion

Takeaway. Adding the heads verified by freezing makes the circuit reproduce more of the full model's fault response.

For the retained-circuit tests, we ask whether the heads that help the full model explain part of the response that the base circuit misses. In Pythia-1.4B induction, we freeze each CoAx set in the full model and add the same set to the 75% base circuit. The fault-response gap D falls from .1512 to .0718, a 52.5% reduction.

We retain the top 75% or 90% of candidate heads by intact effect, along with the fault set. Each method adds ten heads from outside the base circuit. D is the absolute difference between the circuit's and the full model's mean score changes after the fault. Lower D means a closer match. The table reports five-run means for Pythia-1.4B.

Added set75% base circuit: D↓90% base circuit: D↓
Base circuit0.151220.01729
CoAx signed0.071810.00905
CoAx ratio0.088560.01013
Two-state AtP*0.083180.00965
Conditional0.071070.00912
Intact0.086750.01266
Random0.134170.01318

Conditional and signed CoAx give nearly the same D at both fractions. D measures how well the circuit reproduces the response, so we freeze the added heads separately to test whether that response helps the task.

① Pythia-1.4B: the same heads help the task and reduce circuit error. At 75% retention, the CoAx set's ablation effect is 5.2 times larger after the fault than before it. In each run below, we freeze that run's selected heads in the full model, then add those same heads to the base circuit.

SeedSame-set freeze effect RD before → afterMean absolute error per input before → after
1+0.0590.154 → 0.0600.196 → 0.117
8+0.0510.140 → 0.0610.174 → 0.116
15+0.0140.102 → 0.0550.154 → 0.127
22+0.0350.159 → 0.0850.201 → 0.144
42+0.0710.202 → 0.0980.243 → 0.154

Freezing hurts the task in all five runs (mean R = .0461) and more than freezing the two-state AtP* set in every run (+.0179 on average). Mean error per input drops by 32% (.1936 to .1317), so the lower D does not just come from positive and negative errors canceling across inputs. Appendix E.4 pairs the cross-model freeze summary (Table 34) with the five-run completion plot (Figure 12).

② GPT-Neo and Llama 8B: both freezing and completion replicate. We run each model three times at the 75% base fraction. Freezing the CoAx set hurts the task in every run. Adding the respective sets back reduces mean D from .486 to .329 for GPT-Neo and from .230 to .172 for Llama 8B.

③ Per-input error across five models. In all five induction models whose bases pass calibration at both fractions, adding the selected heads reduces per-input error at both fractions. Error falls in all 30 paired runs, with paired resampling supporting a decrease in 27. Fidelity and beneficial repair are separate: at 90% retention in Pythia-410M, D falls from .0032 to .0013, but mean R is slightly negative (−.0054; Table 35). The same-set freeze test is what tells us whether the response recovered by completion helps the task.

AuthorsOpenReview ↗

Response to pi2G (3/5): When Direct Ablation Helps

Part 2. Value of Direct Ablation (W1 / Q1)

We first compare the components selected by CoAx and two-state AtP*. We then use the same task metric to test when direct ablation selects better heads than a local approximation Part 2.1. Next we explain how to separate intact effects from increases after the fault Part 2.2 and how much time scoring takes for heads and neurons Part 2.3.

Part 2.1. Finite vs. Local Effects

Takeaway. Direct ablation finds more useful repair heads among downstream candidates where the local estimate disagrees with the finite effect.

① Joint freezing: how much repair do the heads selected by CoAx and two-state AtP* provide? On GPT-Neo year comparison, freezing CoAx's set hurts the task more than freezing the two-state AtP* set (R = .792 vs. .316). All 11 paired runs across Pythia year comparison, GPT-Neo year comparison, and Llama 8B agreement support a higher R for CoAx.

R is the task-score drop from freezing the selected set's response while keeping the fault fixed. Positive R means the response helps. For agreement, the score is the logit difference between the correct and incorrect verbs. For year comparison, it is the probability mass on later minus non-later year endings. We report means across runs, each with 512 held-out inputs.

MethodPythia-410M years / 3 runsGPT-Neo-1.3B agreement / 3 runsGPT-Neo-1.3B years / 3 runsLlama-3.1-8B agreement / 5 runsLlama-2-13B agreement / 5 runs
CoAx signed+0.0268+0.557+0.792+0.0471+0.0147
CoAx ratio+0.0224-1.36-0.296+0.0202+0.0123
Two-state AtP-0.0058+0.5385+0.316-0.0274+0.0134
Two-state AtP*-0.0058+0.547+0.316-0.0271+0.0135
Conditional-0.0024+0.241+0.531+0.0346+0.0133
Intact-0.0118+0.56-0.285+0.0341+0.0045

CoAx signed has a higher mean R than ratio on all five queries and the highest mean R among the six methods on four.

② Single-head freezing: when does direct measurement help? The joint test compares complete selectors. To isolate directly measuring ablation effects in Pythia-410M induction, we fix the task score and compare finite ablation with two-state AtP* within each candidate pool. We group heads by calibration disagreement, then freeze each ranking's top five within each group, one at a time, on 64 held-out sequences per run.

In the low-disagreement group, both methods select the same heads in all five runs. For the high-disagreement group, the table shows the difference in mean single-head R (finite ablation minus two-state AtP*; manuscript Table 4).

SeedRuleAll headsDownstream heads
1Development0.000+0.108
8Development0.000+0.121
15Fixed−0.011+0.092
22Fixed0.000+0.130
42Fixed0.000+0.023

The gain appears in all five runs in the downstream pool and in none in the full pool. In this test, downstream position determines whether a head's activation can respond to the fault; disagreement then tells us where the estimator changes which responses we find.

Reviewer_kz6vOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COAX, a conditional co-ablation method for identifying backup components that become important after a primary circuit is removed. The method is well motivated by the self-repair phenomenon in transformers and shows promising results on the GPT-2-small IOI circuit.

I find this paper well motivated and consider it a promising working paper. However, I do not think the current empirical evidence is sufficient to support its broader claims. In particular, more convincing validation across different models, tasks, and independently identified backup circuits would substantially strengthen the contribution. I therefore lean toward weak reject at this stage.

Critical strengths
  1. I think the paper is well motivated. The paper addresses a well-motivated problem in mechanistic interpretability. The limitations of intact state attribution in the presence of self-repair are clearly explained.
  2. The proposed method is simple and intuitive, with a reasonable theoretical formulation.
  3. The experiments include several useful controls and causal interventions to validate the identified backup components.
Critical weaknesses
  1. The main concern is the limited empirical validation. Most of the convincing mechanistic evidence comes from the well-studied GPT-2-small IOI circuit. Although the authors include experiments on other models, these primarily evaluate induction circuit completion rather than recovering independently verified backup mechanisms. In the current version, I don't believe the method has good generalization.
  2. The cross-mechanism evidence is relatively weak, with an average Spearman correlation of only 0.123 between the proposed score and causal repair.
  3. The improvement over a two-state baseline is marginal (0.941 vs. 0.938 AUC). Very Small Gain.
Questions for authors
  1. Your figures contain too much information, making them hard to follow.
  2. The writing needs improvement; for example, nobody writes as: "with a hierarchical mean of +0.123 and a nested 95% CI of [+0.047, +0.201]" Line 425. This is ai generated content, please polish the words.
7 comments
AuthorsOpenReview ↗

Response to kz6v (1/2): Cross-Model Repair

Thank you for your careful review and for recognizing the clear motivation, the intuitive, theoretically grounded method, and the causal validation. We first present independent interventions on new models and tasks Part 1, then discuss prediction across mechanisms and repair gains against a strong baseline Part 2, and finally address figures and writing Part 3.

Part 1. Cross-Model Repair (W1 / W3: functional gains)

IOI was our labeled benchmark because the primary and backup heads are documented (Wang et al., 2023). To test generalization beyond IOI and induction, we verify repair on agreement and year comparison in four non-GPT-2 models Part 1.1. We select heads on calibration inputs and verify their responses by intervention on held-out inputs. We also trace a repair response to a single head in natural text and verify circuit completion across three models Part 1.2.

Part 1.1. Agreement and Year Comparison (W1 / W3: functional gains)

Takeaway. CoAx identifies responses that help agreement and year comparison across four non-GPT-2 models.

We generate agreement pairs (Marvin & Linzen, 2018) and adapt year-span prompts (Hanna et al., 2023), requiring a later end year. In GPT-Neo-1.3B year comparison, the model ranks a valid year first on 78.6% of inputs after the primary head is removed. We keep that head ablated and clamp each method's selected heads to their intact activations. Accuracy falls to 0.8% when we freeze CoAx's selection, compared with 29.4% for two-state AtP* and 12.4% for Conditional. In the same model, freezing CoAx's agreement selection also lowers accuracy from 62.3% to 51.3%.

R is the task-score drop from freezing the selected set's response while keeping the fault fixed. Positive R means the response helps. Agreement uses the correct-verb logit margin; years use the probability mass on later minus non-later endings. Compare methods within each query. Each run uses 512 held-out inputs.

MethodPythia-410M years / 3 runsGPT-Neo-1.3B agreement / 3 runsGPT-Neo-1.3B years / 3 runsLlama-3.1-8B agreement / 5 runsLlama-2-13B agreement / 5 runs
CoAx signed+0.0268+0.557+0.792+0.0471+0.0147
CoAx ratio+0.0224-1.36-0.296+0.0202+0.0123
Two-state AtP-0.0058+0.5385+0.316-0.0274+0.0134
Two-state AtP*-0.0058+0.547+0.316-0.0271+0.0135
Conditional-0.0024+0.241+0.531+0.0346+0.0133
Intact-0.0118+0.56-0.285+0.0341+0.0045

Across five calibration-qualified faults, CoAx's R is positive and exceeds the matched-control mean in all 19 held-out runs. It has the highest mean R on four faults and remains positive in all 19 runs with final normalization gain fixed. On Llama 8B agreement, CoAx's response helps (R = +.047), while the two-state AtP* set's response hurts (R = −.027). All six methods use the same downstream pool for both Llamas. Freezing keeps the heads active and isolates their change after the fault.

Part 1.2. Repair Paths and Completion (W1)

Takeaway. The interventions verify repair beyond intact contributions and identify a specific head that responds in natural text.

① Are equally important heads also equally useful after a fault? In GPT-Neo year comparison, ablating the CoAx and Intact sets in the intact model causes nearly identical losses (.978 vs. .977). After the fault, freezing CoAx's set lowers the score by .792, while freezing the Intact set raises it by .285. In separate induction tests, CoAx exceeds the mean response of controls matched on intact contribution and layer composition in all 15 runs across three architectures Reviewer pi2G, Part 1.2.

② Which head compensates? For independently verified backup mechanisms, we trace a response in Pythia-410M. After removing L4H11, only L8H13 in CoAx's set is downstream and can change its activation. Freezing the set isolates L8H13's response and lowers the next-token score in all nine runs across new WikiText-2 articles, HellaSwag contexts, and longer WikiText-2 windows (mean R = .260, .922, and .149). Two-state AtP* selects only heads at the fault's layer or earlier and gives zero freeze effect Reviewer pi2G, Part 3.2. With faults and selections fixed, both Pythia and Gemma retain positive R across all three input settings Reviewer d7Tf, Part 1.2.

③ Do the verified heads explain the missing fault response? In Pythia-1.4B induction, freezing CoAx's set hurts performance in all five runs. Adding the same heads to each run's base circuit reduces the mean fault-response gap from .1512 to .0718 and per-input error by 32%. In three runs each for GPT-Neo and Llama 8B, freezing also hurts and adding the same heads reduces the mean gap Reviewer pi2G, Part 1.3. Thus the heads help the full model and explain part of the response missing from the circuit.

AuthorsOpenReview ↗

Response to kz6v (2/2): Prediction, Repair, and Presentation

Part 2. Prediction and Repair (W2 / W3)

We compare the rankings against the same held-out freeze effects Part 2.1, then ask whether similar IOI AUCs translate into equally helpful responses after a fault Part 2.2.

Part 2.1. Cross-Mechanism Prediction (W2)

Takeaway. Cross-mechanism agreement is modest; freezing tells us which changes in influence correspond to helpful responses.

The original average Spearman correlation is modest (.123), with positive correlations in 11 of 12 held-out cases. We repeated the test over 60 queries. Signed CoAx vs. two-state AtP* is .120 vs. .075 on duplicate-token, .166 vs. .172 on name-mover, .072 vs. .115 on S-inhibition, and .008 vs. .218 on induction. On duplicate-token, signed leads AtP* in all five repeats after averaging the three settings; CoAx ratio also leads the external baselines there. That gain survives changing how we rank the same two-state measurements Reviewer pi2G, Part 3.1.

Why does the score only partly predict repair? A downstream fault can increase the model's dependence on an upstream head without changing that head's activation. Its ablation effect grows, but freezing it to the same value has no effect. Restricting the pool to heads whose activations can respond raises CoAx's correlation in all four mechanisms, most clearly for name-movers (.166 to .323; ratio .178 to .410; AtP* .172 to .314). The diagnostic explains why we use growth to locate changes in dependence and freezing to verify which responses help. The baseline comparison above uses the full pool.

Part 2.2. Gains over AtP* (W3)

Takeaway. CoAx finds more beneficial repair responses than two-state AtP* in the year comparison and induction tests.

① Does the small AUC gap imply a small functional gain? The .941 vs. .938 AUCs rank eight known IOI backups; they are close. Under the same GPT-Neo year fault, accuracy starts at 78.6% and falls to 0.8% when we freeze CoAx's set, versus 29.4% with the two-state baseline set. The blocked response contributes 2.5 times as much to the task score (R = .792 vs. .316), with a larger loss in all three runs. Pythia-410M induction also favors CoAx in all five paired runs (+.318 mean R difference). Thus similar label recovery does not imply equally useful responses after a fault.

② What does direct ablation itself add? The comparisons above test whole selectors. In Pythia-410M induction, we fix the task score, change only the estimator, and freeze each ranking's top five heads individually. For downstream candidates with low calibration disagreement, both methods select the same heads in all five runs. With high disagreement, direct ablation selects heads with higher mean R in every run (+.023 to +.130), including all three fixed-rule repeats. The full pool shows no gain (Table 4; Reviewer pi2G, Part 2.1). Direct measurement is most useful here when estimates disagree for heads that can actually respond.

Part 3. Figures and Text (Q1 / Q2)

Takeaway. The revision separates the figures by purpose and states the correlation directly.

Figures. To address too much information in the overview, we separated the IOI results from the self-repair illustration. Figure 1 now sits beside the opening of the Introduction. Figure 2 shows the three IOI validation panels, with vector plots in (b) and (c). Section 2 explains the method. Figure 5 retains the original comparisons and labels (b) as joint ablation. The baselines are in Appendix E.1–E.2; Table 3 and Appendix E.3–E.4 cover freezing and completion.

Text. Thank you for flagging Line 425. We wrote the original abstract, introduction, methods, and theory, using AI to check wording and flow. AI also helped draft parts of the experimental discussion, which at least two team members reviewed in multiple rounds. The revised AI use statement now discloses assistance with drafting parts of the experimental discussion. For the revision, we read the manuscript ourselves from start to finish, checked the order of the argument, and verified the experimental claims against the figures and tables. Section 5 now states the flagged result directly: the equally weighted mean Spearman correlation across the four mechanisms is .123. The Figure 5(a) caption identifies the plotted interval, and Appendix E.1 explains how it is computed. Section 5 also explains why we freeze responses separately from joint ablation.

Thank you again for pressing for independent validation and pointing out where the presentation was hard to follow. Your questions helped us verify a repair response at the level of a single downstream head and explain how it differs from total contribution under joint ablation. We welcome your feedback on the revision.

Thank you for your reply. I'd like to maintain my original score.

Reviewer_d7TfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

CoAx introduces a method for finding backup components of circuits. That is, those that compensate for the removal of primary components when they are ablated by circuit discovery methods. They present results on the IOI circuit in GPT-2 small and induction in other models.

Critical strengths

The method is simple and discovers backup components in the IOI circuit.

Critical weaknesses

Hard to read, has a very Claude style, lots of colons, “load-bearing” is a classic Claude description, gerund header titles are also typical of LLM writing. Personally, I think using ai for writing is good if it can accelerate the rate of producing good research but I would be careful that it is actually easy for a human to read.

Figure 1 is extremely dense and has lots of overlapping text and figures, quite hard to parse.

It is only tested on IOI and induction, can we have results on other circuits?

Questions for authors

Can we have y axis labels on Figure 3(a)?

Can we have baselines for Figure 4?

5 comments
AuthorsOpenReview ↗

Response to d7Tf (1/3): Beyond IOI and Induction

Thank you for your constructive feedback and for recognizing the simplicity of the method and its ability to identify IOI backups. We first show computations beyond IOI and induction Part 1, then give Figure 4 baselines and functional comparisons Part 2, and finally describe specific changes to the writing and figures Part 3.

Part 1. Beyond IOI and Induction (W3)

On other circuits, CoAx identifies responses that help with subject–verb agreement and year comparison Part 1.1. IOI has published labels for both the primary and backup heads (Wang et al., 2023), so it gives us a direct test of backup recovery. Repeated-token induction (Olsson et al., 2022) lets us test the same computation across architectures. We now vary the task and, in a separate test, keep the selected heads fixed while varying the input text Part 1.2.

Part 1.1. Agreement and Year Comparison

Takeaway. CoAx finds components that compensate for faults in both agreement and year comparison.

We generate our own subject–verb agreement pairs (Marvin & Linzen, 2018) and adapt the year-span prompts of Hanna et al. (2023), where a valid end year must exceed the start year.

For each GPT-Neo-1.3B task, we remove its primary head and keep the CoAx heads active but fixed at their intact activations. Accuracy on the is/are choice drops from 62.3% to 51.3%, and the proportion of inputs on which a valid year ranks first drops from 78.6% to 0.8%.

We measure the drop R in the correct-verb logit margin for agreement and in the probability-mass gap between later and non-later endings for year comparison. Compare R within each query. The last column compares R with controls matched on layer and intact-state ablation energy.

Model / computationFault SRuns × test inputsFreeze effect RGain over matched controls
gpt2-small / subject–verb agreementL0H103×512+1.75+1.98
gpt2-small / year comparisonL0H103×512+0.0776+0.408
pythia-410m / year comparisonL10H83×512+0.0268+0.0251
gpt-neo-1.3b / subject–verb agreementL1H113×512+0.557+0.576
gpt-neo-1.3b / year comparisonL0H83×512+0.792+0.76
llama-3.1-8b / subject–verb agreementL26H235×512+0.0471+0.0503
llama-2-13b / subject–verb agreementL35H385×512+0.0147+0.0136

Across seven model–task combinations that passed calibration in five models, blocking the CoAx response hurts performance in all 25 held-out runs. It also hurts more than blocking the matched controls in every run. The matching tests whether this difference can be explained by heads that were already important before the fault.

① GPT-Neo agreement: do the responses help the model change the verb with the subject? Within each sentence pair, we change the subject's grammatical number and keep the distractor fixed. Freezing the agreement set lowers the proportion of pairs with both answers correct from 24.6% to 9.9%. The selected responses help the model switch between is and are as the subject's number changes.

② GPT-Neo years: are the gains over AtP* confined to a few thresholds? Across three intervals covering all tested thresholds, CoAx's mean R is 2.4–2.7 times that of two-state AtP*. The gain holds across the tested range of start years.

Part 1.2. Inputs and Interventions

Takeaway. Under the same fault, the selected components still contribute to repair on new inputs without reselection.

① Transfer to new text. We keep the original fault and selected set fixed for each of Pythia-410M and Gemma-2-2B. We test next-token prediction on new WikiText-2 articles (Merity et al., 2017), HellaSwag contexts (Zellers et al., 2019), and 128-token WikiText-2 windows. We repeat each model–text combination three times. Freezing hurts next-token prediction in all 18 runs and causes a larger loss than the matched controls in every run. CoAx has the highest or joint-highest mean R in all six combinations. The selected sets thus remain useful on text that played no role in selection.

② The same test applies to two-head faults and individual neurons. For subject–verb agreement in GPT-2, after ablating two heads together, freezing CoAx's selected set reduces is/are accuracy from 87.0% to 53.4%. On natural text, freezing CoAx's selected GPT-2 neuron 2362 in L11 lowers the correct next token's mean log probability by .0187. The neuron tests in GPT-2 and Pythia are in Reviewer RNmd, Part 3.1.

AuthorsOpenReview ↗

Response to d7Tf (2/3): Figure 4 Baselines and Repair

Part 2. Figure 4 Baselines (W3: new tasks / Q2)

We add the requested baselines for Figure 4 for single-head repair prediction and joint ablation Part 2.1. Both original plots remain in Figure 5 of the revision, with panel (b) labeled as joint ablation. Appendix E.1–E.2 reports the full comparisons. We also test whether the selected heads' responses help after the fault and whether the same heads fill a gap in the circuit Part 2.2.

Part 2.1. Original Metrics (Q2)

Takeaway. Both CoAx rankings lead on duplicate-token, and signed CoAx exceeds random joint-ablation controls in all eight models.

① Figure 4(a): predicting repair by individual heads. We retain the original metric: on the same candidate pool, we compute the Spearman correlation between each method's scores and single-head freeze effects measured on held-out inputs. We test three settings per mechanism with five repeats each, for 60 queries.

MechanismCoAx signedCoAx ratioTwo-state AtP*
name-mover0.1660.1780.172
s-inhibition0.0720.0580.115
induction0.0080.0500.218
duplicate-token0.1200.1180.075

Signed and ratio reuse the same two-state ablations, ranking absolute growth and relative amplification, respectively. Both lead the six tested selectors on duplicate-token, suggesting that the advantage there survives the choice of ranking. Name-mover is comparable across the three shown, and AtP* leads both CoAx rankings on induction and S-inhibition. The score measures how much a head's effect grows; freezing tests whether its activation changes help.

② Figure 4(b): joint ablation. We keep the ten-head budget, random reference, and next-induction control. This replication uses separate detection/calibration/test inputs (64/96/256) and full-vocabulary Fisher. The table gives joint-ablation loss ratios to random (1.00), averaged over five runs.

ModelCoAxRatioTwo-state AtP*Next-induction
pythia-160m2.181.412.231.95
pythia-410m9.211.887.446.55
pythia-1.4b4.2627.073.32
gpt-neo-1.3b1.410.475.605.38
gemma-2-2b3.092.191.712.54
qwen2.5-7b7.081.344.11.79
olmo-2-7b2.833.144.323.93
llama-3.1-8b9.44.146.121.47

Signed CoAx exceeds random in all eight models, next-induction in six, and two-state AtP* in four. This shows that the selected heads matter under joint removal across models; the drop includes their intact contributions. To isolate the repair response, we leave the heads active and freeze only their change after the fault.

Part 2.2. Repair and Completion (W3: new tasks / Q2)

Takeaway. The selected responses help on agreement and year comparison, and the verified heads improve circuit fidelity.

① Does blocking the response hurt prediction? In GPT-Neo-1.3B year comparison, the model ranks a valid year first on 78.6% of inputs after the primary head is removed. We keep that head ablated and clamp each method's selected heads to their intact activations. Accuracy falls to 0.8% when we freeze CoAx's selection, compared with 29.4% for two-state AtP* and 12.4% for Conditional.

R is the task-score drop from freezing under the fixed fault. Positive R means the blocked response helped. Agreement uses the correct-verb logit margin, and year comparison uses the probability mass on later minus non-later two-digit endings. Compare methods within each query. Each run uses 512 held-out inputs. Conditional denotes removed-state energy in the revision; Intact denotes intact-state energy.

MethodPythia-410M years / 3 runsGPT-Neo-1.3B agreement / 3 runsGPT-Neo-1.3B years / 3 runsLlama-3.1-8B agreement / 5 runsLlama-2-13B agreement / 5 runs
CoAx signed+0.0268+0.557+0.792+0.0471+0.0147
CoAx ratio+0.0224-1.36-0.296+0.0202+0.0123
Two-state AtP-0.0058+0.5385+0.316-0.0274+0.0134
Two-state AtP*-0.0058+0.547+0.316-0.0271+0.0135
Conditional-0.0024+0.241+0.531+0.0346+0.0133
Intact-0.0118+0.56-0.285+0.0341+0.0045

CoAx has the highest mean R on four of five queries. It beats two-state AtP* in all 11 paired runs across the two year queries and 8B agreement. The sign on 8B matters: blocking CoAx's set hurts, while blocking AtP*'s set helps (R = +.047 vs. −.027). CoAx selects heads whose changes favor the correct verb after the fault.

② Circuit completion in Pythia-1.4B induction. Freezing CoAx's selected responses lowers the task score in all five runs. Adding those same heads to each run's base circuit reduces the mean fault-response gap D from .1512 to .0718 and the per-input error by 32%. This verifies that the heads help the model and explain part of what the circuit misses. The five paired runs are in Reviewer pi2G, Part 1.3; Appendix E.3–E.4 gives the full freezing and completion results.

AuthorsOpenReview ↗

Response to d7Tf (3/3): Text and Figures

Part 3. Text and Figures (W1 / W2 / Q1)

Part 3.1. Text Revisions (W1)

Takeaway. We clarify where AI assisted and have rewritten the experimental discussion to make the evidence easier to follow.

Thank you for pointing out what made the paper hard to read. We wrote and revised the original abstract, introduction, methods, and theory ourselves, using AI in those sections to check wording and the flow of the argument. As the revised AI use statement notes, AI also helped draft parts of the experimental discussion. At least two team members reviewed that material in multiple rounds.

We then reviewed the full manuscript ourselves, checking the order of the argument, how each section leads to the next, and whether the experimental claims follow from the figures and tables. The methods now explain what the primary set specifies before introducing the score, and the theory ties the two propositions back to why we compare the two states. We also rewrote passages in the experimental discussion. In Section 3.3, for example, we replaced load-bearing with the actual result: with the primary heads removed, holding the selected heads at their intact activations lowers the IOI logit margin by about .90. The drop shows that their response helps preserve the margin. Section 5 distinguishes what joint ablation, freezing, and completion each measure. Throughout, we split sentences where colons obscured the reasoning and simplified headings such as “Completing the Circuit” to “Circuit Completion.” The paper title now uses “Recovers.”

Part 3.2. Figures 1 and 3(a) (W2 / Q1)

Takeaway. The revision separates the overview figures and puts the contribution scale beside the head bars.

Original Figure 1. To resolve overlapping text and figures, we split the overview. Revised Figure 1 shows self-repair beside the opening of the Introduction. Figure 2 gives the three IOI validation panels a full-width layout, with vector plots in (b) and (c). Section 2 explains the method.

Original Figure 3(a). Revised Figure 4(a) keeps the original head-level circuit diagram and labels the IO−S logit contribution magnitudes beside the head bars. Each row has its own linear scale, so all bars fit without clipping.

Thank you again for the specific examples in your review. They helped us make the evidence easier to follow in both the text and the figures. We welcome your feedback on the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.