acceptodds
Under review as a conference paper at ICLR 2027

Attribution Vergence: A Mechanistic Indicator across Grokking and Adversarial Transfer

Abstract

Connecting internal computational mechanisms to macroscopic model behaviors, such as grokking and adversarial transfer, is one of the central goals of mechanistic interpretation. However, while existing methods analyze isolated components or global representations, whether the structure of attributed contributions supporting model decisions admits a general macroscopic observable remains fundamentally open. We introduce Attribution Vergence, a measure that quantifies this structural reorganization across feature aggregation by measuring the change in R\'enyi-2 effective dimensionality. Under a unified Aggregation Node formulation encompassing attention and convolution, we ground the measurement via attribution preservation, axiomatic attribution compatibility conditions, an information-theoretic interpretation, and analysis on invariance, bounds, stability, and computation. Experiments on grokking and generative adversarial transfer show stronger behavioral responses than prior analysis tools and additional gains from the formulation. The profiles reveal perturbation-dependent response locations, source indicators shared across transfer targets, and attribution reorganization across layers after training accuracy saturates. Together, these results establish a computable measurement framework linking attribution organization to macroscopic behaviors, as a step toward a scientific understanding of AI models.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability, Attribution, Grokking, Adversarial Example

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qpUkOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Attribution Vergence, a measure of the change in Rényi-2 effective dimensionality of internal attributions before and after feature aggregation. The authors provide mathematical properties of the measure and evaluate its correlations with grokking progress and adversarial transferability.

Critical strengths
  • The paper studies an interesting problem: whether internal attribution organization can provide useful indicators of model behavior across different architectures and tasks.
  • The paper looks well-written and polished.
Critical weaknesses

I didn't identify any critical weaknesses.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for the positive assessment and for recognizing the strengths of our theoretical framework and empirical evaluation. In the revised manuscript, we further clarify the connections to related work, add theoretical analysis and experiments on the Fourier-circuit structure, and evaluate the sensitivity to attribution choices. Our Overall Comments also explain the broader scientific motivation and contribution of the measurement framework. We hope these additions and clarifications provide a clearer understanding of the work.

Dear Reviewer,

Thank you again for your positive assessment. Please let us know if you have any further questions or would like clarification on any aspect of the paper.

Our revision of the manuscript is marked in blue. The main updates are:

  1. Clarified the research context and significance of a common observable (Section 1).
  2. Added comparisons with Fourier-Gini progress measures (Section 4.2, Table 2).
  3. Added Fourier-based circuit associations and routing interventions to investigate task-dependent attribution organization (Section 4.2, Appendices E.3–E.4).
  4. Added robustness analyses across attribution references, objectives, and paths (Figure 5 in Section 4.2, Appendix G.2).
  5. Clarified the meaning and theoretical grounding of “local complexity” (Sections 2–3).
Reviewer_CjQyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Attribution Vergence, a metric that quantifies changes in the effective dimensionality of feature attributions across aggregation operations in neural networks. The metric can be computed across attention, convolution and linear layers. The paper evaluates this metric in two settings: grokking and adversarial transfer and shows stronger correlation with these phenomena than other baselines. The results suggest that changes in attribution organization can be a good indicator of generalization progress and adversarial transferability.

Critical strengths
  • The paper introduces an intuitive metric that quantifies how the effective dimensionality of feature attributions changes across aggregation operations and that shows how attributed contributions are reorganized during neural network computations.

  • The metric works across attention, convolution and linear layers, which makes it suitable to be used in a broad range of architectures.

  • The authors evaluate the metric across grokking and adversarial transfer, showing promising improvements over existing baselines. The experiments are carefully designed, with held-out evaluation, statistical controls, and several robustness analyses.

Critical weaknesses
  • The paper shows that atribution Vergence correlates with intersting phenomema in the field like grokking and adversarial transfer. However, it is unclear whether Vergence tells us something interesting about this phenomena. For example, for the grokking example, Vergence changes after training accuracy saturates and it is interpreted as a reorganization of computational complexity across layers. However, it would be interesting to show the relationship between Vergence and a known causal mechanism (e.g. in grokking). Even if vergence is presented as an indicator and not a causal explanation, I think making this kind of connection would really strengthen the paper.

  • In the definition of vergence, the attribution field S depends on how the attribution is computed. In my opinion it would be important to show how sensitive these experiments are with respect to different attribution baselines, objectives etc.

Questions for authors
  • Can the authors provide additional evidence connecting Attribution Vergence to known computational mechanisms? For example, in the grokking experiments, how does the evolution of vergence relate to the emergence of known circuits (e.g. fourier-based circuits in modular addition)?

  • How sensitive are the experimental results to the choice of attribution baseline, objective, and integration path?

5 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for recognizing the intuitive formulation, applicability across architectures, and careful evaluation with held-out testing, statistical controls, and robustness analyses. Below is our response.


Q: It is unclear whether Vergence tells us something interesting about this phenomena. It would be interesting to show the relationship between Vergence and a known causal mechanism. Can the authors provide additional evidence connecting Attribution Vergence to known computational mechanisms? For example, in the grokking experiments, how does the evolution of vergence relate to the emergence of known circuits (e.g. Fourier-based circuits in modular addition)?

A: Thanks to the reviewer for the suggestion. Overall, as stated in the Overall Comments, the goal of our paper is to first establish a new observable itself and then support concrete interpretations as one of the applications. In the revision, we supplemented a mechanism analysis with the support of Attribution Vergence and the Fourier concentration concept proposed by [1].

In the revised manuscript, we focus on two concrete observations from Appendix E.3 and investigate them in Appendix E.4

  1. Module-dependent directions on the Add P97 modular addition task. At Attention Layer 0, the rank associations of VA\mathbb V_{\mathbb A} with validation accuracy and embedding Fourier Gini are +0.817 and +0.898, respectively. However, at MLP layers, both associations are negative.
  2. Task-dependent directions at Attention L0. Across the three S5 tasks, the correlations at Attention L0 are negative with both validation accuracy, ranging from -0.890 to -0.723, and the two Fourier Gini statistics, ranging from -0.908 to -0.377 (Figure E.3).

Thus, increasing Fourier concentration can accompany different directions of VA\mathbb V_{\mathbb A}. This is reasonable since VA\mathbb V_{\mathbb A} is a general property.

To locate the source of these differences, we decompose VA\mathbb V_{\mathbb A} using the norm expressions from our bound analysis in two forms:

=\mathbb V_{\mathbb A}^{L_1}-\mathbb V_{\mathbb A}^{L_2}, \qquad \mathbb V_{\mathbb A}^{L_p} =2\log_2\frac{\|\mathbf S\|_p}{\|\mathbb A[\mathbf S]\|_p}, \quad p\in\{1,2\},$$ and, with $\boldsymbol v\in \mathbb R^{|\mathcal Z|},\ \boldsymbol v_z=\|\mathbf S_{:,z}\|_2,$ | Quantity | Add P97, first → final | S5 Product, first → final | | ---------------------------------- | ---------------------: | ------------------------: | | $\mathbb V_{\mathbb A}$ | 0.5047 → 0.8737 | 0.5178 → 0.0468 | | $\mathbb V_{\mathbb A}^{L_1}$ | 0.5235 → 0.7791 | 0.5224 → 0.0396 | |$\mathbb V_{\mathbb A}^{L_2}$ | 0.0188 → −0.0945 | 0.0047 → −0.0072 | | $\log_2\mathrm D(\boldsymbol v)$ | 0.5179 → 0.8515 | 0.5176 → 0.0483 | The results show that the difference stems from the distribution of token-wise attribution differences. We formulate the Merging vs. Role-Preserving Transport conjecture, grounded in the algebraic structure of the respective tasks, to provide an interpretation that matches the difference in Attention L0 $\mathbb{V}_A$ 1. Add P97 (Abelian): Symmetric operand roles and Fourier character multiplicativity ($\chi_k(a+b)=\chi_k(a)\chi_k(b)$) favor symmetric merging within heads, raising route balance ($\log_2 \mathrm D(\boldsymbol v) \to 1$) and $\mathbb V_{\mathbb A}$. 2. S5 Product (Non-abelian): Permutation composition requires ordered mapping via point-stabilizer ($S_4$) cosets, precluding symmetric pooling. Attention heads instead evolve into role-preserving selective transport *e.g.*, two heads dedicated to $a$, two to $b$, delegating non-linear composition to MLPs. This single-route specialization per head drives $\log_2 \mathrm D(\boldsymbol v) \to 0$ and $\mathbb V_{\mathbb A} \to 0$. We test the proposed organizations through targeted functional interventions. 1. We designed causal routing interventions on Attention L0, directly contrasting joint intra-head attention (where each head attends to both operands) against head-specialized role separation (where heads disjointly attend to a single operand). 2. The results show that commutative Add P97 strictly favors intra-head operand mixing across all seeds (30/30 pairwise loss advantage), whereas non-commutative S5 Product uniquely requires canonical head specialization matching its learned roles to retain 100% accuracy and minimal loss (5/5 seeds), confirming that their circuits reorganize toward early scalar merging versus role-preserving transport, respectively. Such results support the **Task-dependent Local Complexity Reorganization** (as our response to Xg85, it may not be the computational complexity) hypothesis, as the additive aggregation absorbs the additive processing of modular Add, but not S5 Product.
AuthorsOpenReview ↗

Response to Reviewer #2

On the sensitivity of attribution used

Q: In the definition of vergence, the attribution field S depends on how the attribution is computed. In my opinion it would be important to show how sensitive these experiments are with respect to different attribution baselines, objectives etc. How sensitive are the experimental results to the choice of attribution baseline, objective, and integration path?

A: Thanks to the reviewer for the suggestion. We agree that including only the sensitivity analysis through different integration steps may not be sufficient. We supplement the sensitivity analysis on the adversarial transfer task in our response to reviewer Xg85. Here we provide the results in the Grokking task.

For Grokking, Appendix G.2.1 compares zero and training-mean references, together with a log-probability objective, on common retrained states from all four tasks. Here, results are reported from the average statistics across 4 grokking tasks.

BaselineObjectivePhase type η2\eta^2Train Acc. ∥ρ∥\|\rho\|Train Acc. R2R^2Val. Acc. ∥ρ∥\|\rho\|Val. Acc. R2R^2Δ Val. Acc. R2R^2
ZeroLog odds0.81650.64660.28790.83990.7824—
Training meanLog odds0.77790.65000.25860.80240.7343−0.0482
ZeroLog probability0.81300.68030.27810.82160.8149+0.0325

The main validation association survives both changes, with mean $$rho remaining between 0.80 and 0.83, indicating the consistent response strength across different baselines and objectives. Further comparison (Figure 5) between different baselines shows the strong ranking association (ρ=0.95)(\rho = 0.95).

Note on structural descriptor beyond single interpretation:

As we discussed in the overall comments, our methods offer a new vocabulary for general phenomenon characterization. In [1] The author states that “we might still be able to derive progress measures that don’t require human understanding”, and explicitly proposed “task-independent progress measures”. Our framework takes a step in this direction through a common attribution-based construction evaluated across tasks, modules, and behavioral settings. We added the comparison in the main experiment, showing the comparable response between V\mathbb V and Fourier Gini Measures, addressing the need for progress indicators with reduced reliance on detailed circuit knowledge.

[1] Progress measures for grokking via mechanistic interpretability, ICLR 2023.

Dear reviewer,

We would appreciate your feedback on whether our responses have addressed your concerns, and we would be happy to answer any further questions.

As a summary, the changes in the revised manuscript are marked in blue. (Sorry for the confusion; our previous manuscript update may not have been visible to reviewers, possibly due to an OpenReview issue. We have updated with the marked version) .

The summary of changes is listed as follows:

  1. Significance: We included additional discussion about the key problem solved, including two ICML position papers demonstrating the significance of related problems in Section 1. We added comparisons with task-specific Fourier-Gini progress measures across four Grokking tasks in Section 4.2, paragraph 1 and Table 2 to show the addressed problem of constructing a progress measure with reduced reliance on detailed knowledge of circuit mechanisms.

  2. Response to W1 & Q1 of Reviewer CjQy (mechanism): Added Fourier associations and intervention evaluation on Add P97 and S5 Product to investigate task-dependent local complexity reorganization. Last paragraph in Section 4.2 for brief discussion; Appendices E.3–E.4 for detailed analysis.

  3. Response to W2 & Q2 of Reviewer CjQy and Q1 of Xg85 (sensitivity): Added attribution-choice robustness analysis, covering baselines, objectives, paths, peak stability, and measurement ranking stability. Figure 5, Section 4.2. Detailed in Appendix G.2.

  4. Response to weakness raised by Xg85 (theoretical guarantee): Clarified our relation to historical research in complexity and the terminology of Local Complexity in Sections 2-3, marked in blue.

  5. Response to weakness raised by Xg85 (parameter control): We clarified that the clean reference defines the environmental variable. (Section 4.1). Controlling it makes the indication easier. The corresponding data are in Table G.7 on page 69.

Overall, rather than showing a uniform one-way correlation with Fourier circuit formation, VA\mathbb V_{\mathbb A} uncovers a task-dependent circuit divergence. It tracks symmetric operand merging as Fourier circuits form in commutative Add P97 while capturing role-preserving operand transport in non-commutative S5 tasks.

Reviewer_Xg85OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces attribution vergence, which measures the changes of the dimensionality of attributed contributions. The proposed method formulates attention, convolution and linear projections as aggregation nodes and the measurement is defined as the difference between Renyi-2 entropies of the pre and post aggregation distributions. The experiments on four algorithmic grokking tasks and adversarial transfer settings show the effectiveness of proposed method.

Critical strengths
  • The proposed measurement framework is well-initiated and designed with explicit math conditions.
  • The empirical evaluations are well-designed and includes meaningful safeguards.
Critical weaknesses
  • The assumptions connecting the attribution statistic to a meaningful structural interpretation are not properly validated. The proposed method depends on the attribution objective, reference and a path. In adversarial experiments, generator changes affect both the clean image distribution and the attribution path, and make observed patterns ambiguous. The integration checks validate the numerical precision, but not robustness to these choices. The theoretical guarantees do not establish that the measure captures local computational complexity.
Questions for authors
  • For fixed clean/adversarial endpoints, do the main layer rankings and behavioral associations persist when the generator-derived attribution path is replaced with a common straight-line path?
6 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for recognizing the motivation of our work, the explicit mathematical framework, and the careful experimental design.


On the overall soundness

Q: The assumptions connecting the attribution statistic to a meaningful structural interpretation are not properly validated.

A: Thanks to the reviewer for pointing out the importance of structural interpretation.

We would like to clarify that, as stated in the Overall Comments, the goal of our paper is to first establish a new observable itself, with supporting more structural interpretations as one of the applications. Such problems have been emphasized by a line of recent research, and our cross-task responsiveness directly validates it. Related to the meaningful structure, we clarified the theoretical background in our response to Local Complexity. We also provide a concrete structural interpretation assisted by V\mathbb V in our response to Reviewer CjQy.

Briefly, our framework utilizes the information-perspective complexity concepts to formulate a structural descriptor that shows indicative capabilities across the grokking and adversarial transfer (relating diverse edge-case phenomena in [1] and constructing general macroscopic observable in our abstract), and structure is not our direct goal, but our method.

The relation between structural descriptor and structural interpretation is more like physics and biology. As recent research focuses on the biology of LLMs [2], our results take a step toward the physics of the complex model.

[1] Position: Not All Explanations for Deep Learning Phenomena Are Equally Valuable, ICML 2025.

[2] On the Biology of a Large Language Model. Transformer Circuits Thread.


On the sensitivity of attribution used

Q: The proposed method depends on the attribution objective, reference, and a path. For fixed clean/adversarial endpoints, do the main layer rankings and behavioral associations persist when the generator-derived attribution path is replaced with a common straight-line path?

A: Yes. The main layer rankings and behavioral associations persist.

We agree with the reviewer that the indicator should be simple to make the behavioral association general, and the current evaluation of the attribution parameter sensitivity might not be sufficient. Methodologically, we make a formulation to abstract the diverse neural models and maintain the same attribution and invariance properties. We show the measure's change is bounded when the attribution is perturbed. Here we supplement experimental results on different paths (straight or generator trajectory) and different objectives. The rank correlation scatter plot is given in the revised manuscript, page 9, Figure 5, showing that rank correlation between VA\mathbb V_\mathbb A with attribution alternations is >0.95>0.95 in the Adversarial Experiment and >0.9>0.9 in the Grokking Experiment.

For the adversarial task, Appendix G.2.2 and Tables G.5 report the absolute responses and comparison. The four transfer-direction mean responses of adversarial transfer are as follows. We also clarify that, in our adversarial experiment, the clean reference is severed as an experimental condition, but not the mutable attribution setting.

Attribution ObjectiveGenerator path partial ∥ρ∥\|\rho\|Straight path partial ∥ρ∥\|\rho\|Generative path partial R2R^2Straight path partial R2R^2Δ partial ∥ρ∥\|\rho\|Δ partial R2R^2
Feature displacement (Original Objective)0.87320.86890.63790.6560-0.0042+0.0180
True-class similarity0.85920.85110.61500.6414−0.0081+0.0264
Fixed random projection0.86720.85860.60120.6262−0.0086+0.0250

For the peak response layer that may account for the transfer mechanism, we compare 24 alternative settings, comprise four transfer directions, three attribution objectives, and two response statistics. The results are as follows, showing the consistency of mechanism layer localization.

SourceSame peaks response layerRemaining peak shifts
ViT-B/3210/12One layer in each of two comparisons
ConvNeXt12/12None

More results on grokking are shown in our response to reviewer CjQy.

AuthorsOpenReview ↗

Response to Reviewer # 2

On the parameter control of the interpretation

Q: In adversarial experiments, generator changes affect both the clean image distribution and the attribution path, and make observed patterns ambiguous.

A: We clarify that such ambiguities serve as an enhancement of the experimental setup, which increases the difficulty for measurements to accurately extract underlying patterns. As for our specific interpretation application, the original peak analyses resolve injection-position (tpt_p) responses within each generator while controlling budget and momentum (Appendix F.2, Figure F.1), and show stronger within-generator transfer associations at the identified peaks across all six directions (Appendix F.3, Figure F.2).

We have additionally included within-generator transfer analyses in Appendix G.2.2 (Table G.5). The following table gives four-direction mean partial ρ\rho responses under the original attribution objective. Controlling the generator improves ρ\rho for each setting, and the pooled one is harder since the measurement shall be invariant across generators.

Attribution pathPooledPMF onlyRAE onlySiT only
Generative0.87320.92030.94360.8810
Straight0.86800.92170.95250.8803

On the theoretical grounding of Local Complexity terminology

Q: The theoretical guarantees do not establish that the measure captures local computational complexity.

A: Thanks to the reviewer for raising the issue. We clarify that our framework characterizes Local complexity but not Local computational complexity, and add a brief introduction to historical works on complexity as the second paragraph of Section 3 in our revised manuscript:

Complexity has been studied from the information perspective and from the computational perspective. The former concerns the difficulty of transmitting data in a communication model, as characterized by Shannon entropy (Shannon, 1948), while the latter concerns the difficulty of solving a problem in a computational model (Arora & Barak, 2009). Information-perspective complexity has also been applied to prove the computational complexity problems (Chakrabarti et al., 2001), and has been extended to networks to indicate their structural properties (Dehmer & Mowshowitz, 2011; Li & Pan, 2016). Another line of research introduces computational models into measures of data complexity, including Kolmogorov complexity (Kolmogorov, 1968), and the recently developed Epiplexity (Finzi et al., 2026). We explore an alternative direction by introducing information measures to the algorithm instances in computational models, specifically neural computation graphs, to measure the complexity of attributed circuits and indicate their properties. This direction is less explored since the classical algorithms themselves are well formalized and have fewer interpretation problems. The computational model also serves as a formalization tool for axiomatic attribution and an analysis tool for computational tractability in our framework. (Revised)

The term complexity is introduced earlier in the definition of dimensionality in our revised manuscript. Note that Bhatt et al. adapt the entropy to measure the complexity of explanations from the perspective of using it, while ours formalizes a behavior indicator.

Entropy and sparsity statistics have also been used to characterize the complexity of attribution when evaluating explanation quality (Bhatt et al., 2020; Chalasani et al., 2020). (Revised)

In Section 3.2, we introduce the term local complexity.

Intuitively, larger values indicate greater routing-source heterogeneity, and since the coordinate range of S is restricted to a single network layer, we identify it as a characterization of the Local Complexity. For a component-wise non-negative S, the following theorem shows that VA(S)\mathbb{V}_\mathbb{A}(\mathbb S) exactly quantifies the reduction of Entropy, or the local complexity, induced by reducing the route coordinate Z. (Revised)

We believe that a more global complexity characterization of the large models and agentic systems could be a future direction worth exploring. We have also revised the text in the experiment section to prevent the misconception that we are introducing a new concept of local complexity here.

Dear reviewer,

We would appreciate your feedback on whether our responses have addressed your concerns, and we would be happy to answer any further questions.

As a summary, the changes in the revised manuscript are marked in blue. (Sorry for the confusion; our previous manuscript update may not have been visible to reviewers, possibly due to an OpenReview issue. We have updated with the marked version) .

The summary of changes is listed as follows:

  1. Significance: We included additional discussion about the key problem solved, including two ICML position papers demonstrating the significance of related problems in Section 1. We added comparisons with task-specific Fourier-Gini progress measures across four Grokking tasks in Section 4.2, paragraph 1 and Table 2 to show the addressed problem of constructing a progress measure with reduced reliance on detailed knowledge of circuit mechanisms.

  2. Response to weakness raised by Xg85 (theoretical guarantee): Clarified our relation to historical research in complexity and the terminology of Local Complexity in Sections 2-3, marked in blue.

  3. Response to weakness raised by Xg85 (parameter control): We clarified that the clean reference defines the environmental variable. (Section 4.1). Controlling it makes the indication easier. The corresponding data are in Table G.7 on page 69.

  4. Response to Q1 of Xg85, W2 & Q2 of Reviewer CjQy (sensitivity): Added attribution-choice robustness analysis, covering baselines, objectives, paths, peak stability, and measurement ranking stability. Figure 5, Section 4.2. Detailed in Appendix G.2.

  5. Response to W1 & Q1 of Reviewer CjQy: Added Fourier associations and intervention evaluation on Add P97 and S5 Product to investigate task-dependent local complexity reorganization. Last paragraph in Section 4.2 for brief discussion; Appendices E.3–E.4 for detailed analysis.

In addition, we have revised the discussion on complexity in the last response and in the manuscript to improve accuracy. More related to computational complexity, our observation of empirical difficulty in task-dependent attribution reorganization indicated by V\mathbb V may relate to the grokking version of distribution-specific hardness of learning, which could be regarded as a future direction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall comments on the paper's contribution

We thank the AC for coordinating the review process and the reviewers for their time, thoughtful reviews, and constructive suggestions. Since the reviews report confidence levels of 2, 2, and 1 and were posted around or after the original scheduled end of the limited review period, we would like to share our perspective on the broader research context and help clarify how we see the paper’s contribution within it.

Reusable observable itself remain an important question in model science. Paper[1] (ICLR 2023) discussed progress measures that don’t require human understanding of detailed mechanisms and task-independent progress measures as the future directions. Two ICML 2025 position papers subsequently emphasized new scientific vocabulary [2] and broader explanatory theories [3]. Our contribution takes a step in this direction through a common observable of attribution organization. Supporting specific interpretations of grokking and adversarial transfer, although also important, are applications of this measurement framework. In our responses to the reviewers, we further clarify the connections to related work and hope that these explanations address the concerns raised.

We view common observables as a challenging, non-consensus research direction. Our introduction reviews the graph conductance and modularity as classical works illustrating the scientific role of structural descriptors. As AI-assisted research accelerates theoretical and experimental research, new abstracted concepts and measurements become increasingly important for scaling humans' understanding. Such direction matches the ICLR 2027 CFP Note for veteran submitters’s call for slow science, and through our response, we provide clarifications and new results, and hope the CFP Note's complete and exciting work is met.

References

[1] Progress measures for grokking via mechanistic interpretability, ICLR 2023, Appendix F.

[2] Position: We Can’t Understand AI Using our Existing Vocabulary, ICML 2025.

[3] Position: Not All Explanations for Deep Learning Phenomena Are Equally Valuable, ICML 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.