acceptodds
Under review as a conference paper at ICLR 2027

When Is an Attention Head a Computational Unit? Shared Circuits Despite Orthogonal Representations

Abstract

Mechanistic analyses of transformers often proceed head-by-head, looking for attention maps that correspond to computational roles. But when should an attention head be interpreted as a computational unit? In principle, a single computational function may be distributed across heads, and a single head may participate in several computations, with invariants appearing only after OV routing and residual-stream summation. In language models, where the underlying features and computations are unknown, an apparently uninterpretable head may be polysemantic, part of a distributed circuit, or a sign that the analyst chose the wrong decomposition. We therefore study transformers trained on factored sequence processes, a controlled setting where the target computation is analytically specified: multiple latent factors, represented in orthogonal residual-stream subspaces, receiving independent predictive updates. This lets us ask whether the attention circuits implementing these updates are themselves factorized. We find that they need not be. Per-factor updates constrain only the aggregate contribution across heads, leaving substantial freedom in how computation is distributed. Heads may specialize to individual factors, compose with other heads to implement one factor, or contribute polysemantically across factor boundaries. The regime that emerges depends on the generator's spectrum, the number of available heads, and training dynamics. We introduce effective subspace attention, a scalar quantity that combines attention patterns with OV circuits to measure the amount that heads collectively route from a source position into a given factor subspace. In our controlled setting, this aggregate view recovers the theoretically predicted computation even when per-head maps are illegible. Because only the aggregate is constrained, computation may be best understood at the level of task-relevant subspaces rather than individual heads. We then ask whether effective attention can support circuit discovery in real models and find that its natural application is closely related to existing circuit analysis techniques. When applied to GPT-2, it recovers known and novel circuits on two contrastive next-token tasks.

Primary area:
interpretability and explainable AI
Keywords:
Factored Representation, Attention, Attention Heads, Hidden Markov Model, Computational Mechanics, Toy Model

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

This paper studies whether independent computations are carried out by independent attention heads. It trains single-layer Transformers to predict sequences generated by two independent hidden Markov processes. Section 3 describes attention as accumulating token-dependent evidence with a factor-specific decay over distance: QK supplies the source weights, while OV determines what information is written into each factor's space. The authors show that this computation can be distributed across heads even when the factor representations are separate. They introduce effective subspace attention and derive minimum head counts under specific assumptions. Validation includes 240 training runs across five settings, head counts, and seeds, controlled interventions, and two GPT-2 case studies: IOI and neither/nor prediction.

Critical strengths

I appreciate the interpretation of attention as an approximate update of information about HMM hidden states. It connects distance-dependent weighting to QK and token-dependent writes to OV in this setting. This makes the proposed computation more concrete than an attention-map interpretation alone. The known data-generating process and 240 training runs provide a meaningful setting for testing this view.

Comparing positive and negative decay cases helps explain why a single head may not implement a factor on its own. Combining attention with OV writes reveals structure that raw maps can miss, and the interventions add useful evidence. I find the possible connection to the degree of attention superposition interesting. I regard this as a insteresting and useful direction.

I consider Effective Subspace Attention a valuable contribution of this paper. It not only provides theoretical predictions of the computations the model should perform, but also recovers these computations from trained attention heads, allowing the authors to verify whether multiple heads collaboratively implement the theoretically predicted computations.

Critical weaknesses

Is the paper explaining how a known computation is distributed across heads, or proposing a method for discovering unknown computations? The controlled experiments support the first goal more directly. Eq.(5), Δr_n(d)=∑s=1^d ζ_n^(d−s)gn(z_s^(n)) , has the same weighted-sum form as attention. This does not make the theory trivial: the decay rates and update directions come from the generator. However, the paper should distinguish deriving this target computation, showing that attention can express it, and demonstrating that trained models actually learn it. A concrete sequence example would make the contribution much easier to assess. In Appendix L and Table 3, unequal positive and negative decay rates are grouped into a two-head case. Under the stated requirement of exact realization at all distances, +0.3/−0.5 needs three heads, unlike the tested +0.7/−0.5 case. Finite sequence length must also be distinguished from arbitrary distances. The general counting procedure raises another issue: the number of boundary directions of the target updates need not equal the minimum number of head directions, since other directions can also combine to produce the targets. These concerns affect the theoretical extension, not necessarily the five tested configurations.

The real-model application is limited and does not test the full theory. Section 5 studies GPT-2 small on IOI and neither/nor prediction. These examples are useful, but the method uses an answer-logit direction rather than recovering unknown factor subspaces and their update rules. It is therefore unclear which parts of the HMM-based account carry over, or what the method adds beyond existing OV/logit contribution analysis. The authors acknowledge that this is a demonstration; the main claims should reflect that scope. In the neither/nor case, the evidence also does not clearly separate pairing information in the heads' OV weights from information already present in their inputs. I am not requesting a large new benchmark, but a clearer account of what these examples establish.

Theoretical and experimental differences from closely related work are not established clearly enough. Anthropic's Attention Head Superposition (Jermyn et al., 2023), which the paper already cites, demonstrated computations distributed across heads in toy models. He et al. (2026), Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition (Lorsa), decomposes the attention layer as a whole into sparse, interpretable computational units rather than treating native heads as the units. Wang et al., Dimensional Collapse in Transformer Attention Outputs: A Challenge for Sparse Dictionary Learning, further studies attention outputs concentrated in low-dimensional residual-stream subspaces. The latter two works are not discussed. These results are related but not equivalent: a low-dimensional output space need not be a task-specific computational unit. The authors should explain which theoretical claims go beyond these results and provide a focused comparison on the same controlled model. In particular, how do effective subspace attention and Lorsa differ in recovering factor updates and separating computations, and how do the factor spaces relate to the low-dimensional attention-output space? Such a comparison should report reconstruction quality, full update errors, and held-out results. I am requesting evidence of the differences, not assuming that the prior methods solve the same problem or requiring a full large-model benchmark.

Questions for authors

The questions below concern the current claims, not optional future projects. Their answers could change my assessment.

Is the main contribution an account of how known factor updates are distributed across heads, or a method for discovering unknown computations? Please use a concrete sequence to explain the prediction task, the information supplied by each token, and the combined head updates. Which factor labels, theoretical directions, or target answers must the analyst know beforehand? What does the theory predict beyond the weighted-sum structure of attention? And if it is about interpretability, could you briefly describe what exactly is being interpreted?

Does the two-head claim require the positive component to decay more slowly than the negative one? How does it change for +0.3/−0.5 and for finite versus unlimited sequence length? Why should the general procedure count the target updates' own boundary directions rather than allow a smaller set of other head directions that can also produce those updates?

I wonder whether there are practical methods to quantify the functional similarity and specialization of attention heads. For example, how can we measure how many distinct functions are represented by a single head, and how many heads collectively contribute to a particular function?

3 comments
AuthorsOpenReview ↗

Reviewer ZHUU Response

We would like to thank the reviewer for their thoughtful comments and their close reading of our manuscript. We acknowledge that the clarity of the current manuscript be improved and propose the following outline:

  • Describe the minimum architecture and conic attention head counting theory
  • Apply the general theory to the cases of distinct positive, and positive and negative eigenvalues. These cases are conceptually simpler and motivate introducing effective subspace attention. This also allows the reader time to see some results in between introducing more notation and theory.
  • Demonstrate that ESA recovers the theory across all configurations in the toy model.

Q1: We consider the main contributions to be 1) the precise link between the spectral properties of the data generator and the minimal function the Transformer can implement to get optimal loss and 2) our use of this precise setting to show that the aggregate contribution of attention heads within specific subspaces can be viewed as interpretable even when individual attention head contributions cannot.

While the existence of uninterpretable heads is well established in the field, their origin is poorly understood. Are illegible heads a consequence of poor training, highly entangled representations, or capacity constraints? We demonstrate that none of the above are necessary for head collaboration. In a toy model of data that demonstrably leads to disentangled representations, head collaboration is still present, even when trained to Bayes optimal loss with sufficient capacity. In LLMs, attention head superposition may arise from dimensionality pressure and polysemic/collaborative heads may form because of training dynamics and just having a large amount of heads. Our minimal head count case is neither over nor under-parameterized.

We will make the positioning of our paper next to the related works highlighted by the reviewer clearer. Specifically we will address how this relates to superposition (namely that it is not superposition because there are no capacity constraints), and that instead of looking for discrete computational units (as is the case with the decomposition and dictionary learning methods) we predict the required update. We can also include example sequences from our process in the appendix, to clarify the structure of our dataset.

In terms of what information must be known beforehand: we do need to know which object (the predictive vector) the Transformer tracks to make its prediction. In terms of how we can disentangle collaborative/polysemantic attention heads in real models, the answer is that we must have some candidate subspace/direction already.

As to what is being predicted by the theory, it is the minimum number of attention heads required for optimal loss as well as which update they must perform. The interpretability claim relates to working out which algorithm the model implements. We want to know: “if trained on data X, what is model M going to learn?” We provide a case where this is fully tractable to compute.

Q2: We commend the reviewer on this question! The +0.3/-0.5 case indeed does have a 3-head requirement, with the theory as written. +0.3/-0.5 requires three heads for all distances while two are sufficient for any finite length context length. We will correct the corollary and appendix to distinguish these cases. As for “Why [..] count the target updates' own boundary directions rather than allow a smaller set of other head directions”, this is a relevant point in general for arbitrary numbers of factors (it vanishes with 2 factors only). The reviewer is correct that for an arbitrary number of factors, the latter can be smaller than the former. We will clarify this in the appendix.

Q3: For a specific function, figuring out the number of heads that contribute to it is relatively straightforward if we can use knowledge of the function to find a specific subspace/direction. We can then work out the effective attention similarly to our GPT-2 experiments where we measure OV-writes onto that target direction.

Knowing which functions a given head it implements, is harder. It's possible that an unsupervised method can be used to find consistent subspaces that various heads write to without knowing the function. Weight-level decompositions (such as SVD) could suggest candidate directions, and methods like Adversarial Parameter Decomposition can – given enough data – potentially discover many mechanisms a head participates in. However, none of those methods give any guarantee that we’ve found everything a head does. We consider this area of research – unsupervised subspace/factor detection – extremely important and are actively working on it. However, that project is an entire research question on its own– and not within the scope of the current manuscript.

Once again, thank you very much for your thorough review, especially catching the +0.3/-0.5 case!

Thank you for your detailed response. I generally like this work and will maintain my score. I also encourage you to compare your method with Lorsa in future work.

We realize we could have gone into more detail regarding the comparison with Lorsa, so we do that here.

The main difference is that our work explains which computation attention must implement and how it can be distributed across heads, whereas Lorsa is a concrete decomposition routine that finds the interpretable computational components of attention. You are right to point out that Lorsa also finds that the computational units need not correspond directly to heads -- we do not claim novelty here.

Our novelty lies in studying a setting where the computational parts are pre-specified by the data-generator. This lets us derive the required per-factor updates, the conic characterization of the update rules, and the number of heads required to minimize loss. From there, we can see that even with completely disentangled activations we still find that attention heads can be individually uninterpretable, and that at effective subspace attention is necessary to fully see the implemented computation. Lorsa in contrast lets us discover computation that attention implements, but not why it implemented it. We will add Lorsa (and the Wang et al. paper) to our related works section:

Our work is closest to approaches that explicitly recover computation spanning multiple heads. Parameter decompositions can identify functionally meaningful subcomponents of attention (Christensen & Riggs, 2025), while Bushnaq et al. (2026) find weight-space components spanning heads and separating QK from OV contributions. Wang et al. (2025) show that attention outputs concentrate in low-dimensional subspaces and leverage this structure to improve sparse dictionary learning. He et al. (2026) introduce Lorsa, a sparse decomposition of the full attention module that identifies interpretable units spanning multiple heads. These approaches agree with our work that multiple heads must be taken into account to discover units of computation, but differ in trying to discover computation from trained models. Unlike these approaches, our controlled setting lets us derive the required factor-specific updates analytically and characterize how they can be distributed across heads, even when factor representations are orthogonal.

We thank the reviewer for their continued engagement, and are happy to answer any other questions and provide more clarity.

We also agree with the reviewer that method-specific comparisons could be valuable follow-up work. Specifically, for settings like ours where the implemented computation is analytically derivable from the data-generator, do decomposition methods recover the derived factor-specific computations?

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

This paper investigates attention heads as computational units: whether or not they can be studied as distinct units or as a single component of a more distributed computational structure. To cleanly test this, the authors train transformer models from scratch on “factored sequence processes”, which are sequences generated by several cleanly separated “factors”. Therefore, they can test whether the attention circuits factor in the same fashion as the data representations in the model. The authors conclude that these attention heads do not factor in this way, and that individual heads can be illegible. They also introduce “effective subspace attention”, a construction intended to measure the combined contribution of multiple attention heads.

Critical strengths

The experiments are generally well designed and supported by controlled and falsifiable analyses with known factor dynamics. This is a useful contribution, as many similar attention head studies rely on qualitative attention maps or ad-hoc metrics. The plots are informative and easy to read. “Effective subspace attention” as a concept is interesting; it would be genuinely useful to decompose the collective contribution of multiple attention heads.

Critical weaknesses

While the main experiment is quite clean and somewhat informative, overall this paper does not provide particularly novel insights or actionable results. In the mechanistic interpretability community, as noted by the authors, it is well known that attention heads do not segregate cleanly into single factors. As well, the general applicability and utility of their conclusions is not adequately explored or tested beyond toy settings.

W1. The writing style is generally quite terse and hard to follow at times. I appreciate the authors attempting to rigorously define terms and concretize certain concepts, however the paper would benefit from more prose and less notation. As it stands, the rather limited contribution of the work (that attention heads cannot be cleanly decomposed into separable computational units) does not justify its level of complexity.

W2. Effective subspace attention seems to assume that we know a priori the directions of the factors, which in natural setting is certainly not the case, and could indeed be the most challenging part of assigning meaning to different components of the model. Mapping the heads to these known directions, while descriptively interesting, is not clearly actionable in realistic scenarios. As well, multiplying attention by OV projected onto the logits is very similar to logit attribution, as noted by the authors, although their method is not compared to these prior techniques. Therefore, it is hard to be sure of the novel contribution put forward by the paper.

W3. The paper would stand to benefit from more realistic tasks/experiments The synthetic experiments are clean and well controlled, effectively showcasing the phenomenon that the authors are trying to capture. But, to better understand the generality and utility of their conclusions, the authors should test the implications of their findings in more realistic settings as well as toy settings. The existing natural language experiments are simplistic and given too little space.

Questions for authors

Q1. What do the authors believe the strongest single contribution of their paper to be?

Q2. How useful is effective subspace attention in more realistic scenarios, where the directions of the factors (or even what the factors are) may not be known beforehand?

2 comments
AuthorsOpenReview ↗

Reviewer LBPd Response

We thank the reviewer for their thoughtful remarks and are pleased they see the practical value of the controlled setting used in our experiments as well as the conceptual value of Effective Subspace Attention. We understand their critiques regarding the writing style and note their concern that the complexity of the presentation makes the paper hard to parse, especially given their opinion that the payoff is relatively minor. On the one hand, we think that the manuscript can be made more legible in the revision– this will be, essentially, the main focus of the revision. The lion's share of the terseness is from the original page limit. However, the paper is fundamentally dense: it relies on combining results from disparate fields as well as architectural constraints of a transformer. We don’t expect that it will ever be simple to get through, even with the additional space. In light of this, we also hope to convince the reviewer that our contribution is more significant than their original review suggests.

Q1: Strongest single contribution (TLDR): when distinct computations happen in separable subspaces of the residual stream, we can use these subspaces and Effective Subspace Attention to back out the ways in which attention heads combine to implement these computations. The point, then, is not that attention heads cannot be cleanly decomposed into separable computational units. Rather, the point is that heads that were uninterpretable on their own can become interpretable in aggregate when viewed through their collective contributions to a particular subspace.

To expound, we would like to gently push back on the idea that our main contribution is to show that attention heads cannot be cleanly decomposed into separable computational units. As noted in the review and the paper, the existence of uninterpretable raw head outputs is an empirical fact known to the interpretability community. What is less understood is whether there is a way to look at the attention head outputs which reveals interpretable structure. We show that the aggregate contribution of attention heads within specific subspaces can be viewed as interpretable even when individual attention head contributions cannot. How? First, we leverage the fact that training a Transformer on factored processes carves the residual stream into factor-specific low-dimensional orthogonal subspaces that store and update factor-specific information. This provides a factor-wise signature and enables subspace identification. Our theory, which describes an algorithm for how the information within the subspaces is constructed, tells us what shape factor-specific updates should take at each context position, thereby giving us a definition of what “interpretable structure” means. With these subspaces in hand, we show that Effective Subspace Attention can be used to compute factor-specific effective attention weight profiles by combining fractional contributions across heads. By looking through this subspace specific lens, raw attention patterns that don’t match theory combine within a factor-specific subspace to reveal the expected theoretical update structure. Notably, our result does not depend on an arbitrarily imposed meaning of “collective computation”, which is often the case in superposition or polysemanticity studies of transformer heads. This is because our method uses a data model that predicts the exact structure of the collective contributions and the exact minimum architecture.

Q2: When we do not have candidate factors based on details of the data generator (as is the case for natural language), candidate factors can instead be discovered. Anthropic, Goodfire, and others continue to find geometric manifolds in LLMs that seem to have interpretable roles in downstream tasks. Examples include manifolds useful for line and character counting as well as ones that encode the days of the week, colors, even political bias. While these manifolds may not literally correspond to an ergodic HMM factor within a larger factored process, this is not required for Effective Subspace Attention to be used. As long as one has identified a subspace where it seems like some task-relevant computation is being carried out, then Effective Subspace Attention can be used to compute the per-head contributions to that subspace, thereby understanding how attention heads combine to produce the manifold of interest. We consider the discovery of the relevant subspaces at scale an incredibly important topic, but one that this paper does not address– and we see no reason to link our approach, which dictates what to do with a subspace once you have one, to a particular method of subspace discovery. In short: the reviewer is right that “effective subspace attention” is only one half of an interpretability tool that would be useful in realistic scenarios; however, we believe it is a half that pairs well with many different contemporary and ongoing results in interpretability.

Thank you for the thoughtful rebuttal. I have a much better grasp on the main finding and conclusion of this work now. In particular, the idea that "heads that were uninterpretable on their own can become interpretable in aggregate when viewed through their collective contributions to a particular subspace" is indeed quite interesting.

I will raise my score to a 3, but I strongly encourage the authors to re-write portions of the paper to allow for this takeaway to shine through. The submitted paper is difficult to follow, and the significance of the work is not clearly presented.

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

A good amount of research is done on studying the attention weight matrix. However, the paper argue that attention heads may not be the right unit for analysis. It studies a toy example one-layer transformers trained on two hidden Markov processes. The attention heads can specialize, collaborate at the same loss. Moreover, it shows that effective subspace attention could recovers the original compute unit.

Critical strengths

The paper uses a special toy dataset to probe the attention compute unit, which is really interesting and novel. I think the toy dataset make it much clear for the conclusion.

Critical weaknesses

None

Questions for authors

Could you have try all different kind of setup, like head dim, number of heads and so on? I would like to know how those factors affect the results. Maybe if you use a lot of attention heads, then almost all heads will be specialize or collbrate?

1 comment
AuthorsOpenReview ↗

Reviewer sZvQ Response

We are pleased that the reviewer found the paper interesting and learned something useful.

In general, for increasing either dimension or number of heads, we observe the following:

  • Our original data actually already had enough coverage for us to go back and look at the model circuits as a function of head dimension. In this investigation, we did not observe a strong trend towards or away from specialization or collaboration. As we increase head dimension, models with fewer heads perform better, as they gain more computational power. For some parameterizations of Mess3 (notably those with high alpha or low x) where attention matters less for reducing loss (i.e. much of the loss is reduced by just modelling the stationary distribution), this is enough to get within the noise-range of the Bayes loss. This is not true in general though, and fewer-than-the-required-amount of heads with higher dim is generally worse than the exact amount of heads but with smaller dim.

  • Increasing the number of heads is already in the manuscript draft: if you look at figure 7 in the appendix, you can see the collaboration/polysemy as a function of the number of heads. Given that no parameterization in our paper requires more than 2 heads, anything beyond this on the x-axis is the over-parameterized case. What we observe is dependent on the data-generator, since that determines which solutions can be learned. In general, the number of effective heads goes up. This concurs with the picture that the training dynamics paint: namely that the heads initially all work on the higher eigenvalue and then there is some redistribution after that. There are 2 cases, same magnitude +- and identical negative where this is not the case, but generally speaking, an extra head will do something for one of the factors.

Again: We thank the reviewer for their assessment. Please do let us know if there is anything else we can clarify.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Reviews Summary And Response

We thank the reviewers for their careful and constructive reviews. We are encouraged that they converge on three strengths:

  1. A controlled setting with a known data-generating process, which makes the conclusions clear and falsifiable instead of resting on qualitative attention maps
  2. A concrete account of the minimal division of labor dictated by the transformer architecture.
  3. Effective subspace attention (ESA) as a useful measure of what several heads compute together.

We note that two reviewers asked us to name the main contribution, a strong signal that this needs to be made more explicit in the revision. For clarity: It is not that individual heads can be uninterpretable, which is well known. It is that ESA recovers the predicted update however a trained model divides it, so heads that are illegible one at a time become legible in aggregate, against a target fixed in advance by theory. We show this using a dataset for which the spectrum of the data generator determines, before any training, the update the attention layer must write into each factor subspace and, within the solution class trained models adopt, the minimum number of heads needed to write it, while leaving open how heads divide that update. As an important consequence, our model reveals that completely disentangled activations do not imply interpretable attention heads. This refutes the hypothesis that messy attention maps are a consequence of messy representations, which is something we only could have learned from studying a toy model.

We would also like to point out that our observed “head collaboration” should not be confused with previous accounts of superposition under capacity pressure: we see collaboration/polysemanticity in the regime where the model has exactly the minimal amount of attention heads needed to solve the task (a regime we only have access to because of our choice of data generating process). The reviewers’ concerns can be classified into two main issues:

Scope: reviewers correctly point out that the manuscript does not present a method to identify subspaces in scaled models. This is true: ESA specifies what to do with a subspace once one is in hand, independent of how it was found. It only needs a candidate subspace or readout direction, not an HMM factor specifically. This means it pairs naturally with concurrent work that locates task-relevant subspaces and manifolds in large models. The manuscript already attempts to say this in the conclusion, but we will make sure this is more clear in the revision.

Presentation: As a whole, the reviewers found the manuscript to be difficult to parse. We appreciate that our results require bringing together several technical fields, which makes it a nontrivial endeavor as a reader. Our revision will be devoted almost entirely to legibility, using the extra space to be more explicit and pedagogical rather than adding additional results. We will lead with the minimal architecture and conic theory, then work through concrete cases (currently in the appendix due to length constraints), and only after this introduce ESA. We will also position the work explicitly against attention-head superposition and dimensional collapse in attention outputs.

As the feedback round closes, we would like to thank their reviewers for their continued engagement and their suggestions to our manuscript. We are also happy that all of the reviewers appreciated our work, and that they all now recommend publication.

We have uploaded a new revision of the manuscript that addresses most of the points discussed.

We have:

  • Rewritten the end of the introduction to better highlight our paper's central contribution as recommended by reviewer LBPd.
  • Specified that for the real-model GPT-2 task, the readout direction is supplied by the task, not discovered unsupervised as recommended by reviewer zHUU.
  • Expanded our related works section to address the other attention decomposition methods highlighted by as recommended by reviewer zHUU.
  • Provided a worked example in our conic decomposition appendix to help exemplify what the theorem predicts are commended by reviewer zHUU.

We will also amend the theorem to account for the finite-length vs. infinite length difference and the distinction between whether the positive or negative eigenvalue have the larger magnitude when the two eigenvalues signs are different. (once the experiments for the -0.3, 0.5 case are done and added to the main figure). What remains is a slight re-arranging of the manuscript to present the theorem and its consequences in a more intuitive way, which we also commit to once we have these results.

We would like to thank everyone for their participation in what we consider to be a constructive rebuttal round!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.