acceptodds
Under review as a conference paper at ICLR 2027

From Representation Correspondence To Computational Trajectories In Language Models

Abstract

Geometric analyses of neural networks characterize correspondences between hidden states, but it is unclear when a correspondence between representations extends to a correspondence between the computations that transform them. We study this question by distinguishing a state correspondence at one depth from the transport of a trajectory by a depth-indexed family of correspondences. Across equivalent number representations, affine low-rank correspondences accurately transport hidden states but fail when applied statically to local residual-stream updates. Transporting the two endpoints of an update with the correspondence fitted at each endpoint instead raises update alignment from to , approaching an update-specific correspondence at . The same matched-sample pattern replicates across musical-pitch notation systems (, versus for an update-specific map). In numbers, decomposing the dynamic predictor identifies evolution of the state correspondence across adjacent layers as the dominant additional component. Independent adjacent-map estimation reduces the effect but retains of the static-to-update-specific gap, while shuffled correspondences collapse to near zero. We then ask whether depth-evolving state correspondences predict computation outside the examples used to fit them. In controlled addition, correspondences learned only from baseline states of training operands predict arithmetic-specific trajectories for held-out operands. A strong late-layer correspondence emerges between Question and Compute formulations, and transported Question–Compute states induce aligned arithmetic-specific updates when executed by the actual destination transformer block. Finally, a semantic-shuffle control separates native computational correspondence from local dynamical compatibility: shuffling identity pairings destroys state alignment, native trajectory prediction, and interventional update recovery, while substantial layerwise square closure remains. Thus a representational snapshot does not generally specify a computational trajectory, but the evolution of representation correspondences across depth can contain predictive information about that trajectory. Computational correspondence should therefore be validated against native computation rather than inferred from state alignment or local dynamical compatibility alone.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, representation geometry, hidden representations, transformer dynamics, causal intervention

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZrEVOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates whether correspondences between hidden representations, plausibly of semantically equivalent inputs, also bring to similar representation trajectories across layers. The authors study affine transformations between hidden states at a given layer and between successive representations . They apply this analysis to musical notation, and arithmetic tasks, they explore the extent to which these correspondences can predict the effects of interventions.

Critical strengths

The paper may contain interesting results, and addresses an important problem, but its conceptual contribution is not communicated with sufficient clarity to allow a meaningful assessment of its significance. A substantial rewriting, starting from the abstract and introduction, is in my opinion necessary.

Critical weaknesses

I found the abstract and introduction exceptionally difficult to follow. The authors introduce the notion of a "correspondence between computations" without providing a clear qualitative explanation of what this means, or how it differs from the correspondence between hidden-state trajectories. Since the transformations of hidden states are fully specified by the architecture and trained weights, the meaning of a correspondence between computations needs to be explained much more carefully. The reader should be provided with a qualitative idea of the goal of the research before jumping into formal derivations.

Moreover, the introduction introduces mathematical objects such as "domains", hidden states, and affine correspondence maps without adequately defining them or illustrating their meaning with a concrete example. A simple example involving two alternative representations would make the problem considerably more accessible.

Questions for authors

I think that the papers unclear at a level that cannot be addressed by revising the paper. I suggest a careful and in depth revision aimed at explaining the problem in qualitative and intuitive terms. `

4 comments

Thank you for your frank assessment. We agree that the introduction moves to formalism too quickly, and we are rewriting it. Below is the qualitative explanation that will open the revised introduction.

The idea. Take two ways of writing the same concept, e.g. the number written as "7" and as "seven". At every layer, the hidden states of the two inputs can be related by a map, which we call a correspondence. Such maps are often found to be accurate, and accurate maps are commonly read as evidence that the model "treats the two inputs the same way". We ask whether that inference is justified. As the model processes the two inputs layer by layer, are the changes it makes to each input also related by the correspondence? If so, the model performs corresponding computations on the two inputs, not merely holding similar representations of them at some layer.

On "the transformations are fully specified by the weights": yes. Each transformer block computes the same function for both inputs. What differs is where in state space the two inputs lie, and the same block can act very differently in different regions. A correspondence between computations therefore means that the block commutes with the correspondence (Eq. 2): mapping a state and then running the block gives the same result as running the block and then mapping.

Why it matters. This gives two practical answers. (1) When can we trust that a model computes the same thing on equivalent inputs? Neither similar representations nor identical answers are sufficient. A correspondence that aligns states at 0.954 transports their updates at only 0.096. Three arithmetic phrasings are all answered correctly, but only two of them share a computational trajectory. (2) Where in the network does the information that carries a computation live? Not in any single layer's geometry, nor in one map shared across depth. A correspondence that transports the updates of one layer segment (0.60) fails one layer away (0.11 / 0.27), and one map for all middle layers reaches only 0.23.

Concretely, the revision will include:

  • a two-paragraph qualitative motivation built on these two questions;
  • the running example above with a schematic figure, before any equation;
  • explicit definitions of domain, concept identity and correspondence;
  • a notation table. The revised PDF will be uploaded during the discussion period, and we would be grateful if you could take a look.

Dear Reviewer ZrEV, thank you for your candid review. Following your suggestion, we rewrote the abstract and introduction from scratch, starting with a qualitative explanation and a concrete example before any formalism (pages 1-3, marked in red in the revised PDF). Even a quick look at these pages would help us greatly. We would appreciate your view on whether the paper is now clear enough to assess.

Thanks very much for your reply. Indeed I will be happy to read the paper in its revised version, but I think that such a revision cannot be performed in such a short time, by adding a few paragraphs. This, in my opinion, a clear example of a potentially interesting paper which requires a complete redesign to make it understandable to human readers, avoiding as much as possible the usage of AI.

We take the point about readability seriously, and we partly share it. The submitted introduction moved to formalism too quickly. AI-assisted writeup also tires the readers.

That is why the revision did not add paragraphs to the old introduction but replaced it. Pages 1-3 were rewritten by the author from a blank page, starting from the "7" vs. "seven" example and the question of whether a map that aligns two states also aligns what each layer does to them; AI tools were used afterwards only for grammar. Sec. 3.1 and the table in Appendix A define every symbol. Three further sections report the controls requested by the other reviewers. The revised PDF has been available since Nov 9 (AoE), and the final version since Nov 13. We understand that the discussion window is short. If you are able to read only pages 1-3, that would already tell us whether the rewrite succeeds. And if the thread is still lost, it would help us most to know at which step: the running example, the definition of a correspondence, or the step from state correspondence to update correspondence. That is the kind of feedback we could act on.

Reviewer_QGhtOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether similarities in neural representations also reflect similarities in how language models compute. The authors propose a framework that learns correspondences between representations at the same layer and predicts layer to layer changes. They show that static correspondences fail to align computational updates, while depth dependent correspondences substantially improve alignment. The authors further test this in an arithmetic setting with interventions, and use shuffled controls to show that local dynamical consistency alone does not imply correspondence with the native computation.

Critical strengths
  • The question of how geometry relates to causal computation is important.
  • The approach does not assume a specific structure and can be applied to the tasks where ground truth geometry is unknown.
Critical weaknesses

W1. The correspondence between the question and compute formulations could be caused by prompt similarity. The +0 baseline removes formulation-specific structure but not the contribution of the delta token itself. A control prompt containing the same input tokens without requiring addition would isolate the arithmetic computation from input-token representations.

W2: The paper lacks a validation against a known computational geometry. While the method does not need a known geometry, the authors could show that their method works on known addition geometry as sanity check. Zhou et al. (2024) and Kantamneni & Tegmark (2025) partly showed the causally relevant geometry. Without this kind of validation, we can not know if the proposed method works as expected.

W3: While the paper has causal patching experiments, it ends up with cosine similarity analysis which is a representation similarity measure. Neither model behavior is nor logits are measured for the arithmetic intervention. Therefore, causal analysis inherits all the weaknesses of cosine similarity.

Minor: Since the analysis relies on cosine similarity, a brief discussion of how anisotropy is controlled (e.g., centering and the shuffle baselines) would help readers familiar with this concern.

Informativeness: While the proposed method is interesting, I do not think the paper provides much new insight into how the model actually computes. The results mainly demonstrate geometric alignment rather than identifying the underlying computational mechanisms. I would consider increasing my informativeness score to 3 if the authors could show that the method recovers known computational structures or reveals new insights into the model's computation.

Questions for authors

Q1: The +0 prompts share the full arithmetic template with the evaluated prompts. Could the authors clarify in what sense the maps are learned "without arithmetic-specific trajectory information", and consider a baseline from non-arithmetic prompts containing the same operand (e.g., ‘The number is {n}.’)? This would help address W1.

Q2: Could the authors investigate whether their analysis on changes reflects the phase shifts in the known components of addition, (T2, T5, T10) as identified in the prior work (Zhou et al. (2024) and Kantamneni & Tegmark (2025))? (See W2. This would partly address W3.)

Q3: The authors show the extent to which geometry and computation overlap, but they do not show where they diverge. Could the authors show the situations where they do not overlap? A response to this question would help me reconsider the informativeness score, depending on what it reveals.

7 comments

Thank you for a careful and constructive review. Your questions identified exactly the controls the paper needed, and the results below substantially change how we will present the arithmetic section.

W1 / Q1: Prompt similarity and the +0 baseline. We agree that "without arithmetic-specific trajectory information" (line 366) overstates the claim. The +0 prompts share the arithmetic template; the accurate statement is that the maps are fitted only on Δ=0 states of training operands. We will rephrase it.

Building a token-matched non-arithmetic control was informative in itself. With the span "{n} plus {Δ}" followed by "Answer:", Llama-3.1-8B (base) outputs the sum even when asked to repeat or copy the text (100% of 120 problems), or to name the first number (90-95%). Addition stops only when the expression is quoted. We therefore use two quoted controls whose correct answer is the first operand n, which does not depend on Δ:

  • C1: In the expression "{n} plus {Δ}", what is the first number? Answer: / Write the first number of the expression "{n} plus {Δ}". Answer: (100% / 100% answer n)
  • C2, minimal edits of Question/Compute: What is the first number in "{n} plus {Δ}"? Answer: / Name the first number in "{n} plus {Δ}". Answer: (87% / 98% answer n)

Results (same estimator, operand split and held-out problems as Fig. 2a; layers 20-30; our re-collected arithmetic states match the paper caches with cosine > 0.999 and reproduce 0.56):

  • Question-Compute dynamic correspondence: 0.562 for arithmetic vs. 0.191 (C1) and 0.227 (C2).
  • Direct overlap of the Δ-induced updates, no map: 0.82 for arithmetic vs. 0.31 (C1) and 0.40 (C2).
  • Size of the Δ-induced update relative to the +0 update: 0.73 / 0.64 for arithmetic vs. 0.19 / 0.19 (C1) and 0.32 / 0.22 (C2).

The last line is the most direct answer to W1. Changing the Δ token alone, with the same surrounding tokens, produces an update 2-4 times smaller than when the model adds. Most of the Δ-induced update that Question and Compute share therefore arises from performing the addition, not from processing the shared tokens. Because the control updates are smaller, their lower correspondence may partly reflect a lower signal-to-noise ratio. We therefore read 0.19-0.23 as an upper bound on the token-driven part, not as a precise estimate.

Q1 directly: correspondences fitted only on the non-arithmetic prompts, then applied to the arithmetic Question-to-Compute trajectories, reach 0.341 (C1) and 0.354 (C2) when fitted on Δ=0 states. This evaluation is on arithmetic trajectories, so it is not affected by the small control updates. It is well above the static baseline (0.17) but below the +0-fitted maps (0.56). Part of the correspondence is shared with non-arithmetic prompts containing the same operands, and the rest is specific to the arithmetic template. We will report this split explicitly.

Operand order: an Instruction variant with the Question/Compute operand order ("Add {n} and {Δ}. Result:", 100% correct) gives 0.073, the same as the original Instruction pairs (0.087). Operand order does not explain the Instruction gap. This resolves part of the limitation in lines 437-440; the remaining factor is phrasing.

(continued)

W2 / Q2: Validation against known addition geometry. Thank you for the references, which we will cite and discuss. Kantamneni & Tegmark (2025) study the same model (Llama-3.1-8B) on the same operand range (0-99), which makes a direct comparison possible. We are fitting the helix basis [n, cos(2πn/T), sin(2πn/T)] for T = 2, 5, 10, 100 at each layer. We then test (i) whether the depth-evolving transport preserves the phase of each periodic component in the transported updates, and (ii) whether the late-layer onset of Question-Compute correspondence (layers 18-20) coincides with the emergence of the answer helix on the final token. We will post the results before the end of the discussion period.

W3: Cosine-only intervention read-out. Agreed. We are running the patched forward pass through to the logits and will report answer accuracy, answer log-probability and KL from the native output, against two controls:

  • a norm-matched random perturbation, which addresses the limitation we noted ourselves in lines 432-436;
  • a source-swap interchange: transporting the state of a different problem (n', Δ'). If the transported state carries arithmetic content, the destination prompt should output n'+Δ' well above the matched-norm random control. Results will follow shortly.

Q3: Where do geometry and computation diverge? The paper already contains several such cases, and the new controls add two more. We will collect them in a dedicated paragraph. (1) Accurate state alignment without computational alignment: held-out states align at 0.954, but the same map transports updates at 0.096 (Sec. 4.1). (2) Same answers, different computation: Instruction is answered correctly on the same problems as Question and Compute, yet Instruction-involving pairs reach only 0.09 dynamic alignment vs. 0.56. Behavioral equivalence does not imply a shared trajectory. (3) Local consistency without the native computation: under semantic shuffling, square closure stays at about 0.51 while trajectory and interventional alignment collapse to about 0 (Fig. 2c). (4) New, shared tokens without shared computation: the quoted controls above contain the same operand span as the arithmetic prompts, but their Δ-induced updates are 2-4 times smaller and correspond far less (0.19-0.23 vs. 0.56). (5) New, depth: a correspondence that transports the updates of one layer segment (0.60) fails one layer away (0.11 / 0.27), and a single correspondence for all middle layers reaches only 0.23 (reply to Reviewer TtpX). Together these show that geometric agreement, whether in states, in tokens or in a fixed map, does not by itself establish computational agreement. This is the main claim we will foreground in the revision.

Minor (anisotropy): all scores are centered rowwise cosines on held-out identities, and the shuffle controls preserve marginal geometry. We will add an explicit paragraph on anisotropy.

Results for W2/Q2 and W3, as promised. Same held-out operands and baseline-only maps as Fig. 2.

W2 / Q2: Known addition geometry (Kantamneni & Tegmark, 2025; Zhou et al., 2024). We fit linear (ridge) readouts of the helix features [cos(2πx/T), sin(2πx/T)], T = 2, 5, 10, 100, from final-token states, train on training operands, and evaluate on held-out operands. A linear readout cannot form the helix of n+Δ from separate helices of n and Δ, because this requires a product term. Linear decodability of the sum's helix therefore indicates that the sum has been computed.

(1) The answer helix forms exactly where the correspondence emerges. Units-digit phase accuracy of n+Δ (T=10; chance 0.10):

  • Question: ≤0.23 up to layer 16, then 0.88 at layer 18, 0.98 at layer 20, 1.00 from layer 24.
  • Compute: ≤0.28 up to layer 16, then 0.92 at layer 18, 0.97 at layer 20, 1.00 from layer 27. This is the same window (layers 18-20) in which Question-Compute dynamic correspondence emerges (Fig. 2a).

(2) Instruction, answered correctly on the same problems, never forms a linearly decodable sum helix at this position (max 0.42). The quoted non-arithmetic controls stay at 0.21-0.42. Their margin above chance is expected: for Δ = 10 and 20 the units digit of n+Δ equals that of n. Instruction reaches the same answers without the helix computation that Question and Compute share. This is consistent with its weak trajectory correspondence (0.09) and is a mechanistic instance of "same answers, different computation".

(3) Transport preserves the phase that prior work identifies. Using the destination's own readouts, layers 20-30:

  • State level, transported states: 0.99 / 0.90 / 0.70 phase accuracy for T = 2 / 5 / 10, vs. 0.50 / 0.20 / 0.08 with shuffled maps (native: 1.00 / 1.00 / 0.99).
  • Update level: we add the transported update to the destination's layer-l state and decode at l+1 (T=10). Where the helix forms, the transported update supplies the missing phase:
    • layer 18: 0.64, vs. 0.37 without the update and 0.46 with the static-map update;
    • layer 20: 0.88, vs. 0.58 and 0.68;
    • mean over layers 20-30: 0.98, vs. 0.84 and 0.86 (native 1.00). The depth-evolving correspondence therefore transports the periodic components that prior work implicates in addition, and the layer at which it begins to hold is the layer at which the answer's helix is formed. We will add this analysis as a figure in the arithmetic section.

W3: Behavioral read-out with norm-matched controls. We ran the patched forward pass to the logits (80 held-out problems, all six directed pairs). In no condition does replacing the final prompt-token state at a single layer change the argmax answer. This holds at layers 14-20 and 23-30, and even when the state is replaced with the destination's own native state for a different problem (≤0.2% flips). In this model, the answer is not controlled by that single position at any one layer. We therefore report graded effects.

  • Interchange (Question-Compute, layers 23-30): transporting the state of a different problem (n', Δ') raises the log-probability of n'+Δ' by 1.99 nats relative to a random perturbation of identical norm. This is 66% of the shift produced by native interchange (3.01 nats). KL from the native output: 0.026 vs. 0.002 for the norm-matched control and 0.040 for native interchange.
  • At layers 14-20, the effects are smaller for all conditions (0.40 vs. 1.52 nats; 26%).
  • Transporting the same problem leaves the answer intact (log-prob of the correct answer -0.112 vs. -0.103 native). The norm-matched control shows that the transported state is not merely a small perturbation; it carries the source problem's content. We will state the single-position limitation explicitly and replace the limitation in lines 432-436 with these results.
Reviewer_TtpXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper attempts to quantify the difference between representations at a given layer and the transport of a representation across layers. It does this by identifying maps that push a given representation through a layer. They consider this in two settings, and show that that transported representations are closer to the actual model direction.

Critical strengths

The authors consider an important question which has not previously been sufficiently well addressed in the literature.

Critical weaknesses
  1. It is difficult to understand this paper. Symbols are frequently not defined or explained. Most importantly, the paper hinges on C_static, C_dynamic etc… but these are never explicitly defined. In Eq 2, What is T and what is F? In Eq 3, what is X? What is a concept identity? What is a surface-form domain?
  2. It is not clear why the authors choose the evaluations they do. Can the authors justify this?
  3. It is unclear if the improvement in the transported trajectory is attributable to the fact that the transport contains fittable parameters from the target layer or from the need to incorporate information from the model’s computation.
Questions for authors
  1. Why were the two particular experimental settings considered chosen? The authors provide a brief discussion of previous work on arithmetic but no discussion of musical pitch.
  2. The “dynamic map” contains information about the layer at which the authors evaluate, whereas the “static” map does not. How did the authors control for the fact that the dynamic map has correspondences fitted from the target layer l+1, whereas the static map only has information from layer l?
2 comments

Thank you for the detailed review. We agree that the presentation needs substantial improvement. Your Q2 led to the most informative new result of this rebuttal.

Notation.

  • T: defined in the introduction (line 052) as the layerwise affine correspondence T_l(x) = L_l x + c_l.
  • F: you are right that F_A^l and F_B^l are never defined. F_A^l is the map computed by transformer block l, taking h_A^l to h_A^{l+1} for inputs from domain A. Eq. 2 states that transporting a state and running the block commute. We will define F before Eq. 2 and move Eq. 2 from Related Work into Sec. 3.
  • X: defined in Eq. 3 as X_A^l(c) = h_A^l(c), the matrix of states over concept identities. We will remove the redundant renaming.
  • Concept identity / surface-form domain: a concept identity is the underlying value (the number 7; a pitch). A surface-form domain is a notation that expresses it (digits "7", words "seven", ordinals "seventh"; different pitch notations). These are currently defined only by example in Sec. 3.4. We will define them explicitly in the introduction.
  • C_static / C_dynamic / C_update: these are the centered rowwise cosines on held-out identities (lines 188-189) between the native destination update and the predictors of Eqs. 6, 7 and the update-specific map. We will give the explicit formula.

Q1: Choice of settings.

  • Numbers: a controlled domain with known identities and three surface forms, which permits the exact decomposition of Sec. 4.2.
  • Pitch: an independent domain with both cyclic (pitch class) and ordinal (octave) structure. It tests that the result is not specific to number representations.
  • Arithmetic: adds a computation with a verifiable output, which permits interventions on the native computation. We will add this rationale and discuss prior work on pitch representations.

Q2 / W3: Is the gain due to layer-(l+1) information? We ran the control directly (number domain, layers 7-24, held-out identities). Each static predictor applies one frozen linear map to the update:

  • V P_l (original static): 0.096
  • V P_{l+1} (destination-layer map): 0.268
  • V (P_l + P_{l+1})/2: 0.256
  • V P_pool, one map fitted jointly on the states of l and l+1: 0.603
  • depth-evolving predictor: 0.683; update-specific upper reference: 0.692

Access to layer-(l+1) information alone is not enough. Maps estimated at single layers stay low even when the destination-layer map is used. A map fitted jointly at both endpoints of the update does recover most of the gain. We then asked whether such a correspondence is a single static object, by refitting it away from the segment it is applied to:

  • Shifted by one layer: 0.111 (fit on l-1, l) and 0.268 (fit on l+1, l+2), decaying further with larger shifts (0.076 and 0.164 at shift -8 / +8).
  • Widened depth window: 0.603 / 0.539 / 0.505 / 0.430 / 0.373 / 0.299 for windows of 2 / 3 / 4 / 6 / 8 / 12 layers.
  • One map for all middle layers: 0.226.

The answer to your question is therefore: the extra layer-(l+1) information matters only when used to fit a correspondence for that specific segment. Any fixed correspondence, including one fitted with that information, fails as soon as it is applied away from the segment it was fitted on. What transports the computation is a correspondence re-estimated per layer segment. This is the depth dependence the paper argues for, stated more precisely than in the submission. We will add these results as a figure in Sec. 4.2 and revise the claim to exactly this form.

On fittable parameters: all predictors above come from the same estimator, the same training identities and the same rank. The shuffled-correspondence control (0.003) uses the same layers and the same number of parameters as the depth-evolving predictor. The independent cross-fit (47.1% of the static-to-update-specific gap) rules out shared estimation error as the sole explanation.

Dear Reviewer TtpX, thank you again for your review. As the discussion period ends on Nov 18, we wanted to briefly point to where your questions are addressed in the revised PDF. Q2 (layer-(l+1) information): Sec. 4.2 and Fig. 2 show that static maps using layer-(l+1) information stay low (0.10-0.27), and that the correspondence that transports updates is depth-local. Notation: every symbol you listed (T, F, X, C_static, C_dynamic, domain, concept identity) is defined in Sec. 3.1 and in the table in Appendix A. Q1 (choice of settings): Sec. 3.4. We would be grateful to know whether these address your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision summary and open question

Dear Area Chair,

A short summary as the discussion period closes.

Reviewer QGht (confidence 4) raised their rating from 2 to 3 after the revision. Their review asked for two things: recovery of known computational structure, and cases where geometry and computation diverge. The revision gives both. Trajectory correspondence appears at the layers where the sum’s helix becomes decodable, and the transport carries its phase (Sec. 4.4, Fig. 4). Sec. 5 lists five cases where geometry and computation come apart, including a phrasing that answers every problem correctly without forming that representation.

Reviewer TtpX asked whether the depth-evolving predictor wins because it sees layer l+1. This question produced the best new result in the revision. A direct control (Sec. 4.2, Fig. 2) shows that static maps given layer l+1 stay at 0.10 to 0.27, that a map fitted on one layer segment transports its updates at 0.60 but fails one layer away (0.11 / 0.27), and that a single map over all middle layers reaches 0.23. The claim is now sharper: what transports the computation is a map re-estimated per layer segment. Every symbol they listed is defined in Sec. 3.1 and Appendix A. We would welcome Reviewer TtpX’s view on whether this control answers their question.

Reviewer ZrEV’s concern was presentation. Pages 1 to 3 were rewritten from scratch and have been available since Nov 9 (AoE). We would welcome their reading of the revised pages.

We are grateful for the reviews. The controls they asked for made the paper better.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.