acceptodds
Under review as a conference paper at ICLR 2027

An Exact Advection-Diffusion-Reaction Representation of Transformers

Abstract

Repeated Transformer architecture invites a dynamical-systems view, but leaves open how computation is organized and changes through depth. Building on recent work connecting self-attention to a Laplacian-like difference operator, we systematically analyze how attention, normalization, residual connections, and nonlinear feed-forward transformations compose. We obtain an exact advection–diffusion–reaction (ADR) representation of sequential pre-norm Transformer blocks, retaining the feed-forward response at the attention-updated state. This representation provides a common geometric framework for quantifying component actions, deriving constraints on causal modifications, and examining task behavior through controlled interventions. In the evaluated pretrained decoders, advection generally grows relative to the complete update while reaction's relative magnitude declines, despite absolute reaction growth in most checkpoints. On common inputs, neighboring complete updates are more aligned than a fixed distant-block control, while their transport and reaction actions remain directionally distinct. Separating the identity baseline reveals that apparent diffusion persistence can coexist with changing symmetric transport. An attention-approximation intervention further shows that an opposing feed-forward response can amplify the block-output error. By preserving the complete nonlinear computation, exact ADR provides a quantitative basis for relating component geometry to depthwise behavior and the effects of modifying a block. This connection supplies constraints for architecture design and evaluation criteria for inference approximation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Advection, Diffusion, Reaction, Transformer Blocks, Interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ePEzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper rewrites the Transformer block into a combination of Advection and Diffusion term, calling it ADR representation.
They measure the magnitudes of each component in well known models, and reports the tendency of these magnitudes with respect to depth.

Critical strengths

The paper attempts to tackle the recurring idea of Symmetry-Assymetry decomposition of the Transofmer operator and endeavors to make use of it in model analysis.

Critical weaknesses

The research itself is poorly motivated and it is hard to follow what it is trying to communicate to the community. It was a difficult read not because of the content itself, but because of its motivation.
If I am not reading wrongly, the paper's central mathematical theorem is essentially the composition of two elementary algebraic substitutions, as stated itself in its own proof --- and I am finding it difficult to identify the notable novelty of these theorems in the literature of mathemtical theory of transformers. On the other hand, the main contribution is primarily an empirical application of this decomposition, which seems not well motivated in practical sense as well. I believe they are inspired from layer reuse and local alignments, but the paper does not present an algorithm that utilizes ADR or so-derived measure of geometric quantities.
The paper seems to merely choosing to call the symmetric and skew parts of attention as diffusion and advection, followed by the analysis of their magnitudes.

Questions for authors

What is the practical use of this study's measured quantities of the magnitudes of each component? Please provide explicit example of how they can be used for compression, architectural modification, decision in model selection, etc and support its practical utility with experiment, even possibly providing a useful hindsight whose merit can be evaluated quantitatively.

Did you consider a connection to Helmholtz–Hodge decomposition in the gradient flow formulation of Geshkovski et al, which is mentioned in the introduction of this paper?

6 comments

Part 1/3

Weakness: Research Motivation, Theoretical Novelty and Geometric Interpretation

The research itself is poorly motivated and it is hard to follow what it is trying to communicate to the community. It was a difficult read not because of the content itself, but because of its motivation.

If I am not reading wrongly, the paper's central mathematical theorem is essentially the composition of two elementary algebraic substitutions, as stated itself in its own proof --- and I am finding it difficult to identify the notable novelty of these theorems in the literature of mathemtical theory of transformers. On the other hand, the main contribution is primarily an empirical application of this decomposition, which seems not well motivated in practical sense as well. I believe they are inspired from layer reuse and local alignments, but the paper does not present an algorithm that utilizes ADR or so-derived measure of geometric quantities.

The paper seems to merely choosing to call the symmetric and skew parts of attention as diffusion and advection, followed by the analysis of their magnitudes.

Response to Weaknesses:

Our motivation is to understand how attention, normalization, residual connections, and nonlinear FFNs jointly organize computation within a complete Transformer block, and how this organization changes through depth. Extensive research studies these components individually; their sequential interaction remains challenging to characterize. Building on the connection-walk representation and block-level ADR perspective of Lin et al. (2026) [1], Theorem 2 (§3.2, Eq. (11)) gives an exact representation of the native sequential pre-norm block, preserving attention’s normalization, head-sum scale, affine terms, and the full nonlinear FFN evaluated at the attention-updated state.

The revision makes the mathematical content beyond the symmetric/skew-adjoint split more explicit. Proposition 1 (§3.1, Eqs. (7)–(8)) separates the realized connection as

\qquad \langle x,\mathcal L_ \pi x\rangle_ \pi =\frac12\sum_ {i,j}\pi_ iP_ {ij}|x_i-O_{ij}x_ j|^2\geq0.$$ Here, diffusion is defined by transported disagreement, while the token-diagonal potential accounts for gain/loss. Dissipativity holds for every positive token reference, without assuming stationary routing or orthogonal transports. The split is unique once the routing–transport pair, reference, and discrepancy form are fixed. Theorem 2 then embeds these coordinates into the complete native computation. **The decomposition is the starting point for the analysis.** It enables us to establish causal modification constraints (§4.1), identify which action vectors survive a function-preserving normalization rescaling (§4.2), compare component geometry and complete updates through depth (§§5.1–5.2), and quantify how the FFN changes attention-approximation errors (§§4.3 and 5.3). These are the substantive questions addressed by the paper. Convergence remains a further question; Appendix B specifies the additional assumptions needed for a continuous-depth limit. The operator viewpoint also supports analyses beyond the ADR split. Our transport-spectrum studies (Appendices D.3–D.4) motivate the executable approximation experiment in Appendix D.2, discussed below. **References:** > [1] From Self-Attention to Connection Laplacian: A Unified Operator View of Transformers

Part 2/3

Question 1: Practical Uses of Component Measurements.

What is the practical use of this study's measured quantities of the magnitudes of each component? Please provide explicit example of how they can be used for compression, architectural modification, decision in model selection, etc and support its practical utility with experiment, even possibly providing a useful hindsight whose merit can be evaluated quantitatively.

Response to Question 1:

The quantities distinguish mechanisms within the complete update and provide a basis for controlled modification and evaluation. Equation (17) separates skew action, discrepancy dissipation, and signed potential gain; Eq. (18) distinguishes action magnitude from its directional contribution to the update. Norms alone do not establish causal importance. Indeed, §4.2 shows that some component magnitudes can change under a normalization rescaling that leaves the block function unchanged.

The revised paper provides three concrete uses and their supporting evidence.

1. Constraining architectural modifications. Proposition 3 (§4.1, Eq. (14)) shows that, for a fixed causal connection with nonzero cross-token transport,

\quad\text{has causal support on arbitrary fields} \quad\Longleftrightarrow\quad \alpha=\delta.$$ For the input-dependent native map, the potential’s coefficients must also be considered. Coupling all three gains gives $$I+\gamma(\mathcal A_\pi-\mathcal L_\pi+\mathcal V_\pi) =(1-\gamma)I+\gamma T,$$ which preserves native input causality for a fixed layerwise gain. This supplies an explicit family of admissible modifications and explains why independently changing the geometric components can violate causality. Learning these gains remains a proposed architectural experiment (§4.1; Appendix E). **2. Evaluating approximation after the FFN responds.** An attention perturbation changes both the attention output and the FFN input. Equation (19) therefore evaluates the complete error through $$\delta F=\delta t+\delta r,\qquad |\delta F|^2-|\delta t|^2 =2\langle\delta t,\delta r\rangle+|\delta r|^2,$$ where $\delta r$ is the native FFN response. Across six checkpoints and 128 held-out windows per checkpoint, the FFN response opposes the attention error **yet amplifies the block-output error in 36.3–95.8% of evaluated cases** (§5.3; Appendix C.6, Table 9). This gives a quantitatively testable lesson for approximation: opposition alone does not establish compensation, and local fidelity must be measured after sequential composition. Prediction loss remains a separate endpoint; adding this response measure to the tested loss predictor did not meet the prespecified improvement threshold. **3. Testing an executable compression candidate.** Motivated by the transport spectra, Appendix D.2 approximates one middle-layer attention output projection using a calibration-fitted low-rank basis while preserving its affine term. On GPT-2 Small/Large and Qwen3-1.7B/4B, the approximation meets the declared **0.02-nat/token upper loss-change bound on 64 held-out windows**, including at approximately **25% of the selected projection’s original matrix multiply–accumulate budget** (Figure 7). At that budget, projection working memory falls by 57–71%. The stronger advantage over matched magnitude/random controls is established in the two GPT-2 cases; measured whole-model latency shows no meaningful speedup (Figure 8). These experiments connect operator analysis to concrete modifications and measurable consequences. The compression example is motivated by transport spectra and fitted to calibration outputs; we do not infer compressibility directly from ADR component magnitudes. Appendix D.1 additionally uses geometrically constrained interventions to distinguish native strength, task sensitivity, and the FFN’s response in an exploratory task study.

Part 3/3

Question 2: Connection to Helmholtz–Hodge Decomposition.

Did you consider a connection to Helmholtz–Hodge decomposition in the gradient flow formulation of Geshkovski et al, which is mentioned in the introduction of this paper?

Response to Question 2:

The connection is relevant, with a distinction between the objects being decomposed. Helmholtz–Hodge decomposes a vector field into gradient, divergence-free, and harmonic components under appropriate geometry and boundary conditions [2]. Our construction decomposes the realized attention operator acting on token-feature fields, then retains its sequential composition with the nonlinear FFN.

For fixed routing, transports, and reference, Proposition 1 gives the quadratic energy

=\frac12\langle x,\mathcal L_ \pi x\rangle_ \pi, \qquad -\mathcal L_ \pi x=-\nabla_ \pi E_ {\mathcal L}(x), \qquad \langle x,\mathcal A_ \pi x\rangle_ \pi=0.$$ Thus the frozen diffusion component is a gradient flow of a nonnegative discrepancy energy, while the frozen skew component preserves the weighted norm in continuous time. This is a consequence of our operator identities; it does not establish a Helmholtz–Hodge decomposition of the full nonlinear Transformer dynamics or contraction of the native finite residual step (Appendices A.2 and B). Geshkovski et al. [3] distinguish normalized attention dynamics from an unnormalized surrogate admitting a Wasserstein gradient-flow formulation. Their formulation concerns evolution of token distributions, whereas ours describes a finite native block on token-feature fields. For input-dependent attention, differentiating a candidate energy or computing the divergence also differentiates the operator coefficients. The realized operator is therefore distinct from the full Jacobian, as stated in §2 after Eq. (3). Establishing a nonlinear Helmholtz–Hodge correspondence requires further analysis. **The exact finite-block representation in Theorem 2 does not require such a correspondence or a gradient-flow formulation of the Transformer.** **References:** > [2] Comments on the “Meshless Helmholtz-Hodge Decomposition” > > [3] A mathematical perspective on Transformers
Reviewer_EqW6OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose an exact Advection-Diffusion-Reaction (ADR) representation for Transformer blocks. It include attention, normalization, residual and nonlinear FFN. Authors also study the component behavior across layers. Experiments on seven pretrained models show different depth patterns and FFN error amplification.

Critical strengths

The exact ADR decomposition is mathematically clear. It keep original nonlinear FFN and native block computation.

Experiments cover seven pretrained models. The analysis show interesting difference between adjacent blocks and component directions.

Critical weaknesses
  1. Limited theoretical novelty and interpretation. The main ADR decomposition is interesting, but novelty seems not very strong. The skew-adjoint and self-adjoint decomposition is already standard mathematics. This paper mainly extend attention operator view to complete Transformer block. The FFN is directly put into reaction term, so exactness is partly from algebraic rewriting. I am not very clear what new mechanism this theorem really discover. Also, diffusion here not always have smoothing property. The component definition also depend on token reference and parameterization.

2Experimental evidence and practical usefulness still limited.Although authors test seven models, most main experiments only use WikiText-2 with 64-token context. This is quite short for current LLM setting. The result may change for long context, reasoning tasks or different data distribution. Also, adjacent layer similarity is measured with same input, but different layers normally receive different hidden states. So this comparison maybe not fully represent real model behavior.

Questions for authors
  1. What is the most important theoretical contribution? Since ADR decomposition mainly come from standard symmetric/skew split, and FFN directly included in reaction, can authors explain what new property cannot be obtained by simple algebraic decomposition?

2The advection and diffusion terms depend on token reference and model parameterization. If using different positive reference or function-preserving rescaling, how much the main depthwise conclusion will change? Can authors provide more quantitative sensitivity results for main claims? 3 The adjacent layer similarity use same saved input, but later block normally work on another hidden state distribution. Could this make the adjacent versus distant comparison unfair? Can authors test with different block distances or inputs sampled from later block's normal distribution?

4 comments

Question 1: Theoretical Contribution.

What is the most important theoretical contribution? Since ADR decomposition mainly come from standard symmetric/skew split, and FFN directly included in reaction, can authors explain what new property cannot be obtained by simple algebraic decomposition?

Response to Question 1:

We have refined the theoretical analysis by retaining cross-head terms in the transport Gram operator and absorbing HH into OijO_{ij}. The resulting full-block representation separates skew transport, dissipative diffusion, and total reaction (§3.1, Proposition 1; §3.2, Theorem 2). Crucially, the transported-discrepancy identity establishes that −Lπ-L_\pi is dissipative for every positive reference, without requiring stationarity (Eq. (8))—a property the standard symmetric/skew split alone does not guarantee. This dissipativity concerns the field on which the realized operator acts; it does not imply that −Lπ[X]Z(X)-L_\pi[X]Z(X) decreases the residual-state norm.

Beyond this refinement, the central theoretical question remains how to understand attention as a global operator T=[PijOij]T=[P_{ij}O_{ij}] and the complete Transformer block as a dynamical system (§2, Eqs. (2)–(4); §3.3, Eq. (12)). Our contribution characterizes the joint routing–transport geometry while preserving the nonlinear attention–FFN sequence, including the FFN’s evaluation at the attention-updated state (§3.2, Eqs. (10)–(11)). The symmetric/skew split is one analytical tool within this broader framework for studying transport, dissipation, and amplification.

Question 2: Reference and Parameterization Sensitivity.

The advection and diffusion terms depend on token reference and model parameterization. If using different positive reference or function-preserving rescaling, how much the main depthwise conclusion will change? Can authors provide more quantitative sensitivity results for main claims?

Response to Question 2:

We agree that component magnitudes and depth trends can depend on the reference and parameterization. Appendix C.3 evaluates 64 windows per checkpoint across seven checkpoints and 34 positive references. All 32 bounded random references preserve the advection and native-reaction trend signs observed under uniform weighting. Diffusion is more sensitive: in the explicitly separated head-mean comparison, changing from uniform to received-mass weighting shifts the late-minus-early diffusion norm-ratio contrast from −4.371 to +0.148 in GPT-2 Medium and from −1.269 to +1.548 in GPT-2 XL (Table 6). These results quantify sensitivity within the tested family, rather than establishing reference-independent trends.

In the revised native formulation, HH is absorbed into OijO_{ij}. Writing Oˉij=Oij/H\bar O_{ij}=O_{ij}/H, we have Aπ=HAˉπA_\pi=H\bar A_\pi, whereas Lπ=12I+H22Gˉπ−HSˉπL_\pi=\tfrac12I+\tfrac{H^2}{2}\bar G_\pi-H\bar S_\pi. Advection scales linearly with transport, while diffusion combines quadratic and linear terms. The matched native/RMS comparison illustrates this scale sensitivity: GPT-2 Small’s native diffusion quotient increases by 225.2%, whereas its RMS-normalized counterpart decreases from 0.579 to 0.560; RMS-normalized advection still increases from 0.765 to 1.564 (Appendix C.4, Table 7). RMS normalization changes the operator and is reported separately. For the function-preserving scalar rescaling, §4.2, Eq. (16) establishes that the advection action remains unchanged, while individual diffusion, potential, and total-reaction actions can change.

Question 3: Input Dependence of Adjacent-Block Similarity.

The adjacent layer similarity use same saved input, but later block normally work on another hidden state distribution. Could this make the adjacent versus distant comparison unfair? Can authors test with different block distances or inputs sampled from later block's normal distribution?

Response to Question 3:

We have added comparisons using both earlier and later blocks’ native inputs, with gaps 1,2,4,8,161,2,4,8,16 where available, across seven checkpoints and 64 windows per checkpoint (§5.2; Appendix C.3). Each block recomputes all native operations, and each input anchor uses a matched adjacent control. The adjacent advantage is positive in 27/27 nonadjacent checkpoint/gap comparisons at the earlier anchor and 21/27 at the later anchor, with simultaneous intervals excluding zero. All four GPT-2 checkpoints retain the advantage under both anchors, while Qwen/Llama exhibit input- and gap-dependent behavior (Table 2). These additional tests support local alignment across neighboring blocks beyond the original input protocol.

Thank you very much for the reviewer's results. THank you for having refined the theoretical analysis and the refined diffusion definition and tests with different references and input positions partly address my concerns. a more quick question: Some depthwise findings also depend on the reference and input right?

Thank you very much for your follow up. The motivation behind our depthwise study is to identify stable geometric structure in attention routing, feature transport, and the resulting attention operator across layers under the model’s native operating conditions. It is true that these operators depend on the input, while their geometric characterization also depends on the chosen reference. Inputs far outside the training distribution, such as unstructured random inputs, need not exhibit the same organization across depth. Similarly, changing the reference reweights token interactions and can alter the measured patterns. Our aim is therefore to identify which geometric structures persist across meaningful input and reference variations, and to characterize the conditions under which that stability holds. The sensitivity analyses establish the scope and limitations of these empirical findings. The exact ADR representation and the dissipativity guarantee remain valid under their stated assumptions, but do not imply universal depthwise trends.

Dear Reviewer EqW6,

Thank you again for your follow-up and helpful feedback. When you have a chance, we would appreciate your thoughts on whether our latest clarification addresses your remaining concern. We would be happy to discuss any further questions.

Best regards,
Authors of #52024

Reviewer_xFzHOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a mathematical mapping from repeated transformer architectures to a dynamical systems perspective that allows them to reinterpret the architecture as a advection-diffusion reaction process. This representation is extremely useful and the paper demonstrates how useful this representation is for questions that are commonly asked in LLM interpretability. The questions posed by the authors are clearly answered by the theory groundwork and is supported by the fact that their experiments show the component magnitudes actually changing through depth and local alignment of neighbors. Overall, it was a pleasure to read this paper.

Critical strengths

The paper's theoretical framework is very clear and it explains how apparent stability across layers can be dominated by new terms. The way the paper is written, there are many clear generalizations that the authors can now pursue. In order to establish this, you will have to look a little through the appendix to see how they arrive at Theorem 2.

Critical weaknesses

There is no specific critical weakness. I believe that there are some parts that can be validated more e.g. the connection between the geometric decomposition and how it leads to task computation but this would be a short extension of the paper.

Questions for authors

I have no questions for the authors. I believe the paper is written clearly and I recommend it for acceptance.

Comments on the paper

1 comment

We thank the reviewers for encouraging us to clarify ADR's geometric meaning and its contribution beyond a symmetric/skew-adjoint split (EqW6-W1/Q1; ePEz-W2/W3). Their suggestions led us to separate transported-discrepancy dissipation from gain/loss within the existing representation. The central contribution remains an exact ADR representation of the complete native sequential pre-norm Transformer block at evaluation time.

Refining the diffusion term in response to the reviewers’ questions. The earlier self-adjoint coordinate combined dissipative and gain/loss effects. In response, we make these contributions explicit within the same attention representation.

We adopt the native head-sum convention Oij=HOˉijO_{ij}=H\bar O_{ij}, T=HTˉT=H\bar T, with bars denoting the previous head-averaged convention, and construct the split on the complete attention operator. For fixed realized coefficients, let T=[PijOij]T=[P_{ij}O_{ij}] with nonnegative row-stochastic PP. A fixed positive reference π\pi defines ⟨x,y⟩π=∑iπixi⊤yi\langle x,y\rangle_\pi=\sum_i\pi_i x_i^\top y_i and the corresponding adjoint T∗πT^{* \pi}. Set Sπ=(T+T∗π)/2\mathcal S_\pi=(T+T^{* \pi})/2 and Aπ=(T−T∗π)/2\mathcal A_\pi=(T-T^{* \pi})/2. Define

\mathcal G_ \pi &=\operatorname{blockdiag}_ {j} \left(\frac{1}{\pi_j}\sum_i\pi_iP_{ij}O_{ij}^{\top}O_{ij}\right),\\ \mathcal L_ \pi &=\tfrac12(I+\mathcal G_\pi)-\mathcal S_\pi, \qquad \mathcal V_\pi=\tfrac12(\mathcal G_\pi-I),\\ T&=I+\mathcal A_\pi-\mathcal L_\pi+\mathcal V_\pi. \end{aligned}$$ The self-adjoint generator is thereby resolved as $\mathcal S_ \pi-I=-\mathcal L_ \pi+\mathcal V_ \pi$, with $$\langle x,\mathcal L_ \pi x\rangle_ \pi = \frac12\sum_ {i,j}\pi_ iP_ {ij} \|x_ i-O_ {ij}x_ j\|_ 2^2 \ge0.$$ This identifies $-\mathcal L_ \pi$ as dissipative and $\mathcal V_ \pi$ as token-diagonal gain/loss at fixed coefficients, for general learned transports and any fixed positive reference. In the native block, the guarantee is $\langle Z,-\mathcal L_ \pi Z\rangle_ \pi\le0$ on $Z=N_ 1(X_ \ell)$; finite-step residual-state contraction is a separate property. **Preserving the native attention–FFN composition (EqW6-W1/Q1).** The refined coordinates give the exact full-block update $$F:=X_{\ell+1}-X_\ell =\mathcal A_\pi Z-\mathcal L_\pi Z+R^{\mathrm{tot}}, \qquad R^{\mathrm{tot}} =(I+\mathcal V_\pi)Z+b+\psi(X^a),$$ where $X^a=X_\ell+TZ+b$ and $\psi=\mathrm{FFN}\circ N_2$. The FFN retains its native evaluation at the attention-updated state and remains separately identifiable within total reaction. Total reaction inherits cross-token input dependence through both the potential coefficients and $X^a$. At the same input and reference, the native advection, symmetric-transport and bias/FFN action vectors, together with the complete block output, are preserved. **Connecting the refinement to the measurements (ePEz-W3/Q1).** Responding to the request for clearer interpretation of component magnitudes, Section 5.1 reports the state-dependent diagnostics in Eq. (17), for $\|Z\|_\pi>0$: - **Skew action:** $\omega_ A=\|\mathcal A_ \pi Z\|_ \pi/\|Z\|_ \pi$. - **Diffusion dissipation:** $\rho_D=\langle Z,\mathcal L_\pi Z\rangle_\pi/\|Z\|_\pi^2$. - **Signed potential gain:** $g_V=\langle Z,\mathcal V_\pi Z\rangle_\pi/\|Z\|_\pi^2$, the signed quadratic contribution of the attention potential alone. The profiles cover eight checkpoints under uniform token weighting. Complete component strengths and the native FFN contribution are reported separately. Their joint interpretation uses $$-\mathcal L_\pi Z+(I+\mathcal V_\pi)Z=\mathcal S_\pi Z,$$ which makes the cancellation of the shared Gram contributions explicit. Guided by the reviewers’ questions, the revision makes the analytical value of the existing full-block representation more explicit. It connects the refined coordinates to constraints on component gains that preserve causal support at fixed coefficients (§4.1), action invariance under the specified function-preserving rescaling (§4.2), and approximation-error assessment through the native attention–FFN composition (§§4.3 and 5.3). Together, these analyses explain how to interpret component measurements, identify preserved actions, and evaluate attention modifications at the complete-block level. **Corresponding manuscript updates.** - **Sections 2–3:** state the native head-sum convention, refine Proposition 1 with the explicit Laplacian/potential construction, and express Theorem 2 in the resulting coordinates. - **Section 4 and supporting proofs:** carry the refined definitions through the operator identities and theoretical consequences. - **Section 5.1, Figure 2 and Appendix C.3:** present geometric diagnostics in the main text and complete Strength profiles with supporting FFN measurements in the appendix. - **Introduction, Conclusion and figure explanations:** present the same full-block contribution with its sharpened geometric interpretation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.