acceptodds
Withdrawn from ICLR 2027

Global Exponential Convergence of Two Layer Linear Network Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We prove global exponential (linear) convergence with an explicit rate in the rich scaling for wide two-layer linear networks trained with smooth Polyak–Lojasiewicz predictor losses. Gradient flow in the factors closes exactly in terms of a finite-dimensional Bures flow of the neuron law covariance, in which the predictor dynamics are preconditioned by hidden covariance blocks. Mean-field conservation laws provide uniform spectral lower bounds on the hidden preconditioning blocks when the initial covariance satisfies a spectral support gap condition. This condition encompasses positive definiteness while still allowing for singular initializations. For an initial covariance , the loss converges to the global minimum with linear rate at least , where is the PL constant. We establish stability of this rate under finite-width sampling, as well as global convergence of factor gradient descent for an explicit stepsize interval depending on smoothness, the initial loss, and conserved spectral margins. Our argument extends layerwise to deep linear ResNets, subject to a residual-path bound. In the case of heavy-ball momentum, training dynamics close instead over positions and velocities in terms of a lifted phase covariance. Linear convergence holds under an explicit condition on the energy and damping, specifying a window of admissible dampings. For two-scale white initializations, this interval is nonempty for sufficiently large position scales, with a fixed initial loss gap and velocity covariance. Numerical experiments illustrate the covariance geometry and compare the predicted and observed rates.

Primary area:
optimization
Keywords:
Linear neural networks, global convergence, conservation laws, mean-field analysis, Bures-Wasserstein gradient flows, Polyak-Łojasiewicz inequality, gradient descent, heavy-ball momentum, deep residual networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pzdjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies the gradient flow of matrix factorization problems, i.e. minimization problems of the form ℓ(VU⊤)\ell(VU^\top). They show exponential convergence under a Polyak-Lojasiewicz (PL) condition on the loss ℓ\ell and a condition on the initialization. The rate is explicit in the PL constant and in the initialization.

Extensions are made to gradient descent, deep linear residual networks and the inertial gradient flow.

Critical strengths

The identification and the exposition of the problem structure in Section 2, was new to me, and I found it particularly elegant. It nicely prepares the exposition of Thm 3.1. However, it seems that there are similar results in the literature (see below).

Critical weaknesses

Core contribution lacks novelty, insufficient comparison to the related work. The paper of Min et al. (2021)- on the explicit role of initialization on the convergence and implicit bias of overparametrized linear networks - provide a result (Thm 1) remarkably similar to the Thm 3.1 of this paper. The only difference is that the former restricts to ℓ\ell being quadratic (least-squares problem) while this paper proposes the generalization to the PL case. However, the paper of Min et al. is not cited, and there is no comparison to related work in the paper suggesting such a similarity with previous works.

More generally, the related work provides related papers with vague descriptions, but no comparison with related work is provided when the results are provided. Given the example provided above, this shows that the contributions of the paper are not sufficiently well identified and thus the paper can not be accepted.

Poor writing style. The writing is very dense, lacks some general perspective. In some parts I could not understand the jargon (see first paragraph of the intro, for instance). The paper enumerates many variations of a result (Thm 3.1), rather than explaining its core contribution. There is a surprising repetition between Remark 3.2 and Remark 3.5. These problems might be due to an excessive use of a LLM in writing.

Questions for authors

Explain the novelty (both in terms of setting and of technical difficulty) in Thm 3.1.

Minor suggestions:

  • l.148: should be ∇ΣF=G(Σvu)/2\nabla_\Sigma F = G(\Sigma^{vu})/2.
  • Figure 2(b): I don't understand the caption
  • l.330: what is a 'layer'? Is it one of the factors or two factors with a residual connection?
2 comments
AuthorsOpenReview ↗

Response to Reviewer pzdj

The concern about overlap with Min et al. (2021) is addressed in the common responses above.

l.148: should be ∇ΣF=G(Σvu)/2\nabla_\Sigma F=G(\Sigma^{vu})/2.

Thank you, we will write the argument explicitly to avoid ambiguity.

Figure 2(b): I don't understand the caption

The panel starts from Σ0=Idu+dv\Sigma_0=I_{d_u+d_v} and replaces one diagonal entry in each of the uu and vv blocks by r∈[0,1]r\in[0,1], leaving the other entries unchanged. Thus g(Σ0)=2rg(\Sigma_0)=2r, and the certified loss-decay exponent is 4κr4\kappa r for r>0r>0. The dashed curve shows the r=1r=1 reference rate. At r=0r=0, the support-gap guarantee vanishes. We will revise the caption to specify this initialization and explain the reference curve.

l.330: what is a 'layer'? ...

In the residual network setting, "layer" refers to the residual layer obtained by composition of the predictor (the product of two factors) and a residual connection. We realize this terminology clashes with the rest of the paper, where "layer" is used to refer to one of the factor matrices, and we will clarify this in future versions.

Poor writing style ...

Thank you for this feedback. We will revise the exposition to give a clearer overview of the contributions of the paper. We take responsibility for all work and its presentation, including any AI-assisted material.

Thank you for your answer. I understand your decision to withdraw and resubmit.

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

The paper studies global exponential convergence of Gradient Flow (Theorem 3.1), Gradient Descent (Theorem 3.6), and Heavy Ball Momentum method of two-layer linear networks in the mean-field scaling.

Critical strengths
  1. Theorem 3.6 establishes global convergence of GD under small stepsize assumption with explicit small stepsize size and convergence rate under smooth PL predictor loss assumption (Assumption 1.1).
  2. The heavy-ball momentum convergence result (Theorem 5.2) appears to be a novel result (although the mechanism itself may be not particularly new).
Critical weaknesses

Overall their novelty is not very clear from the paper. In fact, there are a few novelty concerns regarding their main theorems. Important references appear to be missing, and I think the authors should add very explicit comparison with existing results for each of their main theorem:

  1. The main claim of Theorem 3.1 is already established by part 1 of Theorem 1 of Min et al. (https://proceedings.mlr.press/v202/min23d.html) after adjusting for the normalization scale. The limit part may be novel but it seems to be a relatively straightforward consequence of the main claim. I think the authors should provide a very detailed comparison with Min et al.
  2. Theorem 3.6 appears to be a valid extension of Theorem 3.2, Xu et al (https://arxiv.org/pdf/2505.11664). While the submission relaxes strong convexity assumption in Xu et al. to smooth PL assumption, looking at Theorem 3.1 of Xu et al., it leaves impression that not a small portion of the argument has already been established by Xu et al. I think the authors should make it clear their exact contribution and give enough credit to the existing work.
  3. Theorem 5.1 appears to be a novel result but perhaps the underlying mechanism may be well established. It helps understand the novelty further if the authors provide what is known in the existing literature.

There is also a positioning concern:

  1. The authors position this paper as a mean-field analysis. However, neither of their main theorems appears to require mean-field regime. It seems that mean-field interpretation only appears in Appendix B and not essential to prove their main theorems. In fact, none of their main theorems require large width. Although their mean-field interpretation is interesting, I think their results should have been presented as finite-dimensional factorization results. In fact, the normalization scaling by 1/n is the only difference between Theorem 3.1 of the submission and Theorem 1 of Min et al.
Questions for authors
  1. For Theorem 3.1, what exactly is different from Theorem 1 of Min et al.? Please add a very explicit comparison.
  2. For Theorem 3.6, what technical breakthrough made it possible to obtain this result given Xu et al.'s existing work?
  3. For Theorem 5.1, can you clarify the proof is done under existing framework or is there any technical breakthrough?
  4. Can you clarify if there is any genuine mean-field specific argument in your main theorems? I understand Appendix B discusses interesting mean-field interpretation, but it does not seem they are essential for your main theorems.
2 comments
AuthorsOpenReview ↗

Response to Reviewer BVHf

Theorem 3.6 appears to be a valid extension of Theorem 3.2, Xu et al (https://arxiv.org/pdf/2505.11664).

We agree that convergence of GD under strong convexity is established in Theorem 3.2, Xu et al. (2025), we will explicitly credit this work and clarify the scope of our results.

The contribution of our result is to obtain global control under smooth PL assumptions without relying on strong convexity: Xu et al. use strong convexity to bound the predictor's distance to a fixed minimizer (Appendix E, Eqs. (114)–(115)). Instead, our proof uses descent and PL to control the total predictor-gradient length. Under the provisional covariance bounds, with Ek=∇ℓ(Σkvu)E_k=\nabla\ell(\Sigma_k^{vu}), we obtain

∑j<Nτ∥Ej∥F≤I⋆:=4Δ0/κg(Σ0),∥ΣN∥op≤∥Σ0∥ope2I⋆.\sum_{j<N}\tau\|E_j\|_F\le I_\star:=\frac{4\sqrt{\Delta_0/\kappa}}{g(\Sigma_0)},\qquad\|\Sigma_N\|_{\mathrm{op}}\le\|\Sigma_0\|_{\mathrm{op}}e^{2I_\star}.

Together with a bound on ∑j<Nτ∥Ej∥F2\sum_{j<N}\tau\|E_j\|_F^2, this controls the accumulated defect in the transported covariance inequalities and closes the induction.

Theorem 5.1 appears to be a novel result but perhaps the underlying mechanism may be well established. It helps understand the novelty further if the authors provide what is known in the existing literature.

We distinguish Proposition 5.1, which describes conserved conformal symplectic structure, from Theorem 5.2, which establishes global convergence under an explicit initialization condition.

To the best of our knowledge, the contribution of Theorem 5.2 is novel. The mechanism behind conformal conservation (Proposition 5.1) is linked to classical concepts: conformal symplectic structure for damped Hamiltonian dynamics and its relationship to heavy-ball optimization are studied for instance by França et al. (2020), Conformal Symplectic and Relativistic Optimization. However, to our knowledge, it has not been leveraged to study global convergence rates in non-convex optimization. We will make this connection explicit.

For Theorem 5.1, can you clarify the proof is done under existing framework or is there any technical breakthrough?

We distinguish Proposition 5.1, which describes conserved conformal symplectic structure, from Theorem 5.2, which establishes global convergence under an explicit initialization condition.

To the best of our knowledge, the contribution of Theorem 5.2 is a technical breakthrough for the analysis of momentum dynamics and we summarize the strategy below.

We first obtain exponential energy decay conditional on a lower and upper bound on the position spectrum of Σt(u,v)\Sigma_t^{(u, v)} (Lemma C.24), then use this decay to control the accumulated deformation in the position covariance. The final convergence theorem is then provided using a bootstrap argument.

The construction of the modified energy V=E+ϵΔ˙\mathcal{V} = \mathcal{E} + \epsilon \dot{\Delta} used as a Lyapunov function for the conditional convergence argument (see proof of Lemma C.24) has been used previously in the study of second order dynamics, see e.g. Apidopoulos et al. (2021), Convergence rates for the Heavy-Ball continuous dynamics for non-convex optimization, under Polyak-Lojasiewicz condition. We will explicitly mention this in the revised proof.

Proposition 5.1 states the necessary conservation identity for the conformal symplectic setting that we consider. Its proof relies on application of the existing conformal symplectic framework; our contribution here is the specialization to the linear network's closed phase-covariance dynamics. The main convergence argument for the momentum case is in Theorem 5.2: the conservation law of Proposition 5.1 does not prevent degeneration of the position covariance Σt(u,v)\Sigma_t^{(u,v)}.

The authors position this paper as a mean-field analysis...

Our main convergence proofs do not require an infinite-width limit or a measure-PDE argument. We use the measure formulation to describe population and empirical neuron distributions through the same exact covariance dynamics. This viewpoint allows us to specify a population covariance Σ~0\widetilde\Sigma_0 and ask how its convergence guarantee transfers to randomly initialized finite networks sampled from that law. We will clarify that the convergence proofs are finite-dimensional and that the measure formulation provides this population-to-sampling interpretation. We will also revise the title "Mean Field Conservation Laws" of Section 2 to make this distinction clearer.

Thanks for your response. I respect your providing with us detailed answers despite your decision to withdraw.

Reviewer_8YXkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper shows that two-layer linear network gradient flow can be formulated as a Bures flow on weight covariances. It uses conserved spectral quantities to establish exponential convergence for smooth, globally PL predictor losses under a positive initial support gap. It also provides finite-width sampling and small-stepsize GD guarantees, with extensions to deep linear ResNets and continuous-time heavy-ball dynamics.

Critical strengths

The paper shows that finite-width networks initialized by subgaussian sampling retain, with high probability, nearly the convergence rate certified for the population covariance. The required width scales with the population covariance rank, which can be as small as min⁡(du,dv)\min(d_u,d_v) for a specific initialization with one factor set to zero. This is smaller than the width covered by the prior guarantee discussed in my weaknesses, namely previous results required a width of at least du+dvd_u + d_v. However, this quantitative sampling result does not sufficiently offset the overlap in the main convergence guarantee, whose rate coincides with [1] in the overlapping regression setting.

[1] Min et al. “On the Explicit Role of Initialization on the Convergence and Implicit Bias of Overparametrized Linear Networks. ICML 2021.

Critical weaknesses
  1. My main concern is that the central result coincides with an existing and uncited guarantee. The paper notes in Prop. B.3 that the nonzero spectrum of JΣJ\Sigma equals that of the Gram imbalance −Γ/n-\Gamma/n. Combined with (5), this gives g(Σ0)=max⁡(λdu(−Γ0/n),0)+max⁡(λdv(Γ0/n),0)g(\Sigma_0) = \max(\lambda_{d_u}(-\Gamma_0/n), 0) + \max(\lambda_{d_v}(\Gamma_0/n), 0). For full-rank inputs and squared loss, this is exactly the effective level of imbalance Δ‾\underline{\Delta} of [1]. One can also see that theorem 3.1’s exponent 2κg(Σ0)2\kappa g(\Sigma_0) equals their rate 2λr(Σx)Δ‾2\lambda_r(\Sigma_x)\underline{\Delta}, and it is obtained by the same argument. The mean-field normalization does not separate the settings, since rescaling (U,V)↦(U,V)/n(U, V) \mapsto (U, V)/\sqrt{n} turns the flow (1) into standard gradient flow at every width. In parts of the overlap, the result is weaker than prior work. In particular, the extended version [2] adds a margin term, so its rate is strictly larger whenever the initialization has positive margin. The authors need to explain how their results are different or stronger than those in [1] and [2]. I would also appreciate a discussion of how their covariance closure can be distinguished from the lifted block-matrix analysis of [3].

  2. The extensions do not compensate for the overlap with prior work: the ResNet results assume a bound on the residual products along the whole training trajectory, which cannot be checked at initialization and which the paper itself shows positive margins do not imply. The heavy-ball certificate applies only at large initialization scales, where its certified rate is below that of plain gradient flow from the same initialization. Since the main results are not sufficiently novel compared to prior work, the extensions have to be sufficiently strong and significant to justify acceptance, which I am not convinced is the case here.

[1] Min et al. On the Explicit Role of Initialization on the Convergence and Implicit Bias of Overparametrized Linear Networks. ICML 2021.

[2] Min et al. Convergence and implicit bias of gradient flow on overparametrized linear networks. 2022.

[3] Tarmoun et al. Understanding the dynamics of gradient flow in overparameterized linear models. ICML 2021.

Questions for authors
  1. Given the identification g(Σ0)=Δ‾g(\Sigma_0) = \underline{\Delta}, how do you position Theorem 3.1 relative to Min et al. (2021) and its extended version? Can your framework recover the margin contribution?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8YXk

The ResNet results assume a bound on the residual products along the whole training trajectory, which cannot be checked at initialization and which the paper itself shows positive margins do not imply.

We agree that the ResNet guarantees are conditional on a residual-path bound throughout training; positive initial layer gaps alone do not establish this bound. Under this condition, Theorems 4.1 and 4.2 give exponential loss-decay rates that remain uniform in depth, together with convergence of the layer covariances. The factors in these rates, (1−B/S)2(S−1)(1-B/S)^{2(S-1)} at finite depth and e−2Be^{-2B} at continuous depth, are sharp under the respective path bounds. We will make the scope of these guarantees and their assumptions more prominent.

The need for this requirement is illustrated by the orientation-preserving nature of Neural ODEs. We view this as an insight into the mechanics of continuous-depth ResNet training, and not as a drawback or limitation of our analysis. Our counterexample in Remark C.19 shows why positive layer gaps alone are insufficient for an unconditional convergence claim. We will rework our exposition to point this out more prominently in the main text.

The heavy-ball certificate applies only at large initialization scales...

Large initialization scale described in Corollary 5.3 is one sufficient regime for our certificate (and is indeed the one emphasized by our illustrations). However it is not a necessary condition for Theorem 5.2 to apply. The general condition of Theorem 5.2 depends on the initial position covariance Σ0(u,v)\Sigma_0^{(u, v)} and the mechanical excess energy E0=ℓ(S0)−ℓ⋆+K0\mathcal E_0=\ell(S_0)-\ell_\star+\mathcal K_0, where K0=12tr(Σ0(p,q))\mathcal K_0=\tfrac12{tr}(\Sigma_0^{(p, q)}) is the initial kinetic energy. With bounds 0≺a0I⪯Σ0(u,v)⪯b0I0 \prec a_0 I \preceq \Sigma_0^{(u, v)} \preceq b_0 I on the initial position covariance, and fixed s,γ>0s,\gamma>0, the quantities in the theorem satisfy

\frac{a_s\kappa\gamma}{4b_sL+2\gamma^2}>0, \qquad \mathfrak S_s(\gamma)=\frac{4}{\lambda_s(\gamma)} \sqrt{\frac{3\mathcal E_0}{2a_s}}\longrightarrow0 \quad\text{as }\mathcal E_0\longrightarrow0,$$ where $a_s=a_0e^{-2s}$ and $b_s=b_0e^{2s}$. Sufficiently small initial loss gap and kinetic energy can therefore also ensure $\mathfrak S_s(\gamma)<s$ at fixed position scale, including zero velocity initialization. Corollary 5.3 provides an alternative regime that allows fixed initial loss gap and kinetic energy and guarantees a nonempty admissible damping interval for sufficiently large position scale. > ... where its certified rate is below that of plain gradient flow from the same initialization... We do not claim to prove acceleration over gradient flow for all regimes. Outside the large position-scale regime, however, the certified decay exponent of Theorem 5.2 can exceed that of our gradient-flow bound. For instance, take $d_u = d_v = 1$, $\ell(S)=\tfrac12(S-10^{-8})^2$, initial position covariance $\Sigma_0^{(u, v)}=10^{-4}I_2$, and zero initial velocity, so $\Sigma_0^{(p, q)}=0$ and $\mathcal K_0=0$. Then $L=\kappa=1$, $S_0=0$, and $\mathcal E_0=5\times10^{-17}$. With $s=0.1$ and $\gamma=0.01$, Theorem 5.2 gives a higher certified decay exponent than the gradient-flow bound, in the same time variable: $$\mathfrak S_s(\gamma)\approx 0.00322<s,\qquad\lambda_s(\gamma)\approx 1.189\times10^{-3}>2\kappa g(\Sigma_0^{(u, v)})=4\times 10^{-4}.$$ > I would also appreciate a discussion of how their covariance closure can be distinguished from the lifted block-matrix analysis of [3]. Our covariance $\Sigma$ coincides, up to normalization and coordinate conventions, with the matrix $P=RR^\top$ studied by Tarmoun et al. (2021). The distinction between our work and theirs concerns how the covariance dynamics are analyzed: for a squared loss and scalar imbalance, Tarmoun et al. obtain a constant-coefficient matrix Riccati equation, deriving an explicit solution and convergence guarantees. On the other hand, our analysis works directly with the autonomous covariance equation $$\dot\Sigma=-G(\Sigma_{vu})\Sigma-\Sigma G(\Sigma_{vu}), \qquad G(S)=\begin{pmatrix}0&\nabla\ell(S)^\top\\\nabla\ell(S)&0\end{pmatrix},$$ and this holds for general predictor losses and initializations. For general losses and initializations, this equation need not admit the constant-coefficient Riccati reduction obtained under their additional assumptions. We emphasize its interpretation as a Bures gradient flow with congruences preserving $J={diag}(I_{d_u},-I_{d_v})$.
AuthorsOpenReview ↗

Response to Reviewer 8YXk, continued

Can your framework recover the margin contribution?

Yes, by incorporating the sharper spectral estimate of Min et al. (2022), Lemma 8, Proposition 1 and Theorem 3. For simplicity, consider du=dv=dd_u=d_v=d and

ℓ(Σvu)=12∥Σvu−Σ⋆vu∥F2.\ell(\Sigma^{vu})=\frac12\|\Sigma^{vu}-\Sigma_\star^{vu}\|_F^2.

Write Σtvu=VtUt⊤/n\Sigma_t^{vu}=V_tU_t^\top/n, Δt=ℓ(Σtvu)−ℓ⋆\Delta_t=\ell(\Sigma_t^{vu})-\ell_\star, and Γ0=V0⊤V0−U0⊤U0=−W0⊤JW0\Gamma_0=V_0^\top V_0-U_0^\top U_0=-W_0^\top JW_0. Here ℓ⋆=0\ell_\star=0. For this comparison, define

g0=g(Σ0),δ±=[λmax⁡(∓Γ0/n)]+−a±(Σ0),η=[σmin⁡(Σ⋆vu)−∥Σ0vu−Σ⋆vu∥F]+.g_0=g(\Sigma_0),\qquad\delta_\pm=[\lambda_{\max}(\mp\Gamma_0/n)]_+-a_{\pm}(\Sigma_0),\qquad\eta=[\sigma_{\min}(\Sigma_\star^{vu})-\|\Sigma_0^{vu}-\Sigma_\star^{vu}\|_F]_+.

Loss monotonicity and the singular-value perturbation bound give σmin⁡(Σtvu)≥η\sigma_{\min}(\Sigma_t^{vu})\ge\eta. Applying the spectral estimate of Min et al. (2022) to the covariance blocks yields

\ge c_\eta :=\frac{1}{2}\left[ \sqrt{(\delta_++g_0)^2+4\eta^2}-\delta_+ +\sqrt{(\delta_-+g_0)^2+4\eta^2}-\delta_- \right].$$ The covariance dissipation identity then gives $$-\dot\Delta_t ={tr}(E_t^\top E_t\Sigma_t^u) +{tr}(E_tE_t^\top\Sigma_t^v) \ge2c_\eta\Delta_t, \qquad E_t=\Sigma_t^{vu}-\Sigma_\star^{vu},$$ hence $\Delta_t\le\Delta_0e^{-2c_\eta t}$. When $\eta=0$, $c_\eta=g_0$, recovering our exponent; when $\eta>0$, $c_\eta>g_0$, and convergence can be certified even if $g_0=0$. Thus the covariance formulation can accommodate the margin contribution.

Thank you for addressing my comments. I appreciate and understand your decision to withdraw and resubmit.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response to all reviewers

We thank the reviewers for their careful reading of our work and for identifying the overlapping guarantees in Min et al. (2021, 2022, 2023). We were not aware of these overlaps when formulating our contribution claims. As a result, the novelty claim for the gradient-flow convergence guarantee of Theorem 3.1 needs to be retracted. We apologize for the oversight.

  • Min et al. (2021), Theorem 1 establishes exponential convergence for squared-loss regression using the conserved Gram imbalance. In the overlapping full-rank setting, our covariance gap equals their imbalance condition, and the loss-decay exponent in our Theorem 3.1 coincides with theirs.
  • Min et al. (2022), Theorem 3 strengthens the guarantee of Min et al. (2021) by incorporating a margin term. Our stated bound is weaker when the margin is positive. Below, in response to Reviewer 8YXk, we show how their sharper estimate can be incorporated into the covariance formulation.
  • Min et al. (2023), Theorem 1, part (1) gives the two-layer gradient flow guarantee for PL losses, with the same loss-decay exponent as our Theorem 3.1.

We will substantially revise the paper to rectify missing attribution to the prior work we missed. We will separately assess the contributions of our sampling, gradient descent, ResNet, and momentum results.

Since the revisions to the paper will necessarily be extensive, we intend to withdraw the paper from the present review cycle to respect the valuable time of reviewers, ACs and SACs. We address the reviewers' individual questions below.

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

Related papers

Discussion (0)

Sign in to comment.