acceptodds
Under review as a conference paper at ICLR 2027

MultiNormStep: Leaving the Corner of Steepest Descent

Abstract

First-order optimizers are the workhorse of large language model training, and after a decade in which AdamW was the default, Muon has recently become the first optimizer to beat a well-tuned AdamW across model and data scales. Yet Muon acts only on the 2D matrix weights; the remaining parameters (the normalization scales, biases, and most consequentially, the token embedding and LM head) are left as a compromise to a default AdamW. This split looks like two unrelated optimizers bolted together, but through the lens of steepest descent the pieces are more related than they appear: signSGD and Muon are the -norm corner read in two gauges, the box on coordinates and the Schatten- ball on singular values, and Adam is in turn a signSGD that shrinks every coordinate strictly inside that box. SignSGD sits aggressively at the corner and Adam timidly within it, yet neither leaves it. Rather than commit to a single corner, we constrain a of norms through a Young function, obtaining (MN-step): a family with a closed-form coordinatewise update that both the single-norm methods (signSGD, PowerStep, normalized SGD) and new interior maps, a bounded and an unbounded . The identical construction lifts from coordinates to the spectrum, where it recovers Muon, so these maps and Muon are two faces of steepest-descent principle. Under full-parameter training our bounded sigmoid step stays on par with a carefully tuned AdamW across GPT-2 124M/350M and Qwen3 0.6B/4B; its decisive role is the group Muon leaves to AdamW, where, in a Muon-hybrid that fixes Newton-Schulz Muon on the 2D weights and swaps only that branch, the sigmoid step beats the AdamW branch on validation at Qwen3 0.6B and 4B and again outperforms AdamW at two scales of a proprietary mixture-of-experts model (Model-X). These consistent gains show that widening the geometry beyond a single norm genuinely helps, precisely on the parameters Muon leaves to AdamW.

Primary area:
optimization
Keywords:
optimization, large language model, steepest descent

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

The paper introduces MultiNormStep, a family of coordinatewise optimizers. Each takes a steepest-descent step under an Orlicz-modular constraint with updates Δi=−η⋅sign⁡(mi)⋅h(∣mi∣λ)\Delta_i = -\eta\cdot\operatorname{sign}(m_i) \cdot h \left(\frac{|m_i|}{\lambda}\right). Two weight profiles give two different maps: "log: h(x) = ln(1+x)" and "sigmoid: h(x) = x/(1+x)". SignSGD, PowerStep, and normalizedSGD are presented as special cases. Applying the same construction to the singular values gives a "matrix face" whose limiting case is Muon (not implemented).

The theory gives an O(1/T)O(1/\sqrt T) convergence bound up to an O(σ)O(\sigma) neighborhood for momentum applied before the map with λ\lambda held fixed. The experiments have 3 parts:

  1. full parameter training on GPT 2 (124M/ 350M) and Qwen3 0.6B/ 4B. The maps end with about 0.01 nats gained over AdamW on validation and the log map must be clipped to stay stable at 4B.
  2. a Muon hybrid on Qwen3 0.6B/4B and a proprietary MoE, where Newton–Schulz Muon is held fixed on the 2D matrices and the maps replace AdamW on the embedding, LM head, norm and bias parameters, ending 0.001–0.03 nats below the AdamW branch on training loss
  3. ablations of λ's pooling granularity and EMA coefficient, and a measured saving of about one fp32 optimizer buffer relative to AdamW. Code for both codebases are released.
Critical strengths
  1. The paper targets a real gap which appears in practice. In current Muon recipes, the non matrix parameters (embeddings, norms, bias, LM head) are left to AdamW. The paper asks directly what this branch should use, with a drop-in optimizer that keeps one buffer per parameter, and evaluates it across three codebases and model families up to Qwen3 4B and a production MoE, in both full-parameter and hybrid settings, with a per-method learning-rate grid on GPT 2. The memory accounting is exact and honest about its sharding limitation, and the diagnostics that expose the paper’s own confounds are published.

  2. One robust design finding, made visible by the scale of the study. Whether the map is bounded decides stability at scale: at Qwen3-4B the unbounded log map diverges at every learning rate tried (2.5e−4, 3e−4, 5e−4) unless its update is clipped, while the bounded sigmoid map trains normally at the same rate. This does not appear at 0.6B or on GPT-2, so a smaller study would have missed it.

Critical weaknesses
  1. The theoretical framework and both "new" maps exist in uncited peer-reviewed literature! The update x+=x−η∇ϕ∗x^+ = x - \eta \nabla \phi * is dual-space preconditioning (https://epubs.siam.org/doi/10.1137/19M130858X). (https://arxiv.org/pdf/2502.08532) analyze exactly this iteration for separable and isotropic kernels. Their Table 1 contains both of this work's Young functions. Their example 1.6 applied the separable sigmoid kernel coordinate wise and notes it recovers Adam when beta_1 = beta_2 = 0. The (https://arxiv.org/pdf/2510.11312) work follows up with heavy ball momentum and a stochastic analysis. Row pooled second moments that remove Adam's v buffer exist in Adam-mini (https://arxiv.org/abs/2406.16793).

To be precise about the overlap, the final results are rederived with a different method, but the results stand the same. The convergence proof (Lemma 4, 5, Theorem 2, momentum on raw gradient) is the paper's own novel contribution. The experiments are new.

The “remarkably little attention” phrasing of the non-Muon branch omits the published Adam-free recipes for exactly these parameters, Scion (ICML 2025; https://arxiv.org/abs/2502.07529) and modular duality’s embedding dualization (ICML 2025; https://arxiv.org/pdf/2410.21265).

  1. The reported benefits are too small relative to the paper’s own variability. The experimental gains seen are on a single seed and hence is a bit unclear on how to claim benefits with confidence. The Muon-hybrid validation margins over AdamW are 0.010/0.007 nats at 0.6B and 0.001/0.004 at 4B, where all five branches sit within 0.007 nats (Table 11); the paper’s own learning-rate sweep moves the log map by 0.02–0.04 nats between adjacent rates at 124M (Table 10). Model-X is training loss only, on a proprietary model and corpus. The 4B runs see $\approx$2.6 tokens per parameter, so “at scale” and “early in training” are confounded. If possible I would like the authors to perform the Qwen 0.6B and Qwen 4B run with 2 more seeds and at a larger training scale to allow for a confidence interval and substantiate the claims further.

  2. The hybrid comparison confounds geometry with effective step size. All branches share one learning rate, at which the paper’s own Table 12 shows the maps’ per-coordinate update RMS on the vector branch is 3.4× (sigmoid), 5.7× (log) and 6.8× (signSGD) that of AdamW. There is no vector sweep of AdamW or RMS-matched control, and the 4B training loss ordering of the five branches tracks their update RMS. The two controls that Lemma 3 covers identically are not run. It would be very interesting and crucial to know how the identity link h(x) = x with the same row-λ\lambda and momentum (essentially Adam-mini with heavy-ball momentum), which isolates the nonlinearity, and coordinatewise clipping clip(m/λ\lambda, −1, 1), which isolates smooth saturation, operate.

Minor:

  1. the bound is vacuous once σ/G ≥ (1−γ)^{3/2} ≈ 0.032 at the used γ = 0.9
  2. I think App. F.2 should read λ ≈ RMS(g)/√(2γ), not RMS(g)/(√2·γ)? Also, the constant cannot be “absorbed into the learning-rate sweep” for a nonlinear h. This might have to be rephrased.
  3. eq. 8 is printed with the ascent sign
  4. Can the x axis for the training runs be reported in terms of tokens rather than steps? It would be more standardized and would be easier for readers!
  5. Can you report the validation curves and numbers in the main text and defer the training numbers to the appendix?
Questions for authors
  1. How many independent seeds underlie each entry of Tables 4, 9 and 11, and what is the standard deviation of final validation loss for AdamW and the sigmoid map at all model scales (GPT 2 and Qwen 3 runs)?
  2. What does the identity link h(x) = x (same row λ, momentum, β₂ and learning rate) and coordinatewise clipping clip(m/λ, −1, 1) look like? How does this compare with the two explored functions as a baseline? This would let us know how the nonlinearity plays a role in the optimization!
  3. Train vs validation at 4B. Why does the 4B hybrid training-loss margin of 0.030–0.033 nats (Table 4) collapse to 0.001–0.004 on validation (Table 11)?
  4. How does this paper position itself with that of https://arxiv.org/pdf/2502.08532, https://arxiv.org/pdf/2510.11312? What do you claim as new relative to these works?
  5. Why do runs diverge at lower learning rates than the selected one for Model-X (Table 8)? Can you report validation loss or perplexity?
1 comment

Thank you for your review.

Prior work and novelty. [Revised, §2 and intro]

Oikonomidis et al. (ICML 2025, Table 1) give the same scalar log and rational-sigmoid maps and their generating functions; their experiments use radial variants on a toy objective and radial sigmoid on full-batch MNIST. Oikonomidis et al. (NeurIPS 2025) accumulate preconditioned gradients and analyze the base method without momentum, whereas we apply the link after accumulating raw stochastic gradients. We now credit both, together with Maddison et al. (dual-space preconditioning), Adam-mini, Scion, and the modular-duality framework, and we narrowed the "little attention" claim in the introduction to "comparatively little", noting these works touch the companion parameters. Our stated contribution is the unified Orlicz template for these links and their spectral counterparts, plus the LLM-pretraining and Muon-companion evaluation.

Confound between geometry and effective step size. [Revised, Tables 9 and 10; Open]

Two controls separate link shape from magnitude.

  • Table 9 (companion-LR sweep) shows AdamW is at its own optimum at the shared rate.
  • Table 10 adds the identity and hard-clip controls on Qwen3 0.6B. Identity diverges, confirming boundedness is necessary. The hard clip clip(m/λ,−1,1)\mathrm{clip}(m/\lambda,-1,1) beats AdamW (2.895 vs 2.899), so capping the step already helps; but both smooth maps beat the clip (log 2.889, sigmoid 2.892), isolating the Orlicz geometry from mere boundedness.

These are precisely the two controls you asked for. The identity link h(x)=xh(x)=x with the same row-λ\lambda and momentum (Adam-mini with heavy ball) isolates the nonlinearity, and the coordinatewise clip clip(m/λ,−1,1)\mathrm{clip}(m/\lambda,-1,1) isolates smooth saturation; both are now run in Table 10, and the companion-LR sweep of Table 9 supplies the AdamW vector sweep you noted was missing. They also answer the "ordering tracks update RMS" concern directly: the clip caps every coordinate at 11 yet the smooth maps still beat it (log 2.889, sigmoid 2.892 vs clip 2.895), so the ranking is set by link shape, not by raw step magnitude. A full RMS-matched sweep on the vector branch nonetheless remains Open, and Table 16 reports the per-coordinate update-RMS gaps (sigmoid 3.4×3.4\times, log 5.7×5.7\times, signSGD 6.8×6.8\times AdamW) that motivate it.

Seeds, variability, horizon, and held-out evaluation. [Revised, Tables 5 and 6]

We ran the two extra seeds you asked for. Table 5 now reports three seeds (42, 1234, 2026) for the Qwen3 hybrids: validation loss 2.903±0.0032.903 \pm 0.003 / 2.894±0.0052.894 \pm 0.005 / 2.895±0.0042.895 \pm 0.004 at 0.6B and 2.765±0.0032.765 \pm 0.003 / 2.763±0.0032.763 \pm 0.003 / 2.760±0.0032.760 \pm 0.003 at 4B (AdamW / log / sigmoid). This answers the variability concern directly: both maps improve on AdamW in every paired seed at both scales, and at 0.6B the ≈0.008\approx 0.008--0.0090.009 nat margin exceeds the run-to-run std (0.0030.003--0.0050.005). On the 0.020.02--0.040.04 nat swing the log map shows between adjacent learning rates: that measures learning-rate sensitivity for all parameters, not the specific non-matrix branch we focused on. The remaining tables are single-seed point estimates; we state this and treat Table 5 as the reproducibility evidence.

On Model-X being training-loss only: Table 6 now adds held-out evaluation on Wikipedia and a Pile test split at both scales, where both links again improve on AdamW, so the Model-X result no longer rests on training loss alone.

On the 4B horizon confound: we agree "at scale" and "early in training" are entangled at about 2.6 tokens per parameter (vs 17.5 at 0.6B), and we narrowed the scale claim accordingly. The training-loss gains of 0.033 / 0.030 nats shrink to 0.001 / 0.004 on validation because the held-out gap needs a longer horizon to open, not because the effect vanishes; the training-loss margin in fact widens at 4B. Model-X's 55B / 147B-token runs supply the longer-horizon evidence, and a longer-budget Qwen 4B run remains Open due to limited resource.

Theory and formula corrections. [Revised]

  • Eq. 8 now carries the descent sign Δ⋆=−Uh(Σ/λ)V⊤\Delta^\star=-Uh(\Sigma/\lambda)V^\top.
  • App. F.2 is corrected to λ≈RMS(g)/2γ\lambda\approx\mathrm{RMS}(g)/\sqrt{2\gamma} (for R=γdR=\gamma d).
  • We dropped "absorbs into the learning-rate sweep": λ\lambda sits inside the nonlinear link, so we fix γ\gamma and tune the rate separately.
  • We now flag the vacuous-bound regime (noise floor above the trivial bound at the used γ\gamma) as a limitation of the stochastic guarantee.

Tokens on the x-axis and validation. [Revised]

Qwen3 and Model-X curves now use processed tokens (billions); GPT-2 stays in steps. Validation loss / PPL is now in the main text (Tables 5 and 6).

Model-X minor points. [Revised]

Table 12 now marks the untested low-rate Model-X entries as "na" (not searched) rather than "NaN" (diverged), with the caption updated.

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

This paper proposes MultiNormStep, a family of optimizers which arises from looking at steepest descent under an Orlicz-ball constraint which positions existing optimizers like signSGD, Muon, and AdamW. Under this constraint, the authors derive a general solution and convergence proof. Standard LLM training with Muon uses AdamW for embeddings, the output head, normalization parameters, and biases; in their experiments, the authors test two variants of their optimizer (a log and sigmoid update) across full parameter training on GPT-2 and Muon-hybrid training for Qwen3 and Model-X MoE training setups.

Critical strengths
  • The author’s mathematical framework provided by the Orlicz construction yields simple log and sigmoid updates. The scalar derivations and most of the fixed-scale convergence analysis are sound.
  • Experimental settings span several architectures and scales, with ablations on normalization granularity and smoothing. There are also empirical optimizer-state memory savings by pooling the second-moment statistic.
Critical weaknesses
  • Empirical improvements are small and further don’t report over multiple seeds; it is unclear how the results extrapolate. In Table 11, sigmoid improves over Muon+AdamW by 0.007 at 0.6B and 0.004 at 4B so validation loss seems to decrease with scale.
  • There is a gap between the proposed updates given by the mathematical model, where the authors acknowledge in Section 3.3 that efficient spectral log/sigmoid solvers are unavailable in the paper. The large-scale hybrid experiments retain Newton–Schulz Muon and are limited to coordinatewise maps on the remaining parameters.
  • The paper’s claim that steepest-descent methods only commit to fixed geometry updates doesn’t address some prior work that similarly looks at adaptive optimizer families (eg. Massena et al [1] looks at adaptive Schatten exponents).

As a general critique, the writing could be improved, particularly the exposition leading up to the author’s main results taking up the large majority of the main script. I would condense the abstract and introduction and provide better high-level descriptions and connections between the theoretical results and what is provided in the Appendix.

[1] Massena, Thomas, Corentin Friedrich, and Mathieu Serrurier. "From SGD to Muon: Adaptive Optimization via Schatten-p Norms." arXiv preprint arXiv:2605.19781 (2026).

Questions for authors
  • Could the authors address [1] and distinguish their Orlicz framework and non-Muon-branch contribution from existing work?
  • The Muon-Hybrid runs use a single learning rate, which is non-standard but is addressed with experiments in Appendix D.2. However, why is the experiment performed on 400 steps of an AdamW GPT-2 trajectory rather than the evaluated hybrid architectures using Qwen3 and Model-X? Could the authors repeat the diagnostic on Qwen3 0.6B Muon+AdamW trajectories? (i.e. fix Muon’s base rate at 0.002, sweep the companion LR over various multiplicative factors for AdamW, log, and sigmoid, with equal tuning budgets)?

[1] Massena, Thomas, Corentin Friedrich, and Mathieu Serrurier. "From SGD to Muon: Adaptive Optimization via Schatten-p Norms." arXiv preprint arXiv:2605.19781 (2026).

1 comment

Thank you for your review. The revision adds 3-seed validation at both Qwen3 scales, companion-branch controls on Qwen3 0.6B, held-out evaluation at both Model-X scales, and a corrected convergence treatment of the scale actually deployed. All table and equation numbers below refer to the revised paper.

Companion learning rate and validation. [Revised, Tables 5 and 9]

Table 9 now reports full 20k-step Qwen3 0.6B hybrid runs: AdamW companion multipliers of 0.5×0.5\times, 1.0×1.0\times, 2.0×2.0\times give 2.902 / 2.899 / 2.900 validation loss, so the shared rate is at AdamW's own optimum. Table 5 reports three-seed validation; both maps improve on AdamW in each paired run (sigmoid margin 0.008 nats at 0.6B, 0.005 at 4B). We read 0.008 and 0.005 as essentially close validation margin, not a decay with scale: the training-loss margin in fact widens at 4B (0.033 / 0.030 nats), so the geometry is doing more work as the model grows.

Relationship to adaptive Schatten optimization. [Revised, §2]

We corrected the overclaim you flagged: the introduction no longer says steepest-descent methods only commit to a fixed geometry. It now reads that nearly every such method fixes a single norm, and that even recent adaptive variants (Massena et al.) move only within a one-parameter family of homogeneous norms. We added Massena et al. to §2: they select a layerwise Schatten-pp exponent from runtime statistics, moving within that homogeneous family; our Orlicz modular instead expresses several non-homogeneous update rules through one template, so the two are complementary rather than overlapping.

Spectral scope and the companion branch.

§3.3 derives the spectral log and sigmoid maps. Evaluating them by explicit SVD adds a large per-step cost, and the Newton-Schulz iteration used by Muon targets the polar factor only, so applying a nontrivial function to singular values needs a separate approximation. The hybrid experiments therefore keep standard Muon on the hidden matrices and swap only the companion branch (embeddings, LM head, norms, biases; plus over-embedding parameters in Model-X). This branch holds large tensors even when the hidden matrices dominate the parameter count, so its update rule is consequential.

Writing, and the theory-to-appendix connection.

On the general exposition critique, we tightened the introduction (removed the repeated "two optimizers" explanation and a duplicated corner argument) so the setup reaches the main results faster. We also added the high-level signposting you asked for between the main-text theory and the appendix: §3.4 now states the main convergence result (Theorem 1, fixed scalar λ\lambda) in one place and points explicitly to App. C.2, where Proposition 1 carries it through to the online, group-wise scale λt,j\lambda_{t,j} that is actually deployed, and §3.3 points to the spectral construction. So a reader can follow the theoretical thread from the main text into the appendix without losing the thread of what is proved where.

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

MultiNormStep derives log and sigmoid update maps from an Orlicz-constrained steepest-descent framework. These maps achieve competitive full-parameter training performance and improve validation loss when replacing AdamW on the non-Muon parameter branch.

Critical strengths
  1. The Orlicz formulation provides an interpretable framework connecting coordinate-wise and spectral update maps.

  2. Targeting the non-Muon branch is practically useful, with experiments spanning dense and MoE models at several scales.

Critical weaknesses
  1. The hybrid comparison does not sufficiently separate geometry from update magnitude. Despite sharing a learning rate, Table 12 reports sigmoid and log update RMS values approximately 3.4× and 5.7× AdamW’s. The small validation improvements in Table 11 also lack uncertainty estimates.

  2. The convergence analysis covers a fixed scalar λ, while Algorithm 1 uses a gradient-dependent, group-wise EMA scale. Appendix C.2 leaves this adaptive setting unresolved, limiting the guarantee’s applicability to the deployed optimizer.

Questions for authors
  1. Do the hybrid gains persist after independently tuning each non-Muon branch’s learning rate while keeping the Muon branch fixed?

  2. Are the validation improvements in Table 11 reproducible across multiple seeds? Please report mean and variability for a representative setting.

  3. What additional conditions would extend the fixed-λ guarantee to Algorithm 1’s adaptive λ? Please distinguish proved results from heuristic extensions.

1 comment

Thank you for your review. The revision adds 3-seed validation at both Qwen3 scales, companion-branch controls on Qwen3 0.6B, held-out evaluation at both Model-X scales, and a corrected convergence treatment of the scale actually deployed. All table and equation numbers below refer to the revised paper.

Companion learning rate (effective step size). [Revised, Table 9]

We ran the full 20k-step Qwen3 0.6B hybrid with Muon's base LR fixed at 2×10−32\times10^{-3} and the AdamW companion LR set to 0.5×0.5\times, 1.0×1.0\times, and 2.0×2.0\times that rate. Validation losses are 2.902, 2.899, and 2.900. The shared base rate (mul = 1) is best among the three, and doubling the AdamW rate barely moves the result, so the single shared LR does not under-tune AdamW under the existing Muon scaling 0.2max⁡(din,dout)0.2\sqrt{\max(d_{\rm in},d_{\rm out})}. This sweep tunes AdamW locally; independent tuning of every companion method and matching their update RMS remain Open.

Reproducibility across seeds. [Revised, Table 5]

Table 5 now reports validation loss / PPL as mean ±\pm std over seeds {42, 1234, 2026} for the three branches of interest (AdamW, log, sigmoid). Validation losses are 2.903±0.003(AdamW)2.903 \pm 0.003 (AdamW), 2.894±0.0052.894 \pm 0.005 (log), 2.895±0.0042.895 \pm 0.004 (sigmoid) at 0.6B, and 2.765±0.0032.765 \pm 0.003, 2.763±0.0032.763 \pm 0.003, 2.760±0.0032.760 \pm 0.003 at 4B. Both maps beat AdamW in every paired seed at both scales. The advantage is therefore consistent rather than a seed artifact: at 4B the margin is tight but holds across all three seeds, and Table 6 corroborates it on held-out Model-X corpora where both maps again improve on AdamW.

Adaptive, time-varying scale. [Revised, Proposition 1 + Remark 1, App. C.2]

This is the main theoretical gap you flagged, and we have rewritten the analysis to close it. The main-text Convergence subsection now points to the appendix for this; the full treatment is in App. C.2.

  • The original theorem covers a fixed scalar λ\lambda. For a constant λ\lambda the implemented step −η0 sign(m) h(∣m∣/λ)-\eta_0\,\mathrm{sign}(m)\,h(|m|/\lambda) equals −(η0/λ)Pλ(m)-(\eta_0/\lambda)P_\lambda(m), so the scale folds into the effective rate. This absorption is valid only when λ\lambda is constant.
  • The deployed Algorithm 1 applies the online, group-wise λt,j=EMAβ2(g2)\lambda_{t,j}=\sqrt{EMA_{\beta_2}(g^2)} inside the link, i.e. the unscaled map QΛt(u)i=sign(ui)Q_{\Lambda_t}(u)_i=sign(u_i) h(|u_i| / \lambda_{t,j})$ . A time- and group-dependent factor cannot be absorbed into one global learning rate, which is exactly why a separate argument is needed.
  • New Proposition 1 analyzes QΛtQ_{\Lambda_t} directly under a two-sided envelope 0<a≤λt,j≤b<∞0<a\le\lambda_{t,j}\le b<\infty a.s. Then QΛtQ_{\Lambda_t} is 1/a1/a-Lipschitz, and using h(x)≥x/(1+x)h(x)\ge x/(1+x), St≥1S_t\ge1, λt,j≤b\lambda_{t,j}\le b, ∥g∥≤G\|g\|\le G, ⟨g,QΛt(Stg)⟩≥∥g∥2/(b+G)\langle g, Q_{\Lambda_t}(S_t g)\rangle \ge \|g\|^2/(b+G) pathwise. The descent proof then yields O(T−1/2)+O(σ)O(T^{-1/2})+O(\sigma) on the average squared gradient norm with a horizon-dependent step. The envelope is automatic for the EMA scale: bounded gradients cap it above (b≤Gb\le G) and the ϵ\epsilon regularizer supplies the floor aa (Remark 1).

The App. C.2 scope paragraph now states this explicitly, so the object analyzed and the object run are the same map.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your contributions to the review process. As the author–reviewer discussion period ends on November 18, I would like to kindly remind everyone to review the authors' rebuttals and engage in the discussion.

In particular, please indicate whether your main concerns have been adequately addressed and clarify any remaining issues that may affect your final assessment.

Your feedback will be very helpful for the subsequent reviewer–AC discussion and meta-review process.

Thank you for your time and effort!

Dear reviewers,

Thank you again for your time and helpful comments. With the discussion period closing tomorrow, we would appreciate any feedback on whether our rebuttal addresses your concerns. If any points remain unclear, we would be happy to clarify them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.