acceptodds
Under review as a conference paper at ICLR 2027

The Denoising Estimate Is a State Variable: Joint Refinement for Few-Step Flow Language Models

Abstract

Flow language models (FLMs) generate text by repeatedly predicting a clean representation with a denoiser and using it to transport a noisy latent state toward the data distribution. The first-order samplers commonly used in FLMs evaluate the denoiser once at the current state and keep the resulting estimate fixed throughout the update. This approximation becomes inaccurate for large steps, because changing the estimate also changes the transported trajectory and hence the state at which subsequent predictions are evaluated. We introduce , which refines the estimate jointly with the state it induces. Given a candidate estimate, AIM-FLM constructs the induced intermediate state through the closed-form affine transport of the flow, re-evaluates the denoiser at that state, and repeats this Picard iteration toward a joint fixed point. Under linear interpolation schedules this fixed point corresponds to an implicit-midpoint step, but it is solved directly in the denoising-estimate space, preserving a clean representation that can be decoded at the endpoint and retained as feedback across solver steps or distilled hops. Under a local contraction condition, the iteration converges geometrically, and under a margin condition, the decoded tokens stabilize after finitely many iterations without requiring exact convergence of the continuous estimate. The same coupling applies to pretrained deterministic and stochastic samplers without retraining, can be preserved in flow-map distillation by passing the estimate from one step to the next, and can be incorporated into joint training of the denoising field and hop model. Across several flow language model backbones, AIM-FLM improves the trade-off between generation quality and compute. Without retraining, AIM-FLM surpasses the 32-NFE ELF-B sampler in generative perplexity using 17 denoiser evaluations, with comparable entropy and lower repetition. With joint training, it surpasses the same baseline using only 12.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Continuous diffusion language models, fixed-point refinement, training-free sampling, post-training distillation training from scratch, self-conditioning, text generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MhJ6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Flow language models generate text by starting from noise and repeatedly moving a noisy latent toward clean data. At each step a denoiser predicts the clean representation x^\hat{x}, which is used twice: it is what gets decoded into tokens, and it also parameterizes the transport that decides where the next latent lands. Standard first-order samplers compute x^\hat{x} once at the start of a step and freeze it. The authors observe this is inconsistent for large steps: since x^\hat{x} also steers the step, the estimate used to move and the estimate that is correct where you land disagree.

AIM-FLM solves for an x^\hat{x} that is consistent with the trajectory it induces, via a Picard fixed-point iteration that re-evaluates the denoiser at the induced midpoint state. Under a linear schedule this fixed point is an implicit-midpoint ODE step, but it is solved in estimate space, so a decodable clean estimate is kept and passed forward. Each iteration is one denoiser call, making the iteration count a direct compute knob. The idea is applied training-free, inside hop distillation, and in joint training. The authors prove local geometric convergence under a contraction condition and finite-step token stabilization. On OpenWebText (1024-token unconditional generation, scored by GPT-2-Large generative perplexity, entropy, and repetition) they report a better quality-per-compute trade-off across several backbones, matching a 3232-NFE baseline with ∼12–17{\sim}12\text{--}17 denoiser calls.

Critical strengths
  1. The central reframing is correct and cleanly tested. The key point is that the denoising estimate x^\hat{x} is not just a readout but a state variable that steers the trajectory, so refining it at a frozen state (as self-conditioning does) is not the same as refining it along the trajectory it induces. Figure 4 tests exactly this: with backbone, schedule, and refinement budget held fixed, joint refinement beats fixed-state refinement, which rules out the trivial explanation that the gains come from extra compute.

  2. Simple to deploy, and the training-free gains are broad. The method needs no retraining, applies even to backbones without a self-conditioning channel, and shows consistent NFE reductions across several backbones and both deterministic and stochastic samplers, so the mechanism looks general.

Critical weaknesses
  1. The quality claim rests mainly on one proxy metric. The results lean almost entirely on generative perplexity under GPT-2 Large, which rewards fluent, high-likelihood text and does not capture factual accuracy or global coherence; entropy only guards against repetition collapse. The samples (Figs.6-8) are locally fluent and on-topic but show the usual unconditional-FLM drift and hallucinated specifics (invented names/dates). This is consistent with the behavior aggressive few-step sampling is most likely to worsen and a fluency proxy cannot detect it. A task with a verifiable answer (math, code, factual QA), where getting specifics right is required, or a distribution-level measure (MAUVE, held-out likelihood), would make the claim substantially more convincing. We note the backbones are unconditional, so this may require a conditional FLM, but even one such setting would help.

  2. The part of the theory that explains improved quality is narrower than the headline results. Prop. 3.1 (convergence) and Cor. 3.2 (token stabilization) hold generally, but the implicit-midpoint interpretation and O(h3)O(h^3) bound (Prop. A.1), the only result arguing the method is more accurate assume a linear interpolant, an arithmetic midpoint, and the deterministic substep, and the paper explicitly disclaims the re-noised sampler. Yet the main gains are on LangFlow (non-linear log-SNR midpoint) and stochastic SDE ELF samplers, which Prop. A.1 does not cover. So the accuracy argument motivates the method but does not formally explain the reported numbers.

Questions for authors
  1. Is Table 1 matched in total NFE, or in nominal steps? Your method uses multiple denoiser evaluations per step while the baselines use one, so please report the per-cell NFE to confirm the comparison is compute-matched.
  2. Can you report an empirical convergence check in the few-step regime, the contraction constant qiq_i, or ∥x^(k+1)−x^(k)∥\|\hat{x}^{(k+1)}-\hat{x}^{(k)}\| versus kk to show the fixed point is actually reached rather than the iteration merely helping for small kk (Prop. 3.1)?
  3. Implicit-midpoint accuracy result (Prop. A.1) assumes a linear interpolant, an arithmetic midpoint, and a deterministic substep, but the main gains use LangFlow's log-SNR midpoint and stochastic SDE ELF samplers. What, if anything, can be said about accuracy in these settings, and do you observe the same per-step benefit there?
  4. See weaknesses.
2 comments
AuthorsOpenReview ↗

Response to Reviewer MhJ6

We thank the reviewer for the feedback and address the main concerns below.

1. More metrics.\textbf{1. More metrics.} Beyond gPPL, AIM-FLM also achieves higher MAUVE and higher BLEU on conditional translation. See Table R2 and R4\textbf{See Table R2 and R4} in response to reviewer 3GoC.

2. Scope of the accuracy result.\textbf{2. Scope of the accuracy result.} The order argument holds beyond these assumptions. By variation of constants, any step that freezes x^\hat {x} has error

zi+1ODE−zi+1=∫wi(ν)[D⋆(z(ν),ν)−x^]dνz^{\mathrm{ODE}}_{i+1} - z_{i+1} = \int w_i(\nu) \left[ D^\star(z(\nu),\nu) - \hat x \right] d\nu

for any smooth coordinate ν\nu (e.g., log-SNR). Evaluating x^\hat x at the midpoint cancels the first-order term, giving O(h3)O(h^3) versus O(h2)O(h^2) for the pinned update. For ELF's SDE, this holds per step given the noise draws. Empirically, joint refinement outperforms fixed-state refinement on LangFlow (See below Table R5\textbf{See below Table R5}). We will add this to the Appendix in revision.

Table R5. gPPL / entropy on LangFlow with ω = 1.0 and 256 samples. Bold marks the lower gPPL.

Outer steps × kFixed-stateJoint (ours)
8 × 2228.19 / 5.711138.22 / 5.634
8 × 4148.61 / 5.63984.83 / 5.550
16 × 2100.74 / 5.56676.48 / 5.509
16 × 478.82 / 5.51257.99 / 5.444

3. Table 1 is compute-matched.\textbf{3. Table 1 is compute-matched.} Each step in Table.1 is a single hop of the distilled ELF-B student with one network evaluation, so the 1/2/4-step columns correspond to 1/2/4 NFE for all methods (NFE excludes the decoder call throughout, App.B). Wherever multiple evaluations per step are used (ModeA and hop iterations), each one is counted. We will list the NFE in Table 1.

4. Empirical convergence.\textbf{4. Empirical convergence.} Fig.4 shows that increasing kk consistently lowers gPPL. We further iterate the corrector ten times at every step of the 17-NFE ELF-B chain (14 fine steps, App.B) and report q^\hat q, the empirical contraction factor given by the average ratio between successive updates. At the entry and exit steps, successive updates shrink geometrically (q^<1\hat q < 1) and decoded tokens stabilize as predicted by Cor.~3.2. Guidance makes the middle steps near-neutral (q^≈1\hat q \approx 1). See Table R6 below\textbf{See Table R6 below}

Table R6. Empirical contraction factor and token stability along the 17-NFE ELF-B chain.

Fine stepsq̂Tokens changing (k = 0 → 9)
0–1 (entry)0.80–0.8740–72% → 6–13%
2–9 (middle)0.94–1.0730–40% → 10–26%
10–13 (exit)0.84–0.9114–26% → 1.6–6%

Dear Reviewer,

Thank you again for your detailed review. We have posted a point-by-point response to each of your comments, together with the additional results and the revised tables. We would be grateful to know whether our response has addressed your concerns, and we very much welcome any further questions or feedback. We are happy to clarify or run additional checks within the discussion period.

Best regards, The Authors

Reviewer_3GoCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors essentially propose a new sampler for flow language models based on the implicit midpoint method. At each step, the sampler moves the noisy state toward the denoiser’s prediction of the clean sequence. A standard sampler computes this prediction at the start of the step, whereas the proposed sampler computes it at the midpoint of the step. Since the midpoint state itself depends on the prediction, the two are found iteratively: an initial prediction is extrapolated from the previous steps, the midpoint state is computed from it, the denoiser is evaluated at that state, and its output becomes the new prediction. The iterations are repeated until the prediction stops changing; by default, a single iteration is used. The authors prove convergence of these iterations under a contraction condition.

The same principle is applied to distillation into few-step models and to joint training of the model and the student from scratch. Experiments are conducted on OpenWebText with three model families (ELF, LangFlow, and a one-hot FLM). Without retraining, the sampler matches the quality of the native samplers while using 1.7–8× fewer network evaluations. The distilled models improve generation quality at 1–4 steps compared with existing few-step methods.

Critical strengths
  1. Simple training-free sampler. The proposed sampler can be applied to pretrained flow language models without changing their weights. At matched entropy, it reaches the quality of their native samplers with 1.7–8× fewer network evaluations (Table 3).
  2. Theoretical grounding. The authors show that the converged update is an implicit-midpoint step, which is second-order accurate. They prove local convergence of the iterations under a contraction condition and show that the decoded tokens stabilize after finitely many iterations. The analysis also gives a clear interpretation of existing fixed-state refinement as an Euler step.
  3. Generality across models. The method is evaluated on three recent flow language model families that operate in different representation spaces: two embedding-based models (ELF and LangFlow) and a one-hot vocabulary-space model. The gains hold for both deterministic and stochastic samplers.
  4. The same idea extends to distillation, where the student also passes its prediction to the next hop and therefore remains refinable at inference time (Fig. 5), and to joint training of the model and the student from scratch.
Critical weaknesses
  1. No comparison with other second-order samplers. All training-free comparisons are against the native first-order samplers (Euler or ELF’s SDE sampler). With one iteration, the proposed sampler is essentially an explicit second-order predictor–corrector scheme, so a reduction in NFE relative to first-order samplers is expected.
  2. Entropy confound. Generative perplexity decreases with entropy, but several comparisons are made at different entropy: Fig. 3 at equal NFE, where the proposed sampler has lower entropy; Table 1 (5.12 vs. 5.41 for FMLM* at four steps); and Table 2. Fig. 4 does not report entropy at all. Part of the gPPL gain may therefore come from reduced diversity.
  3. Limited evaluation. Quality is measured only by generative perplexity under GPT-2 Large and unigram entropy, with no conditional generation, benchmarks or human evaluation. No wall-clock time is reported, and NFE does not account for differences in per-call cost.
Questions for authors
  1. Could you compare the training-free sampler with standard second-order solvers on the same backbones at equal NFE?
  2. Since unigram entropy measures only within-sample diversity, could you add an across-sample diversity metric (e.g., Self-BLEU) or MAUVE?
  3. Could you report wall-clock time per generated sequence for the proposed sampler, the distilled students, and the native samplers, including the cost of additional heads, adapters, and the decoder?
  4. Could you evaluate on conditional generation, e.g., prefix continuation, or on another task-based setting to support the quality claims beyond generative perplexity?
  5. The abstract claims that the sampler surpasses the 32-NFE ELF-B sampler with 17 NFE and lower repetition, but neither a 17-NFE result nor the Rep-4 metric appears in the tables (at 19 NFE, Table 3 reports 20.43 vs. 20.24). Could you clarify which results support this claim?

Post-rebuttal update. My main concerns regarding generation time and the comparison with second-order solvers are fully addressed: Table R1 shows that standard second-order solvers do not reproduce the gains of AIM-FLM, and Table R3 confirms the speed-up in wall-clock time.

I still have a concern about conditional generation. On WMT14 (Table R4), the comparison is made at equal step counts. At matched BLEU, the reduction in forward passes is only about 1.05–1.4× and decreases with the budget, compared with 1.7–2× in unconditional generation. I would ask the authors to discuss in the final version why the speed-up is smaller in conditional generation, e.g., whether this is due to the less pronounced discretization error, classifier-free guidance, or the noise scale not being tuned for this task.

I am raising my rating to weak acceptance.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3GoC

We thank the reviewer for the feedback and address the main concerns below.

1. AIM-FLM consistently outperforms high-order solvers.\textbf{1. AIM-FLM consistently outperforms high-order solvers.} At equal NFE, ModeA achieves the lowest gPPL on all three backbones against four second-order solvers. See Table R1\textbf{See Table R1} in our response to Reviewer zC4j.

2. AIM-FLM improves gPPL and MAUVE without reducing entropy.\textbf{2. AIM-FLM improves gPPL and MAUVE without reducing entropy.} For each official γ\gamma at 17 and 19 NFE, we report the gPPL, entropy and MAUVE. MAUVE uses GPT-2 Large features with 1000 samples against 1000 OpenWebText references. See Table R2 below\textbf{See Table R2 below}.

Table R2. gPPL, entropy, and MAUVE at equal NFE. Bold marks the better result.

NFEELF-B γELF-B gPPL ↓ELF-B EntropyELF-B MAUVE ↑AIM-FLM γAIM-FLM gPPL ↓AIM-FLM EntropyAIM-FLM MAUVE ↑
171.534.835.2220.1441.529.705.2230.238
2.029.225.1430.0872.024.855.1620.158
2.529.085.1130.0522.2523.035.1320.139
191.531.965.2090.1821.2531.615.2380.303
2.026.855.1330.1022.023.755.1490.170
2.526.715.1070.0642.2521.975.1190.135

3. Wall-clock time.\textbf{3. Wall-clock time.} We report ms/seq at batch 1 and 16, which include the decoder, and ms/call, which isolates the per-call cost of the extra heads and adapters. See Table R3\textbf{See Table R3} below.

Table R3: ConfigurationNFEms/seq (batch 1)ms/seq (batch 16)ms/callSpeed-up
Released ELF-B
Official sampler32371102.949.01.0×
Official sampler64731203.849.40.5×
AIM-FLM (Mode A)1720057.049.51.8×
AIM-FLM (Mode A)1922263.449.51.6×
AIM-FLM (Mode A)32371104.149.31.0×
Distilled student (Mode B)
1 hop1247.152.414.5×
4 hops45816.451.16.3×
4 hops × 2810229.051.33.5×
Jointly trained field (Mode C)
Official sampler32366103.549.31.0×
3 hops × 2 + tail812835.468.42.9×
6 hops × 21220054.368.01.9×

4. AIM-FLM improves conditional generation.\textbf{4. AIM-FLM improves conditional generation.} On WMT14 De$\rightarrow$En with the released conditional ELF-B and its official pipeline (3000 validation pairs, CFG of 2), AIM-FLM achieves higher BLEU at every step budget. See Table R4\textbf{See Table R4}.

Table R4. BLEU on WMT14 De→En. Bold marks the better result.

StepsELF-B ForwardsELF-B BLEU ↑AIM-FLM ForwardsAIM-FLM BLEU ↑
489626.699826.74
326426.286626.45
244825.695026.14
163225.293425.66
122424.832625.33

5. 17-NFE claim.\textbf{5. 17-NFE claim.} This claim is shown in Fig.3 of main manuscript\textbf{Fig.3 of main manuscript}. We apologize for not reporting the numbers in a table and will add the full results to the appendix. We show corresponding result below.

SamplerNFEgPPL ↓EntropyRep-4 (%) ↓
ELF-B official3223.545.1473.55
AIM-FLM (Mode A)1723.035.1322.78

Dear Reviewer,

Thank you again for your detailed review. We have posted a point-by-point response to each of your comments, together with the additional results and the revised tables. We would be grateful to know whether our response has addressed your concerns, and we very much welcome any further questions or feedback. We are happy to clarify or run additional checks within the discussion period.

Best regards, The Authors

Dear Reviewer 3GoC,

Thank you again for your constructive review. As the discussion period comes to an end, we would like to briefly summarize how we addressed your main concerns. high-order solver is addressed in Table R1, more metric issues is addressed in Table R2, Wall-clock time is addressed in Table R3. We also add additional condition generation task evaluation in Table R4 in our response above.

We hope these address your concerns, and we would be glad to clarify anything that remains unclear.

Best regards,

The Authors

Reviewer_zC4jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates few-step sampling in Flow Language Models (FLMs). The authors propose AIM-FLM, which jointly refines the denoising estimate and its induced intermediate state through Picard iteration to improve sampling consistency. Under linear interpolation, the method corresponds to implicit midpoint integration. The framework is further extended to flow-map distillation and joint training. Experiments on ELF, LangFlow, and FMLM demonstrate improved generation quality under reduced NFE budgets.

Critical strengths

This paper investigates the problem of few-step sampling in Flow Language Models (FLMs). The authors point out that traditional samplers fix the denoiser’s clean estimate at each sampling step, which may result in inaccurate state updates for larger step sizes. The authors propose AIM-FLM, which uses Picard iteration to jointly update the clean estimate and its induced intermediate state, ensuring consistency between the two. Under linear interpolation, this method corresponds to implicit midpoint integration. The authors further extend this idea to Flow-Map Distillation (Mode B) and Joint Field-and-Hop Training (Mode C), preserving feedback information between different sampling steps by simultaneously predicting the endpoint state and the outgoing clean estimate. Experiments covering models such as ELF, LangFlow and FMLM achieved a favourable quality–NFE trade-off on the unconditional text generation task using OpenWebText.

Critical weaknesses

The ‘Joint Refinement’ described in this paper corresponds to the classical Implicit Midpoint Method under Linear Interpolation. The authors also explicitly point out that performing a single Picard iteration starting from a first-order estimate can restore the Explicit Midpoint Update. However, the default configuration of Mode A typically performs only a single Picard iteration and utilises historical ‘Clean Estimates’ for an extrapolation warm start. This results in the method used in practice bearing a marked resemblance to existing high-order data-predictive solvers such as the Explicit Midpoint and DPM-Solver++ (2S/2M). Although self-conditioning, time parameterisation and re-noising may lead to differences in the specific formulae, there is currently a lack of direct comparisons to demonstrate that these differences confer independent performance benefits.

Questions for authors

Compared with classical high-order ODE solvers, does AIM-FLM possess any independent advantages?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zC4j

We thank the reviewer for the feedback and address the main concerns below.

1. Picard iteration is a compute knob.\textbf{1. Picard iteration is a compute knob.} As shown in Fig.~4 (main manuscript), increasing the iterations up to k=4k{=}4 consistently lowers gPPL across outer-step budgets.

2. AIM-FLM consistently outperforms high-order solvers.\textbf{2. AIM-FLM consistently outperforms high-order solvers.} At equal NFE, ModeA achieves the lowest gPPL on all three backbones against four second-order solvers (See Table R1 below\textbf{See Table~R1 below}). NFE excludes the decoder call. For LangFlow and FMLM, paired budgets differ by one NFE due to solver structure. ELF-B uses SDE sampling, and LangFlow and FMLM use ODE sampling.

Table R1. gPPL ↓ at equal NFE. Bold marks ours, italics mark the best baseline.

BackboneNFEEuler (native)HeunExp. midpointDPM++ 2SDPM++ 2MAIM-FLM (ours)
ELF-B1155.9750.1354.2059.5467.7733.07
1337.6543.5843.3849.2762.6128.25
1532.6236.7736.5836.7244.1724.82
1729.2232.1931.1232.0142.8023.03
1926.7128.5627.6929.1839.5721.97
LangFlow32/33122.17102.33101.52110.74104.5471.44
64/6580.3472.5071.9174.6273.4151.42
128/12960.0956.3556.0156.9756.6941.94
FMLM32/33237.53244.17251.90335.78228.02166.13
64/65144.68141.53166.32208.72138.9393.78
128/129101.8687.38115.69130.1198.3165.12

Dear Reviewer,

Thank you again for your detailed review. We have posted a point-by-point response to each of your comments, together with the additional results and the revised tables. We would be grateful to know whether our response has addressed your concerns, and we very much welcome any further questions or feedback. We are happy to clarify or run additional checks within the discussion period.

Best regards, The Authors

Thanks for your detailed response and experiment, which help address my main concern.

Dear Reviewer,

Thank you for your positive feedback. We are glad that our additional experiments and clarifications have helped address your main concern.

Thank you again for your time and constructive feedback.

Best regards,
The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response and Rebuttal Summary

We thank all reviewers for their constructive feedback. We are encouraged that they recognize AIM-FLM as simple, theoretically grounded, broadly applicable, and effective across different backbones and samplers (zC4j, 3GoC, MhJ6). They also highlight its training-free applicability (zC4j), and theoretical convergence and token-stabilization guarantees (3GoC, MhJ6).

1. Novelty

AIM-FLM’s novelty lies in treating the clean estimate not only as a readout, but as a variable that determines the sampling trajectory, and therefore jointly refining the estimate and its induced state. This principle differs from both fixed-state refinement and conventional higher-order solvers. More importantly, it provides a unified framework that transitions seamlessly from training-free sampling (Mode A) to flow-map distillation (Mode B) and joint field-and-hop training (Mode C).

2. Higher-order baselines (zC4j, 3GoC)

We added compute-matched comparisons with Heun, explicit midpoint, and DPM-Solver++ 2S/2M. AIM-FLM achieves the lowest gPPL at every reported NFE across ELF-B, LangFlow, and FMLM. See Table R1 in our response to zC4j.

3. Evaluation and efficiency (3GoC, MhJ6)

AIM-FLM improves both gPPL and MAUVE without reducing entropy (Table R2), and achieves higher BLEU at every tested budget on WMT14 De→En conditional generation (Table R4). Wall-clock results further show up to 1.8×, 14.5×, and 2.9× speedups for Modes A, B, and C, respectively (Table R3). Please refer to our response to 3GoC.

4. Theory and clarifications (3GoC, MhJ6)

We clarified that the midpoint accuracy argument extends to smooth coordinates such as log-SNR and conditionally to stochastic ELF steps. Additional experiments confirm that joint refinement outperforms fixed-state refinement on LangFlow (Table R5) and produces empirical contraction and token stabilization (Table R6). We also clarified that Table 1 is compute-matched and provided the exact result supporting the 17-NFE claim.

These results address the reviewers’ main concerns. We will incorporate them into the revised manuscript and welcome any remaining questions from the reviewers or the Area Chair.

Dear Reviewers,

The author-reviewer discussion phase is ending soon. If you have not done so yet, please read the author rebuttals and respond.

Please recall that, as a reviewer, you have agreed to engage in discussion with the authors to clarify and improve the paper. Failure to do so may result in penalties on your own submissions, as per the official ICLR policy. Given the high volume of submissions, your participation is crucial.

Thank you for your time and help.

Best regards,

Your AC

Dear Area Chair and Reviewers,

As the discussion period comes to an end, we sincerely thank all reviewers for their time and constructive feedback.

Our responses add compute-matched comparisons with second-order solvers, distribution-level and conditional evaluations, wall-clock measurements, and an extension of the accuracy argument with supporting experiments. We will incorporate all of these into the revised manuscript.

We would be happy to address any remaining questions before the discussion closes.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.