acceptodds
Under review as a conference paper at ICLR 2027

One Schedule Does Not Fit All: Adaptive Inference Schedules for Diffusion Language Models

Abstract

Diffusion language models commonly rely on a single, fixed denoising schedule across all workloads, even though the cheapest quality-preserving allocation of model evaluations varies substantially across tasks and checkpoints. We introduce Task-Calibrated Denoising Allocation (TCDA), a confidence-controlled calibration framework that selects the lowest-cost schedule from a structured candidate library using a paired calibration sample and a user-specified quality tolerance. TCDA compares candidate topologies, including early prefix decoding, state-reusing online continuation, reduced full-range resampling, and the canonical baseline, deploying the cheapest schedule certified by one-sided simultaneous confidence bounds and task-specific degeneration guardrails. Across three frozen ELF-B tasks, TCDA identifies three distinct operating regimes: early prefix stopping for XSum, state-reusing online continuation for WMT, and guarded full-range resampling for OpenWebText. At representative tolerances ( for XSum and WMT; 20% relative perplexity for OpenWebText), selected schedules reduce denoiser evaluations by to , yielding measured wall-clock speedups between and with held-out violation rates below . Under strict margins, the method safely defaults to the full schedule. These findings establish that optimal inference reduction is workload-dependent, providing a disciplined methodology for budget-constrained diffusion inference.

Primary area:
generative models
Keywords:
diffusion language models, adaptive inference, early stopping, confidence bounds, denoising schedules, efficiency, text generation, calibration, prefix decoding

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_RWoqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Diffusion language models typically default to 64 denoising steps during inference, resulting in significant wasted computation; however, no single short schedule is optimal for all tasks. The paper introduces TCDA, which utilizes approximately 128 calibration samples and a lower confidence bound test with family-wise error rate control to select the lowest-cost, quality-certified schedule from a candidate pool, falling back to the full schedule if standards are not met. Experiments demonstrate that summarization allows for early truncation (3× speedup), translation requires state reuse for continuation (2×), and unconditional generation is limited to coarser grids (1.3×), confirming that the optimal inference schedule is task-dependent.

Critical strengths
  1. By modeling scheduling compression as a non-inferiority test with family-wise error rate control—complemented by a safe fallback mechanism—the violation rate of the retained set is kept below 0.4% (compared to up to 32.6% for naive selection), a capability that is key to the method's practical deployment.
  2. Three distinct optimal scheduling topologies were identified across three tasks and prospectively validated on independently trained ELF-M models using a pre-registration protocol (achieving a 3.26× measured speedup), thereby ruling out inflation due to multiple testing.
Critical weaknesses

Weakness 1: All conclusions are based on a single model family (ELF-B/ELF-M) and four manually designed candidate topologies; there is a lack of evidence regarding their applicability to other diffusion-based language models or a broader scheduling space, limiting the external validity of the claim regarding "task dependency."

Weakness 2: The candidate pool is entirely pre-specified by humans and tied to specific task structures (e.g., the specific parameters of prefix19 and H32); TCDA merely selects from existing candidates rather than discovering new schedules, making the method's contribution appear limited when compared to existing calibration or multiple-testing tools (such as Westfall–Young or Holm).

Weakness 3: The writing, especially the plot is hard to read. also why put the SHA here (Appendix K)?

Questions for authors
  1. How are the specific parameters in the candidate pool (e.g., prefix19, H32, resamp_K24) determined? If the candidate pool were larger or more granular, would TCDA’s family-wise error rate control become overly conservative, thereby negating the speedup benefits?

  2. The method was validated solely on the ELF model family; does it hold equally true for other diffusion/flow-matching language models (e.g., those with different noise schedules or decoding heads)? Is there any preliminary evidence?

  3. The calibration set (n=128) is fixed and randomly sampled; if the input distribution at deployment differs from the calibration distribution, can the certified schedule maintain its non-inferiority? Has a recalibration trigger mechanism for distribution shift been considered?

  4. Guardrail thresholds (such as n-gram repetition and entropy for OpenWebText) were added after degradation was observed; does this constitute a form of post-hoc adjustment? How can it be ensured that the guardrails themselves do not introduce new multiple-testing issues?

2 comments

Thank you for your practical, deployment-focused review. Your questions directly inspired our new LLaDA experiments.. All updates are in the revised PDF (Appendix A lists all changes).

W1 & Q2 (Validation Beyond ELF on Other Diffusion Families): We validated TCDA across 2 model families, 3 model scales, and 7 datasets:

  1. Second Model Family, LLaDA-8B-Instruct (masked discrete diffusion with an absorbing-state noise process and unmasking head; 1,000 items/task; 21 schedules & samplers; 500 splits):
TaskQuality Tolerance (Calib. n)Certified ScheduleSplits CertifiedPass / Wall-Clock SpeedupFalse Certs
WMT14 De→Enε = 1.0 chrF (n = 256)Fast-dLLM (τ = 0.8)451 / 5008.0× / 7.5×0 / 500
GSM8Kε = 5.0 acc pts (n = 512)Fast-dLLM (τ = 0.8)398 / 5004.05× / 4.25×0 / 500
XSumε = 0.5 ROUGE-L (n = 128)Fast-dLLM (τ = 0.8)424 / 5001.7×0 / 500

(On LLaDA XSum, a static back-loaded 16-step schedule at 4× is acceptable on the full pool, so at n = 128 TCDA conservatively certifies Fast-dLLM instead.)

  1. Prospective ELF-M & ELF-L on OpenWebText (§5.3, App. G): Using deterministic quantile time grids (removing ELF’s random per-run grid variance; App. H) on 3 fresh seeds × 1,000 untouched samples, ELF-M certifies prefix48 (Δ = −0.002 nats) and ELF-L certifies H48 (Δ = −0.027), both achieving 1.33× wall-clock speedup even at a strict 5% PPL tolerance. (Comparisons against SharpEuler, Heun, and adaptive early exit are also added in §5.)

W2 & Q1 (Hand-Picked Candidates vs. Automated 127-Schedule Pool): Originally, prefix19 and H32 came from small ELF-B pilot runs. To test whether a large, fine-grained pool makes error control overly conservative, we generated 127 schedules automatically (prefix at every step, grids every 4 steps, and 33 continuations):

  • Speedup nearly doubles: At n = 128, ε = 1, certified speedup rises from 3.37× → 6.75× on XSum (prefix8 already matches the full 64-step schedule!) and 2.0× → 3.8× on WMT, with ≤ 3/500 false certifications across all settings.
  • Why TCDA works where standard Westfall–Young fails: Standard studentized bootstrap breaks on large discrete-metric pools when bootstrap samples have zero variance—causing 406/500 needless fallbacks on this 127-schedule XSum pool. Standardizing by the original sample’s standard error fixes this while capturing correlation across nearby schedules, and pairing it with joint guardrail testing (Proposition 1) gives a complete deployment guarantee.

Q3 (Distribution Shift & Online Recalibration Trigger):

  • Zero-Shot Transfer (ε = 1): Certified schedules stay non-inferior out-of-domain: XSum → CNN/DailyMail (prefix19: Δ = −0.17 [−0.36, +0.01] ROUGE-L); WMT14 → newstest2016 (H32: +0.09 [−0.03, +0.21] chrF) and → IWSLT17 (+0.04 [−0.12, +0.22] chrF). Recalibrating directly on shifted data (n = 128) certifies prefix19 on CNN/DM (379/500 splits; 491/500 at n = 256; shifting safely to prefix26 at ε = 0.5) and H32 on newstest2016 (491/500) and IWSLT17 (401/500), with ≤ 1/500 false certs. (We now state the exchangeability assumption explicitly.)
  • Recalibration Trigger: Running the full schedule in the background on 10% of live traffic ("shadow monitoring") produces 0 false alarms on natural streams while flagging a −3.5 ROUGE-L quality drop after a median of 62 shadowed requests (96% detection; 79 requests at 100% shadowing; a label-free source-length KS test flags CNN/DM shift in 35 requests).

Q4 (Post-Hoc Guardrails & Multiple Testing): OWT guardrails were originally set after observing resamp_K16 degenerate. we have tested the following:

  1. Joint Certification (Proposition 1): Tests task quality (α = 0.04) and all guardrails (α = 0.01) together inside the family-wise test, keeping the asymptotic probability of deploying an infeasible schedule ≤ 0.05 with no uncounted multiplicity. On ELF-B (n = 128), OWT false certifications drop from 36/500 → 0/500 (463/500 match resamp_K24; XSum 500/500, WMT 486/500 matches, 0/500 false certs across one unified 500-split cohort).
  2. Prospective OWT Study: Froze all guardrail thresholds in writing before generation on untouched seeds and new scales (ELF-M, ELF-L), confirming every selection (§5.3, Added in App. G).

W3 (Readability, Plots & SHA Hashes): We regenerated all figures with ≥ 9 pt fonts to make it easier to view.

Dear Reviewer RWoq,

Thank you again for your review. The discussion period closes soon, so we wanted to check whether our response answered your questions and whether anything remains unclear.

In brief, the revised paper addresses each point:

  • Other model families (W1, Q2): on LLaDA-8B (masked discrete diffusion, with a different noise process and decoding head), TCDA certifies 8.0× fewer forward passes on WMT14, 4.05× on GSM8K and 1.7× on XSum, with 0/500 false certifications.
  • Hand-picked candidates (W2, Q1): with a mechanically generated 127-schedule library, error control does not become overly conservative: the certified speedup rises (XSum 3.37× → 6.75×, WMT 2.0× → 3.8×) with at most 3/500 false certifications.
  • Distribution shift (Q3): certified schedules stay within tolerance on CNN/DailyMail, newstest2016 and IWSLT17; we also evaluate recalibration and a shadow monitor as a trigger.
  • Post-hoc guardrails (Q4): guardrails are now certified jointly with utility (Proposition 1), which reduces OWT false certifications from 36/500 to 0/500; a prospective test froze every threshold before generation and confirmed every selection on fresh seeds.
  • Readability (W3): all figures regenerated with text of at least 9 pt; hashes moved to the supplement.

If any concern remains, we would be glad to address it before the discussion closes. If our response has resolved your concerns, we would appreciate it if you would consider updating your assessment.

Reviewer_bcohOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TCDA, an offline calibration framework that selects the cheapest inference schedule satisfying a prescribed quality tolerance. It compares predefined schedules using paired confidence bounds. Experiments on ELF models report 1.30–2.91× wall-clock speedups across summarization, translation, and unconditional generation.

Critical strengths

TCDA provides a simple, training-free procedure with explicit quality constraints and meaningful measured speedups.

Paired comparisons, conservative fallback, calibration-cost accounting, and prospective ELF-M validation strengthen the empirical evidence.

Critical weaknesses
  1. Insufficient engagement with existing sampling theory and unclear methodological novelty.

The related-work discussion overlooks substantial progress on diffusion sampling with approximate scores and discretization errors (Chen et al., 2023; Li et al., 2024a, 2024b, 2025; Gatmiry et al., 2026). More directly, Chen et al. (2024) already optimize continuous diffusion time-stepping using sampling-error bounds.

For discrete masked diffusion, Chen et al. (2026) provide an information-theoretic characterization of optimal parallel unmasking schedules, while Ben-Hamu et al. (2025) study entropy-controlled unmasking. Yao et al. (2026) additionally establish parallel-in-time acceleration for discrete CTMC sampling. These are mathematically different from ELF's continuous latent flow, but their implications and boundaries deserve discussion.

Crucially, TCDA neither optimizes a sampling-error bound nor learns a state-dependent schedule. It selects among handcrafted candidates using standard statistical tests. Its optimality is therefore only relative to the specified library and utility tolerance. The methodological advance beyond existing schedule heuristics and statistical model selection remains unclear.

  1. Narrow experimental scope and insufficient support for generality.

All principal experiments use ELF-B, while ELF-M validation covers only XSum. Each claimed scheduling regime is illustrated primarily by one task. It is therefore unclear whether the findings reflect general diffusion behavior or ELF-specific architectures, checkpoints, and evaluation metrics.

Furthermore, the highlighted OpenWebText setting permits 20% perplexity degradation, and Appendix H acknowledges post-hoc application of decisive degeneration guardrails. Section 6 also reports violation rates below 0.4% across all settings, whereas Appendix L includes 1/100 (1%) violations. These issues weaken the claimed reliability and warrant clarification.

Questions for authors
  1. How does TCDA compare with error-based time-step optimization (Chen et al., 2024) or higher-order sampling methods (Li et al., 2024b, 2025)? Can local discretization or denoising errors explain the selected schedules?

  2. Do these scheduling regimes persist beyond ELF, or across additional tasks and checkpoints? What evidence supports a general claim about diffusion language models?

  3. Could the authors reconcile the reported violation rates and explain the OpenWebText reference-match discrepancy between Table 3 (90%) and Appendix L (0.06)? How do post-hoc guardrails affect the certification claims?

  4. Could the collapse-and-refine perspective (Huang et al., 2026) explain why different noise stages require different computational budgets? Similarly, DPRM (Bu et al., 2026) uses terminal reward to determine token order in discrete diffusion. Could an analogous reward-to-go criterion guide ELF's continuous integration schedule?

References

Ben-Hamu et al. (2025). Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking.

Bu et al. (2026). DPRM: A Plug-in Doob h transform-induced Token-Ordering Module for Diffusion Language Models.

Chen et al. (2023). Sampling Is as Easy as Learning the Score: Theory for Diffusion Models with Minimal Data Assumptions.

Chen et al. (2024). Adaptive Time-Stepping Schedules for Diffusion Models.

Chen et al. (2026). Optimal Inference Schedules for Masked Diffusion Models.

Gatmiry et al. (2026). High-Accuracy and Dimension-Free Sampling with Diffusions.

Huang et al. (2026). Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine.

Li et al. (2024a). Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models.

Li et al. (2024b). Accelerating Convergence of Score-Based Diffusion Models, Provably.

Li et al. (2025). Faster Diffusion Models via Higher-Order Approximation.

Yao et al. (2026). Accelerating Discrete Diffusion Models with Parallel-In-Time Sampling.

3 comments

Thank you for connecting our paper to the continuous and discrete sampling theory literature

W1 & Q1:

  • Can local discretization error explain the selected schedules? No: Measuring latent trajectories shows that at step 19 of 64, relative latent error is similar on XSum (0.20) and WMT (0.29). Yet 85% of XSum output tokens already match step 64 (50% never change after step 4!), whereas WMT decodes stay truncated until step 64—even at step 63 where latent error is just 2.4%. We add a Decoding-Margin Proposition showing that discrete output tokens lock in once remaining latent drift falls below the decoder’s top-1 vs. top-2 logit margin divided by its Lipschitz constant. Thus, terminal t = 1 sampling bounds (Chen et al., 2023, 2024; Li et al., 2024a,b, 2025; Gatmiry et al., 2026) are sufficient, not necessary, for discrete task quality.
  • Head-to-Head Baselines at Matched NFE (XSum 1,000; WMT 3,000 items):
    • XSum: Prefix decoding at 20 NFE (+0.46 ROUGE-L; prefix19 runs 3.28× faster) outperforms a SharpEuler-style error-optimized step grid (−0.57, 3.12×; flow analogue of Chen et al., 2024) and 2nd-order Heun (−1.6 on a uniform grid, −8.2 on the logit-normal grid, 3.11×; cf. Li et al., 2024b, 2025), because integrating to t = 1 wastes steps after tokens have already committed.
    • WMT: SharpEuler matches H32 at 32–48 NFE (−0.04 vs. +0.06 chrF) but drops to −0.75 at 16 NFE (vs. +0.22 for continuation). Per-example adaptive early exit works on XSum (+0.45) but fails badly on WMT (−5.7 to −45.6 chrF).
  • What TCDA Contributes: TCDA is not a competing sampler, it is a layer that can wrap any candidate library, including SharpEuler, Fast-dLLM, EB-Sampler (Ben-Hamu et al., 2025), or an automated 127-schedule grid (which boosts certified speedup from 3.37× → 6.75× on XSum and 2.0× → 3.8× on WMT). Methodologically, our revised selector certifies task quality and degeneration guardrails jointly (Proposition 1: probability of deploying an infeasible schedule ≤ 0.04 + 0.01 = 0.05, asymptotically) while standardizing by the original-sample SE to avoid zero-variance bootstrap failures on discrete metrics. We have expanded our related work no to discuss all the 11 papers you cited (including Chen et al., 2026; Yao et al., 2026).

W2 & Q2: We expanded to 2 model families, 3 scales, and 7 datasets:

  1. LLaDA-8B-Instruct (masked discrete diffusion; 1,000 items/task; 21 schedules & samplers; 500 splits):
TaskQuality Tolerance (Calib. n)Certified ScheduleSplits CertifiedPass / Wall-Clock SpeedupFalse Certs
WMT14 De→Enε = 1.0 chrF (n = 256)Fast-dLLM (τ = 0.8)451 / 5008.0× / 7.5×0 / 500
GSM8Kε = 5.0 acc pts (n = 512)Fast-dLLM (τ = 0.8)398 / 5004.05× / 4.25×0 / 500
XSumε = 0.5 ROUGE-L (n = 128)Fast-dLLM (τ = 0.8)424 / 5001.7×0 / 500

(On LLaDA XSum, a static back-loaded 16-step schedule at 4× is acceptable on the full pool, but at n = 128 TCDA conservatively certifies Fast-dLLM instead.)

  1. On OpenWebText, the certified topology shifts with scale (prefix48 on ELF-M vs. continuation H48 on ELF-L); certified schedules also transfer zero-shot to CNN/DailyMail, newstest2016, and IWSLT17. Accordingly, we drop the "three universal regimes" claim and show instead that safe acceleration depends jointly on task, model, and sampler—requiring per-deployment certification.

Q3 & 20% PPL Tolerance:

  • "≤ 0.4%" came from a 500-permutation summary whereas Appendix L’s 1/100 came from a 100-split grid; Table 3 (90%) vs. Appendix L (0.06) used guardrail-filtered (resamp_K24) vs. unfiltered (resamp_K16) references. All results now use one 500-split cohort and one guardrail-screened reference π*.
  • Testing guardrails jointly per split cuts ELF-B OWT false certifications from 36/500 to 0/500 (463/500 match resamp_K24, vs. 500/500 degenerate certifications without guardrails; XSum 500/500 and WMT 486/500 matches with 0/500 false certs).
  • Prospective OWT at 5% PPL Tolerance (§5.3, App. G): Freezing all guardrails in writing before generation on untouched seeds (3 seeds × 1,000 samples) and using deterministic quantile time grids (App. H), both ELF-M prefix48 (Δ = −0.002 nats, bound −0.005) and ELF-L H48 (Δ = −0.027, bound −0.030) pass even at the strictest 5% PPL tolerance (ε = 0.049 nats) with 1.33× speedup.

Q4: This perspective aligns directly with our data: strong conditioning (XSum) settles coarse "collapse" early (85% of tokens fixed by step 19; later steps slightly lower ROUGE-L by 0.42), whereas unconditional OWT changes 58% of tokens after step 24/32. A DPRM-style reward-to-go stopping rule is a promising candidate for TCDA to certify (noting train–serve skew in our pilot learned predictors); both are discussed in Related/Future Work.

Thank you for the rebuttal. My main concern remains the limited novelty. TCDA essentially applies standard statistical model selection and confidence bounds to a predefined library of inference schedules. The additional decoding-margin proposition follows a conventional stability argument, while the risk-control guarantees rely on existing statistical techniques. I do not see a substantially new insight into diffusion sampling errors, convergence, or optimal scheduling beyond the existing literature.

The additional experiments do not address this fundamental concern.

Thank you for the follow-up. We agree with the factual part of your assessment: TCDA's statistical machinery (bootstrap max-T, intersection–union testing) is established, and the decoding-margin proposition is a standard stability argument. We do not claim new risk-control theory or a new convergence result, and the revision says so (§3, §6).

Our claim is different: the paper reports findings about diffusion sampling that the theory-driven literature does not cover, and that decide what can be deployed.

  1. Terminal sampling error is not the binding quantity for discrete outputs: At similar latent error (relative residual 0.20 vs 0.29 at step 19), 85% of XSum output tokens are already final, while WMT outputs stay truncated even when the latent is within 2.4% of its endpoint. As a result, at matched NFE, stopping early beats both an error-optimized grid (the flow analogue of Chen et al., 2024) and a second-order solver on XSum (+0.46 vs −0.57 and −1.6 ROUGE-L at 20 NFE) and is catastrophic on WMT (−27 to −49 chrF). Schedules that minimize a terminal error bound integrate to t = 1 and cannot express this trade-off.
  2. The safe schedule is not a property of the model alone: On LLaDA, early commit at 2× is free on WMT14 (−0.005 chrF) and costs 51.9 points on GSM8K; on OWT the certified topology flips between ELF-M (prefix) and ELF-L (continuation). Results that characterize optimal schedules for a model class (e.g., Chen et al., 2026) do not tell a deployment which sampler is safe for its task at a given tolerance; our data show that this answer changes across tasks for the same model.
  3. The standard tools fail when applied as-is: The textbook studentized max-T falls back in 406/500 splits with a 127-schedule library and never deploys Fast-dLLM on LLaDA WMT14 (TCDA deploys it at 7.5× wall-clock, 0/500 false certifications). Utility-only certification deploys a looping OWT schedule in 500/500 splits; joint certification in 0/500. The fixes are small, but they decide whether the rule is usable.
  4. A reproducibility finding: ELF's reference sampler draws its time grid at random once per run, which moves OWT policy effects by up to 32 standard errors; schedule comparisons on ELF are not reproducible unless the grid is fixed.

We understand if you weigh new theory above these results. We ask only that the paper be judged on our claims: an empirical and methodological study of when cheaper diffusion samplers are safe to deploy, not a new sampling theory.

Reviewer_nc4rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Task-Calibrated Denoising Allocation (TCDA), an offline method for selecting a cheaper inference schedule for a DLLM. The authors define four kinds of schedules, including prefix stopping, state-reusing hybrid continuation, resampled full-range schedules, and uniform reduced schedules. Candidate and canonical schedules are evaluated using paired examples, shared seeds, and shared initial latents. The paper then applies simultaneous one-sided confidence bounds and chooses the cheapest certified candidate, with fallback to the full schedule when no candidate passes The performance of the proposed method is evaluated on Xsum/WMT14 De-En/OpenWebText, showing the efficiency of the proposed method.

Critical strengths
  1. The decision problem is clearly formulated. The method optimizes inference cost subject to a user-specified quality tolerance and has an explicit full-schedule fallback.

2 The task-dependent finding is interesting. XSum favors early prefix stopping, WMT requires endpoint-reaching continuation, and OpenWebText favors a reduced full-range schedule. This directly supports the authors' main claim that one schedule does not fit every workload.

  1. The prospective ELF-M experiment is comparatively strong. It uses a preregistered candidate, untouched examples, fresh seeds, frozen guardrails, and hierarchical resampling. Generally, this shows clear research discipline.
Critical weaknesses
  1. The research scope is narrow. The primary study covers three tasks from one ELF-B model family, while the prospective transfer validates only one XSum candidate on one new model. This supports transfer of prefix19 on XSum, not broad transfer across tasks, architectures, or scales. Also, the experiment need to be adopted to more dataset to make the results more convincing.

  2. Several percentages are incompatible with the stated denominator. The paper says the Fixed-n evaluation uses 100 splits per setting, but reports XSum and WMT reference-match rates of 95.4% and 92.8%. Appendix Table 7 instead reports 96% and 93%. The counts and weighting must be clarified.

  3. Lack of comparison with other method. The main experimental results only compare the proposed method with baseline, which is not convincing.

Questions for authors
  1. Which OpenWebText target is the true “reference policy”: resampK16 or the guardrail-corrected resampK24? Please recompute all reference-match results under one definition. Can the OpenWebText result be repeated on untouched examples and seeds with all degeneration guardrails frozen in advance?

  2. What is the exact mathematical definition of the Endpoint Dependence Index? How are its uncertainty intervals computed?

See also the weakness part.

3 comments

Thank you for your careful review.

W1 We expanded from 1 model family and 3 tasks to 2 model families, 3 model scales, and 7 datasets:

  1. Second Model Family : LLaDA-8B-Instruct (masked discrete diffusion; 1,000 examples/task; 21 schedules & samplers; 500 calibration splits):
TaskQuality Tolerance (Calib. n)Certified ScheduleSplits CertifiedPass / Wall-Clock SpeedupFalse Certs
WMT14 De→Enε = 1.0 chrF (n = 256)Fast-dLLM (τ = 0.8)451 / 5008.0× / 7.5×0 / 500
GSM8Kε = 5.0 acc pts (n = 512)Fast-dLLM (τ = 0.8)398 / 5004.05× / 4.25×0 / 500
XSumε = 0.5 ROUGE-L (n = 128)Fast-dLLM (τ = 0.8)424 / 5001.7×0 / 500

(On LLaDA XSum, the cheapest acceptable schedule on the full pool is a static back-loaded 16-step schedule at 4×, but at n = 128 TCDA conservatively certifies the easier-to-verify Fast-dLLM instead. Also, 2× early commit is near-lossless on WMT14 [−0.005 chrF] but loses 51.9 points on GSM8K. now it aligns with our claim task-, model-, and sampler-dependence)

  1. Added ELF-M & ELF-L on OpenWebText, 3 domain-shift datasets (CNN/DailyMail, newstest2016, IWSLT17, they all stay within ε = 1), and an automated 127-schedule pool (raising certified speedup from 3.37× → 6.75× on XSum and 2.0× → 3.8× on WMT, with ≤ 3/500 false certifications).

W2 & Q1

  • You are right, 95.4% and 92.8% came from a 500-permutation run, whereas Appendix Table 7 (96%, 93%) used 100 splits. Likewise, OWT previously used guardrail-filtered resamp_K24 in Table 3 (90%) but unfiltered resamp_K16 in Fig. 4 (0.06). Every table and figure now uses one 500-split cohort (reported as exact counts) and one reference policy: the cheapest schedule passing both the quality tolerance and degeneration guardrails on the full pool (resamp_K24 on ELF-B OWT).
  • Previously, guardrails filtered the reference policy offline but were not tested inside each calibration split (causing 36/500 false OWT certifications of resamp_K16). Our revised selector tests task quality (α_u = 0.04) and guardrails (α_g = 0.01) together on every split (Proposition 1: P(deploying an infeasible schedule) ≤ α_u + α_g = 0.05, asymptotically):
ELF-B Setting (n = 128)Old: Ref. MatchOld: False CertsRevised: Ref. MatchRevised: False Certs
XSum (ε = 1)485 / 5000 / 500500 / 5000 / 500
WMT14 (ε = 1)472 / 5000 / 500486 / 5000 / 500 (2/500 at ε = 0.5)
OWT (20% PPL, resamp_K24)433 / 50036 / 500463 / 5000 / 500
  • Prospective OWT with Frozen Guardrails (§5.3, App. G): We froze all candidate schedules, guardrails, and tolerances in writing before generation on untouched seeds. Fixing ELF’s random per-run time steps to a deterministic quantile grid (documented in App. H, with zero changes to the frozen guardrails), every selected policy was confirmed across 3 fresh seeds × 1,000 samples even at a 5% PPL tolerance (ε = 0.049 nats): ELF-M prefix48 (Δ = −0.002 nats, lower bound −0.005) and ELF-L H48 (Δ = −0.027, lower bound −0.030), each giving 1.33× speedup. (ELF-M XSum was also repeated with deterministic grids on 1,000 untouched items across 3 fresh seeds: +0.32 ROUGE-L, lower bound +0.22, all guardrails pass.)

W3

  • On LLaDA: Compared against Fast-dLLM, EB-Sampler, and Prophet. On 200 GSM8K items (full schedule: 79.5%), official Prophet reaches 51.5% at 1.4× (README default) or 73.0% at 1.22× (answer-prefix), behind TCDA-certified Fast-dLLM τ = 0.8 (75.5% at 4.26×) and EB-Sampler (78.0% at 3.49×). Default Fast-dLLM (τ = 0.9) gives 3.26× (GSM8K), 6.6× (WMT), and 1.5× (XSum); EB-Sampler (γ = 0.1) gives 1.9×, 4.1×, and 1.16×.
  • On ELF: Compared against a SharpEuler-style error-optimized grid, Heun, and per-example adaptive early exit. On XSum, prefix decoding at 20 NFE (+0.46 ROUGE-L; prefix19 runs 3.28× faster) beats SharpEuler (−0.57, 3.12×) and Heun (−1.6 on a uniform grid, −8.2 on the logit-normal grid). On WMT, adaptive early exit is unsafe (−5.7 to −45.6 chrF), showing why TCDA certification is needed.

Q2 Comparing an endpoint-reaching schedule against prefix stopping at the same step budget K (fixing our prior unequal-budget comparison) over paired examples with shared initial noise:

EDI(K)=1n∑i=1n[u(i)(end,K)−u(i)(prefix,K)]\text{EDI}(K) = \frac{1}{n} \sum_{i=1}^{n} \big[ u^{(i)}(\text{end}, K) - u^{(i)}(\text{prefix}, K) \big]

with 95% paired percentile bootstrap intervals (10,000 resamples). At matched budget K: XSum H32 − prefix32 = −0.21 [−0.34, −0.08] ROUGE-L; WMT +32.2 [31.6, 32.8] chrF; OWT resamp24 − prefix24 = +0.80 [0.78, 0.82] nats. Across K = 8 to 60, EDI(K) is negative at every K on XSum (−1.28 to −0.14) and positive at every K on WMT (+27 to +46).

Thank you for the rebuttal. My major concern remains the limited research scope and multiple confusing statements about the performances. The rebuttal material does not actually address my concern.

Thank you for reading our response. We address the two remaining concerns as concretely as we can.

Confusing performance statements: Every inconsistency you identified had one of two causes: two separate evaluations reported side by side (a 100-split grid and a 500-permutation analysis), and two definitions of the OWT reference policy. The revision uses one 500-split cohort reported as exact counts, one reference policy, and two cost measures with fixed names: NFE ratio (denoiser evaluations of the full schedule ÷ those of the deployed schedule) and wall-clock (measured). All headline results in one table (n = 128 unless noted; Δ = deployed − full schedule; Selected = splits in which TCDA deploys that schedule; False = splits whose deployed schedule violates the tolerance or a guardrail on the full pool):

ModelTask, toleranceDeployedΔSelectedFalseNFE ratioWall-clock
ELF-BXSum, 1 ROUGE-Lprefix19+0.42499/5000/5003.37×3.28×
ELF-BWMT14, 1 chrFH32+0.06466/5000/5002.00×1.94×
LLaDA-8BWMT14, 1 chrF (n=256)Fast-dLLM τ=0.8−0.21451/5000/5008.03×7.53×
LLaDA-8BGSM8K, 5 pts (n=512)Fast-dLLM τ=0.8−0.7398/5000/5004.05×4.25×
LLaDA-8BXSum, 0.5 ROUGE-LFast-dLLM τ=0.8−0.07424/5000/5001.70×1.70×
ELF-MOWT, 5% PPL, prospectiveprefix48−0.002 natsconfirmed on 3 fresh seeds–1.33×1.33×
ELF-LOWT, 5% PPL, prospectiveH48−0.027 natsconfirmed on 3 fresh seeds–1.33×1.33×
ELF-MXSum, 0.5 ROUGE-L, prospectiveprefix19+0.32confirmed on 3 fresh seeds–3.37×3.26×

Scope: The evidence covers two model families with different noise processes (the continuous latent flow ELF at three scales plus an independently trained ELF-M, and the masked discrete diffusion model LLaDA-8B), seven datasets spanning summarization, translation, math reasoning and unconditional generation, three shifted test sets, and two prospective tests on fresh seeds under protocols frozen before generation. We agree that no finite study covers every architecture; if a specific model family or task would change your assessment, please tell us which.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.