acceptodds
Under review as a conference paper at ICLR 2027

ZENO: Zone Expanding Decoding Operator for Diffusion Language Models

Abstract

Diffusion Large Language Models (dLLMs) generate text by unmasking tokens from anywhere on the generation canvas. This maximizes generation flexibility but minimizes context buildup by forcing models to commit to premature conclusions before establishing reasoning chains. This lack of directional grounding leads to poorly formed trajectories, high inference latency and rigid denoising schedules. We introduce ZENO (Zone Expanding decodiNg Operator), a training-free sampler that transitions from autoregressive grounding to parallel diffusion within a single generation instance at inference time. ZENO enforces early directional priors by restricting generation to a geometrically expanding prefix window, then shifts to a fixed block once sufficient context crystallizes. Evaluated on six standard benchmarks, across base and instruction-tuned checkpoints of LLaDA and Dream families, ZENO achieves up to a and speedup in per-query latency, with token throughput gains of up to and and cuts reverse diffusion steps by up to %. ZENO also improves generation quality, yielding a mean accuracy gain of up to points over vanilla decoding. Furthermore, enabling KV caching reduces per-query latency by up to with small accuracy trade-offs, establishing ZENO as a fast and efficient decoding strategy for dLLMs.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion, Inference Acceleration, Training Free Sampling, Masked Diffusion, Efficient Inference, Diffusion Language Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EVC3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ZENO is a training-free decoder for masked diffusion language models. It first allows generation only near the start of a response, then expands this region geometrically and can switch to fixed-size increments. Tokens are filled according to confidence. Tests on LLaDA and Dream show faster decoding, with the largest quality gains on instruction-tuned reasoning tasks. Gains on base models and knowledge tasks are less consistent.

Critical strengths
  1. ZENO offers useful speed and accuracy trade-offs against adaptive samplers. On Dream Instruct, ZENO(NB)+C achieves 60.2% mean accuracy at 2.8 seconds per query, compared with KLASS at 59.9% and 4.4 seconds.
  2. The evaluation covers varied tasks and reports both gains and regressions. The experiments cover four checkpoints and six benchmarks, with parameter sweeps, five-seed stochastic tests and paired correctness analyses against vanilla decoding. The paper also reports knowledge-task regressions and the accuracy cost of approximate KV caching.
Critical weaknesses
  1. The methodological novelty is incremental. Fast-dLLM already combines fixed-block decoding, confidence-based token selection, early block completion and optional KV reuse. ZENO's main addition is a geometrically expanding prefix, optionally followed by fixed increments. I view this as a scheduling extension of existing block decoders. The 'AR-to-diffusion paradigm switch' framing in Section 1 overstates this change.
  2. The experiments do not show how much this new schedule adds. In Tables 1 and 2, Block uses 256 model evaluations, while ZENO also changes token selection and stopping. Tables 9 and 10 vary ZENO's parameters, but do not compare fixed-size and expanding regions under the same decoding rules. A direct comparison with the complete Fast-dLLM decoder is also missing. Table 4 applies its KV-cache mechanism to ZENO rather than evaluating Fast-dLLM itself as a baseline.
Questions for authors
  1. Could you compare fixed-size blocks and expanding regions using the same confidence threshold, per-step token limit and stopping rules? One representative Instruct setting would be sufficient.
  2. How does ZENO compare with the complete Fast-dLLM decoder in accuracy and per-query decoding latency under the same evaluation protocol and comparable caching configurations?
4 comments

We thank the reviewer for the feedback. As suggested, we have conducted the recommended experiments and addressed each point below to strengthen the paper.

  • Point 1. We acknowledge that ZENO shares components with Fast-dLLM (confidence-based token selection, early completion and optional KV reuse), and we have revised Section 1 to present ZENO as a change in decoding schedule. Our controlled ablations show that the schedule is not a minor detail. With all other decoding rules held fixed, geometric expansion matches or exceeds the accuracy of fixed-increment (linear) expansion with 12.6–21.3% fewer NFE (Point 2), and the full ZENO(NB)+C decoder matches Fast-dLLM's accuracy (53.3 vs 52.8 on LLaDA-Instruct; 60.2 vs 60.1 on Dream-Instruct) with 13.8% and 28.4% fewer NFE. The schedule also changes how the model generates: with left-anchored expansion, the answer is committed after 96–97% of the content is written, compared with 21–41% for vanilla (Fig. 13), and most of ZENO's accuracy gains come from improved content rather than answer completion. We therefore see the contribution as showing that the order in which a diffusion LLM commits tokens is an effective, low-cost lever for reasoning quality and we have framed it in these terms in the updated manuscript. [Revised: Introduction, L076–084; Contribution 1, L086–089; Appendix B.1, L816–822]

  • Point 2 + Q1 (fixed-size vs. expanding regions). Unlike Block in Tables 1–2 (256 NFE), this ablation changes only the region schedule and keeps every decoding rule identical to ZENO(NB): confidence threshold τ\tau=0.8, per-step cap M=5, maximum generation length 256, and the same token selection, fallback, stopping and cleanup rules. With fixed-size causal blocks of 64 tokens, accuracy is 34.6 on LLaDA-Instruct and 26.2 on Dream-Instruct, compared with 54.2 and 61.6 for ZENO(NB) and stays essentially unchanged with M=64.

  • We also run a linear schedule that grows the window by a constant 8 tokens per phase (w.x+bw.x+b with w=8;b=8w = 8; b = 8, i.e. 16, 24, 32, ... tokens) until the generation limit is reached. On LLaDA Instruct, it scores 53.4 at 127 NFE, compared with 54.2 at 111 NFE for ZENO(NB). On Dream Instruct, it scores 61.2 at 138.5 NFE, compared with 61.6 at 109 NFE. ZENO matches or slightly exceeds its accuracy while using 12.6% and 21.3% fewer function evaluations, so the geometric schedule itself accounts for the efficiency gain. We have added this ablation to the revised version. [Revised: Table 12, Appendix F.9, L1226–1264]

  • Q2 (Fast-dLLM decoder evaluation). We evaluated the Fast-dLLM decoder as a standalone baseline, rather than only applying its KV cache to ZENO as in Table 4, under the same evaluation protocol on LLaDA-Instruct and Dream-Instruct (mean over six benchmarks), with τ\tau=0.8 and without KV caching for both methods. Fast-dLLM scores 52.8 at 65 NFE, compared with 53.3 at 56 NFE for ZENO(NB)+C on LLaDA-Instruct, and 60.1 at 86.6 NFE compared with 60.2 at 62.0 NFE on Dream-Instruct. ZENO therefore matches or slightly exceeds Fast-dLLM's accuracy with 13.8% and 28.4% fewer function evaluations. We report NFE as the hardware-independent cost measure for this comparison; wall-clock timing for the Fast-dLLM run will be added in the final version. We have added all of these comparisons in the revised version of the submitted manuscript. [Revised: Table 12, Appendix F.9, L1266–1269; Section 5.1, L291–293]

We thank you for your thorough review and insightful comments on our work. We hope these additions address your concerns and would be grateful if you would consider them in your final assessment. Thank You!

AuthorsOpenReview ↗

A Gentle Reminder to review our responses

Dear reviewer,

Thank you again for your time and feedback on our submission. As the discussion period progresses, we wanted to send a gentle reminder regarding the response we’ve submitted. We have made a lot of effort to address the questions, and really feel that the questions that were asked are really valuable.

Your feedback is highly appreciated. Kindly let us know if you have any further queries.

Thanking you, The Authors

Thank you for your rebuttal, as well as for conducting the additional experiments. Your responses have satisfactorily addressed my concerns, and I have updated my review score accordingly.

Dear Reviewer,

Thank you for your positive feedback, for taking the time to evaluate our new experimental results, and for updating your score. We truly appreciate your questions and suggestions, which have significantly helped us strengthen our paper.

Thanking you, The Authors

Reviewer_1acdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ZENO is a training-free decoding schedule for masked diffusion language models. It restricts early eligible positions to a short prefix zone, expands the zone geometrically, and optionally switches to fixed increments. The evaluated variants combine this schedule with confidence-based commitment, reveal caps, forced progress, finite per-phase budgets, and a final cleanup pass.

The paper evaluates LLaDA and Dream base/instruction checkpoints on six benchmarks and reports substantial instruction-model gains over vanilla full-canvas decoding. For example, LLaDA-Instruct GSM8K accuracy increases from 54.8% to 75.3%, with mean function evaluations falling from 256 to 56.9 for NB+C. Additional studies cover stochastic seeds, generation budgets, schedule parameters, approximate KV caching, and DiffusionGemma transfer. Paired correctness counts and significance analyses support substantial improvements over vanilla on instruction checkpoints. The main unresolved question is how much benefit is attributable specifically to zone expansion rather than its accompanying commitment policy.

Critical strengths
  1. Substantial gains with a simple training-free intervention. The instruction-model improvements in Tables 1-2 are large enough to make the phenomenon worth studying. A sampler that improves useful quality while reducing latency could be practically consequential without additional training or an auxiliary model.

  2. The empirical study includes meaningful robustness and failure analysis. Generation-budget, hyperparameter, and five-seed stochastic studies broaden the evidence. Appendix I includes paired outcomes, McNemar tests, bootstrap intervals, and failure examples; Appendix F.1 acknowledges approximate caching. These strengthen the case that the main instruction-model gains over vanilla are a real empirical phenomenon rather than a single favorable configuration.

Critical weaknesses
  1. The central scheduling contribution is not isolated from the commitment policy. Appendix H states that N/NB use threshold 0.8 and cap five, while NB+C uses the same threshold and cap 64. The fixed Block baseline instead uses 256 evaluations. Differences in commitment rules and phase budgets therefore accompany geometric expansion. A fixed-block or comparable non-geometric schedule with identical threshold, cap, fallback, cleanup, and timing rules is needed to establish the scheduler's incremental value. This is consequential because scheduling is the principal claimed contribution, and the existing baseline comparisons do not identify which component causes the gain.

  2. The grounding and largest efficiency claims need stronger operational evidence. Algorithm 1 permits incomplete zones to carry masks forward and finally fills every remaining mask without the threshold or cap; the frequency and contribution of these operations are not quantified. The interpretation of a fully settled grounding prefix is therefore not guaranteed by the evaluated algorithm. Also, Dream-Instruct's roughly 4.0x per-query speedup and 20.2x token-throughput ratio require output-length and aggregation context. Appendix I discusses missing answers, but per-task lengths, truncation/extraction failures, and cleanup contributions are needed to interpret quality at latency. Caching trade-offs should also be described proportionately: Table 4 includes 6.6- and 5.3-point mean accuracy losses for Dream-Instruct NB and NB+C, respectively.

Questions for authors
  1. Evidence needed for assessment: How much gain remains against a fixed-block or non-geometric schedule using exactly the same threshold, cap, progress fallback, phase-budget convention, cleanup rule, and timing implementation? This would identify the contribution of zone expansion.

  2. Clarification: How often do phases terminate with unresolved masks, how many tokens are written by the final cleanup pass, and do those cases account for a substantial part of the reported gains or errors? This would clarify what grounding the evaluated algorithm actually enforces.

  3. Clarification / supporting evidence: For the largest reported gains, what are the per-benchmark generated lengths, truncation and answer-extraction failure rates, and latency aggregation rules? Please distinguish improved content correctness from improved completion/formatting and make clear how these factors affect the throughput multiplier.

9 comments

We thank the reviewer for the feedback. As suggested, we have conducted the recommended experiments and addressed each point below to strengthen the paper.

  • Point 1 + Q1: For block diffusion (causal block size 64) with confidence threshold τ=0.8\tau=0.8, per-step cap M=5 and maximum generation length 256 (identical to ZENO(NB), including fallback, cleanup and timing rules), LLaDA-Instruct scores 34.6 at 63 NFE, compared with 54.2 at 111 NFE for ZENO(NB), a gap of 19.6 points. On Dream-Instruct, block scores 26.2 at 67.4 NFE, compared with 61.6 at 109 NFE for ZENO(NB), a gap of 35.4 points. With M=64, block scores 34.6 at 21 NFE on LLaDA-Instruct and 25.8 at 27.2 NFE on Dream-Instruct. Block's accuracy is essentially unchanged across a 2.5-3× range of NFE, so the gap is not due to compute budget. We added these results in the revised manuscript.[Revised: Table 12, Appendix F.9, L1226–1259]

  • As suggested, we also perform the linear scheduler experiment. The window grows linearly as w.x+bw.x+b with w=8;b=8w=8;b=8 (16, 24, 32, ... tokens) until the generation limit is reached. The other parameters kept exactly same as ZENO(NB) experiment (\tau=0.8, M = 5, Max generation limit=256 ). On LLaDA Instruct, the linear scheduler scores 53.4 at 127 NFE, compared with 54.2 at 111 NFE for ZENO(NB). On Dream Instruct, it scores 61.2 at 138.5 NFE, compared with 61.6 at 109 NFE. ZENO matches or slightly exceeds the linear schedule's accuracy while using 12.6% and 21.3% fewer function evaluations. Under identical commitment rules, these ablations isolate the scheduler's contribution. We have added both ablations in the revised version. [Revised: Table 12, Appendix F.9, L1261–1264; Appendix B.1, L816–822]

  • Point 2 + Q2: We instrumented the evaluated threshold-mode sampler. For ZENO(NB) (M = 5), at least one phase ends with unresolved masks on 12.7% (LLaDA) / 21.9% (Dream) of questions, with 2.8 / 4.8 carried masks per question on average; most are resolved within the next zone. The final cleanup pass commits content on 5.5% / 12.0% of questions ($\sim$1–2 tokens per question averaged over all), almost entirely on MMLU / MMLU-Pro. These questions have lower accuracy (~25% vs 57% / 66%), so the reported gains do not come from cleanup-committed tokens. To make the grounding interpretation hold by construction, we evaluated a strict-fill variant as well in which a zone cannot grow until it is fully unmasked, so cleanup never fires. Strict-fill achieves 54.8 vs 54.2 (LLaDA) and 63.2 vs 61.6 (Dream) for ZENO(NB), and 53.8 vs 53.3 (LLaDA) and 61.9 vs 60.2 (Dream) for ZENO(NB)+C, at equal or lower NFE. We do not adopt strict-fill as the main configuration, as it was introduced during the review as a diagnostic and has so far been evaluated only in the zero-shot Instruct setting, whereas all other experiments (Base models, KV caching, longer budgets, temperature sampling, DiffusionGemma and ablations) use standard ZENO; its gains support our claim that left-anchored grounding, not the cleanup mechanism, drives ZENO's improvements. We have reported these statistics and the strict-fill variant and the caching trade-offs in the revision. [Revised: Appendix H, L1404–1414; Table 12; caching: Section 5.3, L367–378, Appendix F.1, L1002–1018, Abstract L025–026]

  • Q3 and Clarifications: Aggregation rule- Tok/s is a ratio of sums (total generated tokens / total decode time, per benchmark); s/q is the mean wall-clock decode time per question at batch size 1, excluding tokenization, answer extraction and unit tests. Both are unweighted means over the six benchmarks. Detailed explanation is added in the revised manuscript. [Revised: Appendix C, L873–889]

  • On Dream-Instruct, vanilla mostly answers directly (27.4 vs 130.2 tokens for ZENO(NB)), so the 20.2× tok/s ratio of ZENO(NB)+C factors into a 3.96× per-query speedup and ~5× longer outputs; on LLaDA the length factor is only 1.4×. We therefore treat s/q as the primary efficiency measure. [Revised: Section 5.2, L357; Appendix I, L1484–1488]

  • In the revised version, Table 16 reports per-benchmark canvas use, no-answer and truncation rates, and accuracy on questions both methods answered. Most gains are content: on questions both answered, ZENO(NB) leads by +11.3 pp (Dream) and +5.3 pp (LLaDA) pooled, e.g. +36.0 of +39.0 on Dream GSM8K. LLaDA SVAMP is the main completion-driven case (vanilla gives no answer on 40%). The gain also holds under lenient extraction (+11.4 / +13.3 vs +12.3 / +12.8 strict), so it is not a formatting artifact. The Dream MMLU / MMLU-Pro losses (−5.1 / −2.4) come from truncated reasoning: ZENO(NB) leads by +0.4 / +2.2 on questions both answered and recovers at L = 512 (Table 6). [Revised: Table 16; Appendix I, L1457–1493]

Thank you for your thorough review and insightful comments on our work. We hope these additions address your concerns and would be grateful if you would consider them in your final assessment. Thank You!

AuthorsOpenReview ↗

A Gentle Reminder to review our responses

Dear reviewer,

Thank you again for your time and feedback on our submission. As the discussion period progresses, we wanted to send a gentle reminder regarding the response we’ve submitted. We have made a lot of effort to address the questions, and really feel that the questions that were asked are really valuable.

Your feedback is highly appreciated. Kindly let us know if you have any further queries.

Thanking you, The Authors

AuthorsOpenReview ↗

A Reminder and a Request to review our Responses

Dear Reviewer,

Thank you again for your detailed review. We carefully ran all the experiments and made the changes suggested. Because the discussion period is ending soon, here is a short summary of what we changed in response to each of your concerns.

  • We ran block diffusion under the same commitment rules as ZENO(NB): threshold 0.8, cap M = 5, length 256, and the same fallback, cleanup and timing. ZENO(NB) leads by 19.6 points on LLaDA-Instruct (54.2 vs 34.6) and 35.4 points on Dream-Instruct (61.6 vs 26.2). Block’s accuracy barely changes across a 2.5–3× range of NFE, so the gap is not due to compute budget (Table 12, Appendix F.9).
  • We also added a linear-growth schedule as a non-geometric baseline under identical settings. ZENO(NB) matches or slightly exceeds it (54.2 vs 53.4 on LLaDA-Instruct, 61.6 vs 61.2 on Dream-Instruct) with 12.6% and 21.3% fewer NFEs (Table 12, Appendix B.1).
  • We instrumented the sampler. Phases end with unresolved masks on 12.7% (LLaDA) / 21.9% (Dream) of questions, and most of these masks are resolved within the next zone. Cleanup commits tokens on 5.5% / 12.0% of questions, about 1–2 tokens per question, almost entirely on MMLU / MMLU-Pro. Those questions have lower accuracy, so the reported gains do not come from cleanup.
  • We also evaluated a strict-fill variant in which a zone cannot grow until it is fully unmasked, so cleanup never fires. At equal or lower NFE, it scores 54.8 vs 54.2 (LLaDA) and 63.2 vs 61.6 (Dream) against ZENO(NB), and 53.8 vs 53.3 (LLaDA) and 61.9 vs 60.2 (Dream) against ZENO(NB)+C. This supports the claim that left-anchored grounding drives the gains (Appendix H, Table 12).
  • The caching trade-offs are now described in proportion to their size in Section 5.3, Appendix F.1 and the abstract. We clarified the aggregation rules for tok/s and s/q (Appendix C). The 20.2× tok/s on Dream-Instruct breaks down into a 3.96× per-query speedup and outputs about 5× longer. We now treat s/q as the primary efficiency measure (Section 5.2).
  • Table 16 now reports per-benchmark canvas use, no-answer and truncation rates, and accuracy on questions both methods answer. On jointly answered questions, ZENO(NB) leads by +11.3 pp (Dream) and +5.3 pp (LLaDA), so most gains are in content. The gains also hold under lenient extraction, so they are not a formatting artifact. The Dream MMLU / MMLU-Pro drops come from truncated reasoning and recover at L = 512 (Appendix I).

We have also explained all the results in detail and updated the revised version of the paper as well. We have made a lot of effort to address the questions, and really do feel that the questions that were asked are really valuable.

Your feedback is highly appreciated. Kindly let us know if you have any further queries.

Thanking you, The Authors

Reviewer_XtM8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ZENO is a training-free schedule for masked diffusion language models. It unmasks inside a zone anchored at the start of the answer, a few tokens wide at first, widened by a growth base, then advanced in fixed blocks after a switch width (Section 3.2). ZENO(NB)+C also commits every eligible token whose confidence clears a threshold, and a later variant adds an approximate prefix cache. On LLaDA-8B and Dream-7B, base and instruct, the abstract reports up to 4.3× and 4.0× lower per-query latency, up to 6.2× and 20.2× token throughput, up to 78% fewer reverse steps, and up to 12.8 points of mean accuracy over vanilla decoding. The accuracy gains are concentrated on the instruction-tuned checkpoints.

Critical strengths
  • The default schedule is small: initial width 4, growth base 2, switch width 128, post-switch block 64, confidence 0.8. Algorithm 1 is the sampler.
  • The ablations cover the growth base, the post-switch block length, the confidence threshold, generation length and step budget, few-shot prompting, and five seeds at temperature 0.7 (Table 13).
  • A paired per-question McNemar test and a paired bootstrap give +12.5 points for ZENO(NB) over vanilla, interval [11.1, 14.0], and +11.4 [9.9, 12.8] for ZENO(NB)+C, both with p < 10−4.
Critical weaknesses
  1. The accuracy gains are measured against vanilla and against baselines that often fail to produce an answer. On the 256/256 setting, vanilla GSM8K is 54.8 for LLaDA-Instruct and 40.8 for Dream-Instruct, while Block is 27.5 and 13.5 (Table 6), and Block leaves 240 of 400 LLaDA-Instruct GSM8K items, and 256 of 400 Dream-Instruct items, with no extractable answer (Table 15). Fast-dLLM is the closest published confidence-and-cache sampler and is not run as a baseline. The “up to 12.8 points” is Dream-Instruct ZENO(NB) at 61.6 against vanilla at 48.8 (Table 2).

  2. Consistent gains on base and instruct models are not what Table 2 shows. It is said ZENO gives consistent and substantial improvements on both variants, but base means move from 54.3 to 54.6 for LLaDA and stay at 61.9 for Dream under ZENO(NB), and ZENO(NB)+C is 54.3 and 61.7 (Table 2). The introduction already calls the base effects more variable. On Dream-Instruct, ZENO(NB) is 5.1 and 2.4 points below vanilla on MMLU and MMLU-Pro.

  3. The token-throughput headline is inflated by output length. The abstract’s 6.2× and 20.2× are larger than the per-query speedups of 4.32× and 3.96×. Tok/s counts tokens actually generated divided by decode time (Table 2 caption). Dream-Instruct vanilla is about 11.2 × 2.7 ≈ 30 tokens per query and ZENO(NB)+C about 2.8 × 54.5 ≈ 153; LLaDA-Instruct goes from about 12.0 × 13.6 ≈ 163 to 2.8 × 83.7 ≈ 234 (Table 2).

Questions for authors

N/A

7 comments

We thank the reviewer for the feedback. As suggested, we have conducted the recommended experiments and addressed each point below to strengthen the paper.

  • Point 1: We have checked for no-answer-generated cases for vanilla and ZENO(NB). For LLaDA Instruct model on Gsm8K with vanilla sampler out of 400 no answer generated cases are 21 with 219 correct and 160 incorrect; for ZENO(NB) sampler there were 9 instances where answer was not generated with 299 correct and 92 incorrect. For Dream Instruct model the vanilla sampler numbers are: No answer generated: 41; Correct:163; Incorrect: 196 and for ZENO(NB) sampler the numbers are - No answer generated: 6; Correct: 319; Incorrect: 75. Alongside these for the block diffusion case we ran further analysis and found that on the items where Block does answer, its accuracy is 68.8% on LLaDA-Instruct and 37.5% on Dream-Instruct, compared with 76.5% and 81.0% for ZENO(NB) on the items it answers. This analysis shows that Block's low scores are not only due to truncation, and ZENO(NB) remains ahead even on answered items. [Revised: Table 16, Appendix I, L1457–1493]

  • As suggested, we ran the Fast-dLLM baseline for comparison with ZENO sampler, with both methods run without KV cache. The average comparison score for Fast-dLLM vs ZENO(NB)+C is 52.8 vs 53.3 (at 65 NFE vs 56 NFE) on LLaDA-Instruct and 60.1 vs 60.2 (at 86.6 NFE vs 62.0 NFE) on Dream-Instruct. In this case also, ZENO matches the baseline's accuracy at 13.8% and 28.4% lower NFE. [Revised: Table 12, Appendix F.9, L1266–1269; Section 5.1, L291–293]

  • We clarified the abstract to read "12.8-point gain over the vanilla baseline for Dream Instruct." [Revised: Abstract, L022–024]

  • Point 2: This point helps us state ZENO's contribution more precisely. By "consistent and substantial improvements on both variants," we meant that ZENO improves both model types, in different ways. On instruct models, ZENO improves both accuracy and efficiency, with per-query speedups of 4.32× on LLaDA-Instruct and 3.96× on Dream-Instruct (ZENO(NB)+C). On base models, the improvement is in efficiency. Under ZENO(NB), LLaDA-Base moves from 54.3 to 54.6 and Dream-Base stays at 61.9, while generation cost is substantially lower (NFE: 256 to 95 and 106). We have revised the sentence in the paper to state this distinction explicitly. [Revised: Introduction, L079–084; Contribution 3, L096–099]

  • On Dream-Instruct, ZENO(NB) scores 56.2 on MMLU and 37.6 on MMLU-Pro, compared with 61.3 and 40 for vanilla, which are gaps of 5.1 and 2.4 points. Our analysis shows that this gap in ZENOs performance comes mainly from reasoning that does not finish, not from reasoning that goes wrong. ZENO's left-anchored decoding produces reasoning trace before the answer, including on knowledge-heavy questions (e.g., history, law, biology) and on these benchmarks, that reasoning sometimes does not complete within the 256-token canvas. As a result, ZENO(NB) produces no answer on 12.5% of MMLU and 12.4% of MMLU-Pro questions, compared with 2.2% and 1.7% for vanilla (Where in most cases there is direct answer without reasoning). On questions that both samplers answer, ZENO is ahead by 0.4 points on MMLU and 2.2 points on MMLU-Pro. The paired counts support this (60 vs 58 on MMLU; 46 vs 38 on MMLU-Pro). So when ZENO's reasoning finishes, it usually helps. The cost comes mostly from reasoning that runs out of budget and only occasionally from reasoning that leads away from a correct direct answer. We have added a dedicated paragraph making it explicit in the revised manuscript. [Revised: Section 5.1, L280–285; Table 16 and Appendix I, L1486–1490]

  • Point 3: We reported tok/s as a complementary measure of generation rate alongside per-query latency, but as the reviewer notes, it combines speed with output length. On Dream-Instruct, 20.2× ≈ 3.96 (per-query speedup) × 5.1 (longer outputs) for ZENO(NB)+C. On LLaDA-Instruct, 6.2× ≈ 4.32 × 1.44. To avoid any ambiguity, we have removed "with token throughput gains of up to 6.2× and 20.2×" from the abstract. [Revised: Abstract, L021–026; Section 5.2, L357; Appendix C, L873–889]

  • The longer outputs come from ZENO producing complete reasoning, not from padding. In the full-canvas setting, vanilla Dream mostly answers directly, averaging ~27.4 tokens per query against ~130.2 for ZENO(NB). On LLaDA, vanilla fixes its final answer when only 21–41% of the response has been generated, compared with 96–97% for ZENO (Fig. 13). Vanilla commits to an answer before most of its reasoning exists, while ZENO's left-anchored decoding writes the reasoning first and then answers. This is also the source of ZENO's gains in accuracy on instruct models. [Revised: Appendix I, L1484–1488; Table 16]

Thank you for your thorough review and insightful comments on our work. Your acknowledgment will be highly appreciated and inspired if our responses are satisfactory and lead you to reconsider the overall assessment of our work. Thank You!

AuthorsOpenReview ↗

A Gentle Reminder to review our responses

Dear reviewer,

Thank you again for your time and feedback on our submission. As the discussion period progresses, we wanted to send a gentle reminder regarding the response we’ve submitted. We have made a lot of effort to address the questions, and really feel that the questions that were asked are really valuable.

Your feedback is highly appreciated. Kindly let us know if you have any further queries.

Thanking you, The Authors

AuthorsOpenReview ↗

A Reminder and a Request to review our Responses

Dear Reviewer,

Thank you again for your detailed review. We carefully ran all the experiments and made the changes suggested. Because the discussion period is ending soon, here is a short summary of what we changed in response to each of your concerns.

  • We added no-answer, correct and incorrect counts for vanilla and ZENO(NB). On the items Block actually answers, its accuracy is 68.8% and 37.5%, against 76.5% and 81.0% for ZENO(NB), so Block's low scores are not only due to truncation (Table 16, Appendix I).
  • As you suggested, we added Fast-dLLM as a baseline. ZENO(NB)+C matches its accuracy (53.3 using ZENO(NB)+C vs 52.8 with Fast-dLLM on LLaDA-Instruct; and 60.2 vs 60.1 on Dream-Instruct) with 13.8% and 28.4% fewer NFEs (Table 12, Appendix F.9).
  • The abstract now states that the 12.8-point gain is over vanilla on Dream-Instruct.
  • We revised the Introduction and Contributions. They now say that ZENO improves both accuracy and efficiency on instruct models, while on base models the gain is in efficiency at roughly unchanged accuracy (NFE 256 --> 95/106).
  • The MMLU and MMLU-Pro drops on Dream-Instruct come mainly from reasoning that does not finish within the 256-token budget. On questions both samplers answer, the ZENO sampler is ahead. We added a paragraph on this in Section 5.1.
  • We removed the 6.2× and 20.2× token-throughput figures from the abstract, so the headline efficiency claim is now per-query latency. We also show that the longer outputs are complete reasoning, not padding (Appendix I, Fig. 13).

We have explained all the results in detail and updated the revised version of the paper as well. We have made a lot of effort to address the questions, and really do feel that the questions that were asked are really valuable.

Your feedback is highly appreciated. Kindly let us know if you have any further queries.

Thanking you,

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.