acceptodds
Under review as a conference paper at ICLR 2027

Commit Locally, Exit Globally: Coordinating Adaptive Sampling and Early Exit in Diffusion Language Models

Abstract

Diffusion language models predict tokens in parallel but still require many iterative forward passes. Adaptive sampling accelerates local token commitment, while recent global stopping methods seek to end generation early. Our trajectory analysis shows that short answers can stabilize early in decoding, whereas reasoning outputs often require more steps to settle. A fixed progress threshold therefore struggles to avoid both redundant computation on short-answer tasks and premature termination on reasoning tasks. We introduce C, a training-free framework that combines local decoding acceleration with adaptive global early exit. Confidence-Verified Early Exit (CVEE) determines when to end generation, skipping the remaining steps once its stopping criterion is met. It combines confidence and temporal agreement to adapt stopping to the observed revision history. Commit-Core-Then-Confirm (CCTC) accelerates token commitment while generation continues, reducing the decoding steps needed to complete each block. It extends confidence-threshold parallel decoding with a boundary-anchored core and one-step confirmation of deferred predictions, constraining additional commitments to mitigate premature freezing under unresolved context. With CVEE calibrated once on LLaDA, the same numerical settings transfer across all 12 zero-shot tasks and to Dream without recalibration. Under this configuration, C reduces decoding steps by 64–95% and achieves measured decoding-loop speedups of 2.6x–18.6x, with task performance close to full decoding. Component analyses show complementary contributions: early exit supplies most savings on short-answer tasks, while local acceleration contributes more on reasoning and code generation. These results demonstrate that C effectively combines local token commitment and global early exit to accelerate diffusion decoding while largely preserving task performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Adaptive Sampling, Early Exit, Efficient Inference

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_tv79OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C⁴, a training-free framework that accelerates diffusion language model decoding by combining local token commitment with adaptive global early exit. Its two components are Commit-Core-Then-Confirm (CCTC), which extends confidence-threshold parallel decoding with a boundary-anchored core and one-step confirmation of deferred predictions, and Confidence-Verified Early Exit (CVEE), which stops generation once the extracted answer is both confident and stable, with patience that grows with the number of observed revisions.

Critical strengths
  1. The paper gives a clear empirical motivation for revision-adaptive stopping.
  2. The method is robust under a single frozen configuration.
Critical weaknesses
  1. Limited technical novelty. CCTC is essentially an engineering extension of Fast-dLLM’s confidence-threshold scheme, simply layering on a positional heuristic (left boundary) and a one-step confirmation pass. Similarly, CVEE’s confidence + run-length check closely mirrors the stability mechanisms in SchED, LESS, and STDec. The only real departure is the patience term ⌈γ⋅changes⌉\lceil\gamma \cdot \text{changes}\rceil, but this feels purely heuristic and lacks any ablation against natural adaptive alternatives.

  2. Questionable attribution of gains & missing baseline. On reasoning and code tasks, the speedup is almost entirely driven by CCTC rather than the proposed CVEE joint gate (Table 3 reveals the gate rarely fires, using 250/256 steps on LLaDA GSM8K and 256/256 on Dream). In effect, C4C^4 degenerates into a minor Fast-dLLM variant on half the benchmarks. Crucially, Table 1 fails to benchmark uncached Fast-dLLM at the same threshold (τ=0.8\tau=0.8), making it impossible to isolate CCTC's actual contribution over Fast-dLLM. Furthermore, the confirmation branch seems marginal—shaving off barely over 1 step on LLaDA GSM8K (85.2 vs. 83.9).

Questions for authors

see weakness above

5 comments
Reviewer_tv79OpenReview ↗

respone to author

Thanks for the reply. Based on the current reply, I am satisfied.

One more question, I would like to know how the author views this paper's contribution to the DLM community.

AuthorsOpenReview ↗

Response to Reviewer tv79 (1/2): What C⁴ contributes

Thank you for the follow-up. In short, we see three contributions to the community. First, an empirical observation: the point at which the answer stabilizes varies widely across inputs, so the computation a DLM spends should follow the decoding trajectory rather than a fixed schedule. Second, a concrete decoder built on this observation, which coordinates what to commit and when to stop under one configuration, with measured gains against a strong baseline. Third, a broader point that our experiments raise: acceleration methods are not necessarily compatible, and compatibility deserves attention in both design and evaluation. This comment covers the first two; the next covers the third.

1. Empirical observation: when the answer stabilizes

We measure when the extracted answer stops changing along the decoding trajectory (Sec. 4.1). The decoding progress by which 90% of trajectories have stabilized is 0.04 on short-answer tasks and 0.96 on long-reasoning ones. The computation that can be saved therefore sits in different places: on short answers it lies after the answer has settled, where the whole generation can end early; reasoning and code settle late, so savings must come from inside the blocks. A single rule tied to decoding progress cannot fit both, which is why the stopping condition has to adapt to the revisions actually observed. We believe this observation, and the way of measuring it, can be used by later work independently of C⁴.

2. Design of C⁴

DLM decoding involves two decisions: which tokens can be committed at each step, and when the whole generation can end. C⁴ designs one rule for each and has them operate on the same decoding trajectory:

  • CCTC handles commitment: among positions above the confidence threshold, it immediately commits those near the already-resolved boundary (the core) and defers the others; a deferred position is committed at the next step only if, under the updated context, it still clears the threshold and still predicts the same token.
  • CVEE handles stopping: it looks at both the confidence of the answer and the history of its revisions to decide whether the whole generation can end.
  • What the two give together: a short answer stabilizes early, so the remaining computation is skipped; while the answer of a long reasoning chain is still changing, CVEE stays conservative and lets generation continue, and the blocks that do need to run are accelerated by CCTC.

The difference from Fast-dLLM is therefore more than an adjusted threshold rule: CCTC builds on Fast-dLLM's threshold rule so that it fits the inference pipeline people already use, and what is added is the coordination between what to commit and when to stop. Confirmation reuses the next forward pass that would be executed anyway, rather than adding a dedicated one; the whole method is training-free and needs no auxiliary model.

3. Experimental results

  • Against the official Fast-dLLM + DualCache (Table R2; same model, examples, grading code and GPU): C⁴ has higher accuracy and fewer forward passes on all six datasets, and total decoding time summed over examples is 20.5% lower. Per example, C⁴ is faster on four datasets; on the two code tasks Fast-dLLM is faster, but C⁴ is more accurate (HumanEval 41.5% vs. 35.4%; MBPP 35.2% vs. 32.0%).
  • Against each backbone's own full decoding (Table R3): total decoding time over the six evaluated datasets falls by 81.2% on LLaDA and 82.7% on Dream (MMLU uses a 1,000-example subset). Accuracy is within one point of full decoding in 9 of the 12 settings and 1.8 points higher on LLaDA HumanEval; the two code tasks on Dream involve a quality trade-off (HumanEval −3.7, MBPP −1.8).
  • In the matched-threshold comparison with CVEE disabled (Table R1), CCTC trades a few extra steps per example for higher accuracy, a trade-off between quality and computation.
  • The same main numerical settings, calibrated once, are used for all evaluated tasks and both backbones; answer extraction is still handled per output format. This matters for practical deployment.

The third contribution, the point about compatibility, follows in the next comment.

AuthorsOpenReview ↗

Response to Reviewer tv79 (2/2): Compatibility between acceleration methods

4. Current state of decoding-level acceleration

At the decoding level, acceleration for DLMs largely follows a few patterns: adaptive sampling decides which tokens to commit at each step, early exit decides when the whole generation ends, and speculative decoding drafts and then verifies. Most current work advances within one of these patterns and is evaluated mainly by its standalone results. Combinations have been tried, for example Prophet reports a combination with Fast-dLLM on GSM8K, and Fast-dLLM itself pairs parallel decoding with caching; but whether different samplers and stopping rules, once joined, keep their quality across tasks has rarely been examined systematically.

5. Practical deployment combines optimizations at several levels

In practice, inference has to be optimized at every level at once: beyond the decoding algorithm there are KV caching, attention acceleration, quantization and hardware-level considerations. Open-source inference engines such as vLLM and SGLang, built for real deployment, integrate optimizations from many levels rather than a single acceleration algorithm or operator. DLMs still lack mature open-source deployment infrastructure of the kind vLLM provides (Li et al., 2025), and a recent survey of DLM inference concludes that "efficiency gains in dLLM inference are rarely attributable to a single knob" (Gwak et al., 2026). How fast a method is on its own is therefore not the full answer; whether it can coexist with other components matters as much.

6. Compatibility as observed in our experiments

Methods are not necessarily compatible: adopting A may rule out B, or cost quality. We observe this in our experiments (Table 2, Tables R1 and R2b):

  • Some combinations lose substantial quality on some tasks: in our settings, Prophet attached to Fast-dLLM or LESS loses a large part of its accuracy on reasoning and code, and CVEE with a fixed setting attached to WINO also loses accuracy on reasoning and code, though not on short answers.
  • Other combinations work: CVEE at its default setting attached to Fast-dLLM and LESS keeps the sampler's accuracy on the tasks tested (within one problem of LESS on HumanEval) and removes most steps on short answers.
  • The difference lies in the decoding trajectory: different samplers produce different trajectories, and with them the reliability of the signal a stopping rule depends on. This involves the assumptions the two components make about the trajectory, for example whether committed tokens can still change; we do not claim to have identified the cause of every failure, but it shows that compatibility is not only a matter of tuning.
  • Levels interact as well: in the HumanEval setting of Table R1, adding an approximate KV cache lowers accuracy for every decoding rule we tested.

C⁴ in turn shows that a compatible design can be achieved: it integrates early exit into adaptive sampling successfully, with both operating on one trajectory (design and results in the previous comment).

7. Possible directions

As the design of individual methods matures, we think a natural next step is to treat compatibility as a primary objective of design and evaluation:

  • In design, stating the assumptions a method makes about the decoding trajectory, and favoring methods with few of them (training-free, no auxiliary model, no per-task tuning), so that others know what it can be combined with.
  • In evaluation, reporting quality and efficiency after combination in addition to standalone results, together with diagnostics that reveal why a combination fails.
  • In the longer run, co-designing several decoding accelerations with caching, attention and hardware-level optimization, toward inference engines for DLMs that maximize inference speed at every level. This is especially relevant to agentic systems, where many calls have short outputs and latency accumulates across steps, and short outputs are where combining early exit pays most.

Our paper studies the combination of two decoding accelerations and one caching scheme. We hope that C⁴, as a concrete set of decoding rules together with an empirical analysis of combinations, lets later work both use the method directly and draw on when token commitment should be accelerated and when the answer still needs more generation. We would be glad to hear your view.

References:

Li et al., A Survey on Diffusion Language Models, arXiv 2508.10875.

Gwak et al., Accelerating Masked Diffusion Large Language Models, arXiv 2607.12829.

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

The paper proposes C4, a training-free method to accelerate diffusion language model inference by reducing decoding steps within blocks and stopping generation early when the answer stabilizes. C4 combines Confidence-Verified Early Exit (CVEE), which uses prediction confidence and temporal agreement to determine when to stop, and Commit-Core-Then-Confirm (CCTC), which accelerates token commitment using confidence, position, and confirmation across steps.

The authors evaluate C4 on LLaDA and Dream across 12 zero-shot benchmarks. The authors report results that reduce decoding steps and achieve 2.6-18.6 times speedups while generally maintaining task-level accuracy.

Critical strengths

The paper presents a practical, training-free approach to combining local token commitment and global early stopping, which addresses key inefficiencies in dLLM decoding. It also presents analyses of the trade-offs between decoding speed and output quality, including when early stopping is effective, how the two mechanisms interact, and where they can fail, and these provide useful insights of the challenges and opportunities.

Critical weaknesses

The experimental setup is not well explained or justified. The paper uses only 100 examples per reasoning benchmark, zero-shot prompting, and fixed generation lengths without sufficient justification. The choices of answer extraction rules and hyperparameter calibration are also not adequately explained. These choices raise questions about the reliability and generalizability of the results.

The benefits of the proposed commitment mechanism against existing methods are not sufficiently shown. Although C4 is compared against Fast-dLLM, the experiments do not clearly show whether the CCTC provides a better accuracy-speed trade-off than standard confidence-based parallel decoding, independent of CVEE and differences in commitment thresholds.

Questions for authors

See weakness

6 comments
AuthorsOpenReview ↗

Follow-up to Reviewer CQUm

Dear Reviewer CQUm,

Thank you again for your review. As the discussion period closes on 18 Nov, we would like to check whether our responses have addressed your concerns.

To summarize what has been added since your review:

  • On the experimental setup (W1): six datasets are rerun at full test-set scale, against Fast-dLLM and Prophet on LLaDA (Table R2) and as a component ablation on both backbones (Table R3). The conclusions are unchanged.
  • On the benefit of CCTC (W2): Table R1 disables CVEE and compares CCTC with Fast-dLLM's threshold-only parallel decoding at the same threshold, with and without caching, which isolates CCTC from both CVEE and threshold differences as you requested.

You may also find our discussion with Reviewer tv79 relevant. That review raised a related question about what CCTC contributes beyond Fast-dLLM, and the thread, including our view of the paper's contribution to the DLM community, is visible in the forum. Reviewer tv79 has indicated that our responses addressed those concerns.

If anything remains unclear, or if there is a specific experiment you would still like to see, please let us know and we will respond promptly. We would be grateful if you could consider the responses and new results in your final assessment.

Thank you for your time.

AuthorsOpenReview ↗

Final follow-up to Reviewer CQUm

Dear Reviewer CQUm,

As the discussion period closes today, we would like to follow up once more on our responses.

In brief, on the experimental setup (W1), six datasets are rerun at full test-set scale in Tables R2 and R3, with unchanged conclusions. On the benefit of CCTC (W2), Table R1 compares CCTC with Fast-dLLM's threshold-only parallel decoding at the same threshold with CVEE disabled, as you requested. A related question on what CCTC adds beyond Fast-dLLM is discussed in the thread with Reviewer tv79, who has updated their assessment after that discussion.

If any concern remains, we are available today and will reply as quickly as we can. We would be grateful if you could take the responses and new results into account in your final assessment.

Thank you for your time.

Dear Reviewer CQUm,

We apologize for writing once more as the discussion period closes. We would like to share a brief update, as it may be relevant to your assessment.

Both of the other reviewers have now responded to our rebuttal. Reviewer tv79 wrote that they are satisfied with the reply, and Reviewer NEV3 wrote that their concerns are "largely resolved"; both have increased their scores. Several of the points they raised are related to yours, in particular what CCTC contributes beyond Fast-dLLM.

For your two concerns specifically: the six datasets are rerun at full test-set scale in Tables R2 and R3 (W1), and Table R1 compares CCTC with Fast-dLLM's threshold-only parallel decoding at the same threshold with CVEE disabled, as you requested (W2).

If any concern remains, we would be glad to address it right away. We would be grateful if you could take the responses and the new results into account in your final assessment.

With sincere appreciation and respect, Authors of Submission 2679

Reviewer_NEV3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a training-free method called C4, which reduces two types of computational overhead during the decoding process of diffusion language models. CCTC accelerates token commitment within a block by employing boundary-anchored cores and one-step confirmation, while CVEE determines when to terminate the generation of an entire sequence based on confidence in the extracted answers, sequential consistency, and modification history. The authors report significant speedups in the decoding loop across 12 tasks using LLaDA-8B and Dream-7B, with overall task performance remaining comparable to that of full-scale decoding.

Critical strengths
  1. The motivation is clear. The paper distinguishes between local block updates and global halting, demonstrating through trajectory analysis that short answers tend to stabilize earlier, whereas long-form reasoning and code generation rely more on accelerating the blocks that still require execution. This provides a direct rationale for a two-level design.
  2. The experimental analysis are relatively thorough. The authors evaluate two types of backbones across three categories of tasks and report actual inference speeds.
Critical weaknesses
  1. The incremental innovation over existing work is limited. Concepts such as high-confidence parallel submission, spatial and temporal stability, and global early exit have already been explored in prior research. For instance, Prophet has been evaluated in combination with Fast-dLLM, and STDec also leverages spatial and stride-based stability. While C4 employs a distinct set of rules, it has not been compared against these simpler combinations under identical experimental settings, making it difficult to gauge the specific marginal gains provided by its "coordination" design.
  2. CVEE’s design for answer consistency is rather limited. This approach is only suitable for tasks where answers are easily localized; furthermore, the fact that an extracted answer remains unchanged does not guarantee its correctness. In experiments, applying the same CVEE rule to WINO caused the accuracy on GSM8K to drop from 72% to 36%, further demonstrating that this stopping rule lacks robustness across different local decoding trajectories.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Follow-up to Reviewer NEV3

Dear Reviewer NEV3,

Thank you again for your review. As the discussion period closes on 18 Nov, we would like to check whether our responses have addressed your concerns.

To summarize what has been added since your review:

  • Prophet + Fast-dLLM, which you suggested as a baseline, is implemented and reported under identical settings in Table R1; the same test with LESS is in Table R2b.
  • A comparison against the official Fast-dLLM and Prophet on six full test sets is in Table R2.
  • Your points on the scope of CVEE and on the WINO experiment are answered in our replies (2/3) and (3/3).

You may also find our discussion with Reviewer tv79 relevant. That review raised related concerns about novelty with respect to prior work, and the thread, including our view of the paper's contribution to the DLM community, is visible in the forum. Reviewer tv79 has indicated that our responses addressed those concerns.

If anything remains unclear, or if there is a specific experiment you would still like to see, please let us know and we will respond promptly. We would be grateful if you could consider the responses and new results in your final assessment.

Thank you for your time.

AuthorsOpenReview ↗

Final follow-up to Reviewer NEV3

Dear Reviewer NEV3,

As the discussion period closes today, we would like to follow up once more on our responses.

In brief, the baseline you suggested (Prophet + Fast-dLLM) is reported in Table R1, the same test with LESS in Table R2b, and the comparison on six full test sets in Table R2. Our view of the paper's contribution, which relates to your concern about novelty, is set out in the thread with Reviewer tv79, who has updated their assessment after that discussion.

If any concern remains, we are available today and will reply as quickly as we can. We would be grateful if you could take the responses and new results into account in your final assessment.

Thank you for your time.

Thanks for the rebuttal and my concerns are largely resolved. I have increased the score.

Dear Reviewer NEV3,

Thank you very much for your kind response and for taking the time to reconsider our work. We are truly grateful that our rebuttal has addressed your concerns.

We sincerely appreciate your thoughtful feedback and your willingness to increase your rating. Your suggestion to evaluate Prophet together with Fast-dLLM led to one of the most informative additions to the paper, and we will include these results in the revised version.

Thank you again for your time, consideration, and support.

With sincere appreciation and respect, Authors of Submission 2679

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.