acceptodds
Under review as a conference paper at ICLR 2027

dVoting: Fast Voting for dLLMs

Abstract

Diffusion Large Language Models (dLLMs) represent a new paradigm beyond autoregressive modeling, offering competitive performance while naturally enabling a flexible decoding process. Specifically, dLLMs can generate tokens at arbitrary positions in parallel, endowing them with significant potential for parallel test-time scaling, which was previously constrained by severe inefficiency in autoregressive modeling. In this work, we introduce dVoting, a fast voting technique that boosts reasoning capability without training, with only an acceptable extra computational overhead. dVoting is motivated by the observation that, across multiple samples for the same prompt, token predictions remain largely consistent, whereas performance is determined by a small subset of tokens exhibiting cross-sample variability. Leveraging the arbitrary-position generation capability of dLLMs, dVoting performs iterative refinement by sampling, identifying uncertain tokens via consistency analysis, regenerating them through voting, and repeating this process until convergence. Extensive evaluations demonstrate that dVoting consistently improves performance across various benchmarks. It achieves gains of %-% on GSM8K, %-% on MATH500, %-% on ARC-C, and %-% on MMLU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
dLLMs, efficiency, reasoning

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_DU4zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes dVoting, a training-free test-time scaling method for diffusion language models. It preserves tokens that agree across sampled responses, regenerates the remaining positions using entropy-threshold parallel decoding, and aggregates candidate answers through majority voting with early stopping. Experiments on LLaDA-8B-Instruct, Dream-7B-Instruct, and LLaDA-1.5 demonstrate improvements over single-generation baselines and favorable accuracy-cost tradeoffs across mathematical reasoning and question-answering benchmarks. The principal contribution is cross-sample token reuse to reduce the cost of obtaining additional votes.

Critical strengths

dVoting offers a simple, training-free way to exploit arbitrary-position regeneration in diffusion language models, with practical benefits supported across multiple backbones and benchmarks. The end-to-end measurements are particularly useful: on GSM8K at generation length 256, it achieves 83.78% accuracy in 38.3 seconds, compared with 82.33% in 96.8 seconds for majority voting. The remasking ablations and evaluation on an RL-enhanced model further support its potential as an inexpensive inference-time improvement.

Critical weaknesses
  1. The comparisons do not sufficiently isolate the proposed contribution. dVoting uses adaptive entropy-threshold decoding, whereas the main voting baselines use fixed denoising budgets. Adaptive stopping and entropy-based acceleration already have precedents, making controlled attribution essential. Table 7 is helpful, but the MATH500 gain over confidence-based remasking is only 0.4 percentage points while requiring more steps, without uncertainty estimates. Matched-decoder and matched-budget comparisons are needed to establish the value of cross-sample consistency.
  2. The central reuse mechanism is insufficiently validated and specified. Fixing tokens after agreement between two samples makes later votes dependent on earlier outputs and may preserve shared errors. NUPR establishes repetition, but does not establish that freezing those tokens preserves useful exploration. The paper provides no systematic analysis of error persistence or diversity loss, while Algorithm 1 leaves competing repeated-token choices unspecified and does not clearly explain the illustrated stopping after two matching answers.
Questions for authors
  1. Under identical entropy thresholds, temperatures, block sizes, and hardware allocation, how does dVoting compare with independent sampling plus answer-based early stopping at matched latency? Please include repeated-run uncertainty estimates.
  2. Which token is retained when multiple token values each occur at least twice at the same position? Can preserved tokens subsequently change? Also, with n=5, Algorithm 1 requires three matching answers; what explains Appendix B’s stopping after two samples?
  3. How often does token reuse lock in an incorrect answer or prevent a correction achieved by independent sampling? Please quantify this alongside answer diversity, and clarify whether NUPR excludes padding and end-of-sequence tokens.
Reviewer_gV2AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a method called dVoting, which uses the remasking ability of dLLMs to avoid regenerating redundant content. For each sampling round, positions whose tokens have the same predictions as in previous samples are kept fixed. The remaining positions are remasked and regenreated. It also stops sampling early when a majority of the samplings cap agrees on an answer. The experiments on LLaDA-8B-Instruct, LLaDA-1.5 and Dream-7B across GSM8K, MATH500, ARC-C, GPQA and MMLU show that dVoting gains 3-15 points over the base models and has about 10x lower end-to-end latency than HEX.

Critical strengths
  1. Simple and training-free. This method needs no training and is easy to implement in different dLLMs.
  2. Well-motivated by analysis. The voting consistency analysis and the NUPR@k metric make the redundancy intuitive.
  3. Broad evaluation and comprehensive ablations. The main results cover three models and five benchmarks. And it has the ablations of the main hyperparameters to isolate the core idea. It also reports the real latency.
Critical weaknesses
  1. No compute-matched comparison with sequential scaling. For the base model, a natural way to improve is sequential scaling, like a single generation with a longer budget. However, the base model is only evaluated at a fixed length L, while dVoting could draw up to n samples fo length L.
  2. The isolated contribution is modest. The improvement over early stopping is small. In Table 7, confidence-based remasking uses fewer steps on MATH500 and is only 0.4 points lower in accuracy.
  3. The consistency is measured by absolute position. The agreement is computed position by position. Two samples that differ in a single inserted or deleted token are misaligned, which is fragile for longer cot outputs.
Questions for authors
  1. Could you report the results over multiple seeds to avoid randomness? Since many comparisons differ by only 0.4-1.5 points.
  2. Could you give the base models the same budgets to make them comparable? In Table 2 for MATH500, the base model has the result 35.00(256) v.s. 34.80(292.1) for dVoting. Could you explain this case?
  3. I see that all base models use the same number of steps as their generation length. Maybe the generation length is not enough for the base model. Could you give me more experiments for longer generation lengths?
Reviewer_fTjbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes dVoting, an efficient majority voting algorithm for diffusion LLMs. Specifically, dVoting utilizes redundant token predictions and spends compute on a small subset of tokens that exhibit cross-sample variability and have a larger impact on reasoning quality. Empirically, dVoting shows comparable performance to baseline test-time scaling methods like majority voting and HEX, while using fewer decoding steps.

Critical strengths
  1. The proposed method is intuitive. It is natural to utilize the redundancy of token predictions and spend compute on those core positions with high variability. The empirical results also demonstrate its effectiveness.
Critical weaknesses
  1. Some experimental settings are unclear and raise some concerns about the fairness of comparison. Please see my questions for details.
Questions for authors
  1. Could the authors clarify the following details for the default decoding, majority voting, and HEX?

    • unmasked positions: at each decoding step, are masked positions selected randomly, by the highest confidence, or by another rule?
    • token sampling: How is the token at each position sampled? By greedy argmax or sampling? If by sampling, could the authors specify the temperature, top-p, and top-k settings?
  2. Similar entropy-threshold parallel decoding has already been used in [1,2], and although not exactly the same, it is not a brand new feature that first appears in dVoting. The main comparisons use the entropy-threshold parallel decoding for dVoting while using fixed-step decoding for the baselines (1 token per step for default and 2 tokens per step for MV and HEX), which can raise concerns about the fairness of comparison. Specifically, to my knowledge, directly applying a fixed 2-tokens-per-step decoding usually brings non-negligible performance degradation, and its efficiency can also be suboptimal. In addition to the existing results, could the authors compare default decoding, majority voting, HEX, and dVoting using the same entropy-threshold decoding rule? Current experiments can not measure the real benefit brought by the unique design of dVoting.

  3. Another concern about the fairness of comparison is the use of early-stopping voting. According to the paper, dVoting stops sampling new candidates when the number of consistent candidate answers exceeds half of the sampling upper bound. However, majority voting and HEX always run out of the candidate budget and can lead to redundant computation, which will make their decoding steps actually higher than necessary. Could the authors also provide results for majority voting and HEX with the same early-stopping voting setting?

  4. What does "Early Stopping" refer to in Table 7? Could the authors provide a detailed explanation?

Reference:

[1] Ben-Hamu, H., Gat, I., Severo, D., Nolte, N. S., & Karrer, B. Accelerated sampling from masked diffusion models via entropy bounded unmasking. NeurIPS 2025.

[2] Chen, Z., Fang, G., Ma, X., Yu, R., & Wang, X. dparallel: Learnable parallel decoding for dllms. ICLR 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.