acceptodds
Under review as a conference paper at ICLR 2027

Soft Confidence Order in Masked Diffusion: Quality–Diversity Trade-offs Across Models

Abstract

Masked diffusion language models predict all masked tokens in one forward pass, leaving the decoder to decide which predictions to commit at each model call. Most-confident-first selection is common, but its strength is rarely isolated from other decoding choices at a fixed call budget. We study a one-parameter family of position-selection rules spanning confidence-free selection, soft confidence-weighted order, and hard most-confident-first ranking. We separate position selection from the time grid (tokens committed per call) and token revision. On two 124M checkpoints and LLaDA-8B-Base, we compare rules at equal call budgets using perplexity under a fixed GPT-2-large scorer, bigram diversity, and MAUVE. At temperature 1.0 and 32 calls on the 124M model, soft order is the best fixed-count rule tested: it lowers scorer perplexity by about 12% relative to both the standard ancestral sampler and a confidence-free fixed-count control, whereas hard ranking roughly doubles it. On LLaDA-8B-Base at the same temperature, the reduction relative to the ancestral sampler is 11–23% at 16–64 calls, with all 95% intervals excluding zero. On the 124M model, the gain over the ancestral sampler persists among outputs matched on per-document bigram diversity. One-shot revision trades diversity for a further 31% reduction in perplexity, while time grids fitted to pilot-run confidence lower ancestral-sampler perplexity by about 4% at the smallest budgets. Cooling to favors harder weighting. Decoding gains should be reported alongside call budget, count rule, and diversity.

Primary area:
generative models
Keywords:
masked diffusion language models, confidence-based decoding

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_4dvgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper measures how confidence-weighted position selection affects masked-diffusion decoding at fixed model-call budgets. A Plackett–Luce family interpolates between uniform selection and hard confidence ranking, with a fixed-count control used to isolate ordering on 124M models. The study also varies token revision, time grids, temperature, and diversity, and reports transfer to LLaDA-8B-Base. The main positive result is a scorer-perplexity improvement from soft weighting at selected operating points; revision helps the small model but hurts LLaDA.

Critical strengths
  1. The study makes several distinctions that are often obscured in decoding comparisons: model calls versus tokens committed per call, order versus revision, and scorer quality versus diversity. The fixed-count β=0 control is particularly useful. Multiple decode seeds, alternative scorers, and explicit accounting for extra revision calls make the small-model comparisons more informative than a single sampler leaderboard.

  2. The paper reports important negative and conditional findings. The preferred weighting depends on temperature; revision does not transfer successfully to LLaDA; and a diversity-calibrated nucleus baseline outperforms one-shot revision on both scorer perplexity and MAUVE. These results help delimit when a decoding intervention is useful and give the work value as an empirical study.

Critical weaknesses
  1. The key controlled finding has limited support in a practical generation regime. The fixed-count β=0 and hard-ranking controls are evaluated on internal 124M checkpoints whose outputs the paper itself describes as far from fluent. On LLaDA, soft order changes both position weighting and the count rule relative to the ancestral baseline; Appendix B explicitly acknowledges that neither control was run there. Thus, the public-model result does not establish that confidence weighting itself causes the reported improvement or that the small-model preference over hard ranking transfers. Since the contribution is a controlled measurement of an existing sampling family, completing that identification on a useful public model is important to its significance.

  2. The quality conclusion remains tied to proxy operating points. Larger scorers preserve several rankings, which is helpful, but perplexity under a reference language model does not establish useful generation quality. Matching outputs after generation on one bigram statistic does not hold the output distribution fixed, and joint diversity matching is too small to resolve the effect. The nucleus comparison also shows that the apparent revision advantage can reverse under another reasonable diversity control. I appreciate that these limitations are disclosed, but the current evidence does not yet establish enough new knowledge about useful quality–diversity trade-offs to support acceptance. I therefore lean toward rejection.

Questions for authors
  1. On LLaDA, does soft confidence weighting still improve over β=0 with exactly the same fixed-count schedule, and how does hard ranking compare under identical token sampling? These are the comparisons most likely to change my recommendation.

  2. On the same public model and call budget, does the advantage remain when temperature or nucleus sampling is tuned to comparable diversity without post-selecting generated documents? A modest human or task-based evaluation of those operating points would help determine whether the scorer improvements correspond to useful quality.

  3. Please reconcile Appendix D’s statement that its public-model setup isolates the order channel with Appendix B’s explicit statement that the fixed-count control was not run on LLaDA. The exact count rule for each public-model arm should be unambiguous.

1 comment

Thank you again for your time and your thoughtful review. We would be grateful to hear whether our response has addressed your concerns, and we are happy to provide any further clarification.

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

This paper studies how masked diffusion language models select which masked positions to fill during decoding. The authors propose a confidence-based position selection method controlled by a parameter β, ranging from random selection to hard confidence ranking. The paper also examines how position selection interacts with decoding schedules and token revision. Experiments are conducted on two 124M MDLM checkpoints and LLaDA-8B-Base, using perplexity, diversity, and other generation quality metrics. Results show that soft confidence-based selection reduces perplexity by approximately 12% compared to baseline methods, while hard confidence ranking performs substantially worse. The improvements are consistent across different models, decoding budgets, and experimental settings. The authors also investigate revision strategies and decoding schedules, finding additional improvements in some settings, often with trade-offs in diversity.

Critical strengths

This paper has the following major strengths:

  1. Careful experimental design and evaluation methodology. The authors carefully control for potential confounding factors by comparing decoding strategies under matched model-call budgets and introducing a fixed-count baseline to isolate the effect of confidence-based selection. They also control for output diversity, use additional scoring models, and report bootstrap confidence intervals. These choices strengthen the reliability of the experimental conclusions.

  2. Insightful experimental findings. The experiments reveal several useful insights into masked diffusion decoding. The authors also identify limitations of revision strategies and confidence-based time schedules. These findings provide practical guidance for balancing generation quality, diversity, and computational cost.

Critical weaknesses

This paper has the following major weakness:

  1. Limitations in evaluation metrics and experimental settings. The main findings rely heavily on a small 124M model with relatively low generation quality. The evaluation primarily uses perplexity, without human judgments or downstream task results, making it difficult to assess actual generation quality. The paper also lacks comparisons with practical LLaDA decoding configurations and existing selection methods. These limitations raise concerns about the generalizability of the findings to real-world decoding settings.

  2. Limited analysis of how different selection strategies affect decoding. The paper primarily reports performance results without examining how different selection strategies influence the decoding process. For example, it remains unclear which positions or token types are prioritized by soft and hard confidence ranking, and why hard ranking performs substantially worse. A deeper analysis of these differences would help explain the observed improvements and provide insights into whether the findings generalize to other models and decoding settings.

Questions for authors
  1. Is soft order with a fixed β equivalent to MaskGIT’s confidence-plus-Gumbel selection with τ = 1/β? Could the authors clarify the differences between these methods?

  2. Hard ranking (β → ∞) is only evaluated on the 124M model at T = 1.0, while the results at T = 0.7 favor stronger confidence weighting. How does hard ranking perform at T = 0.7 on both models and at T = 1.0 on LLaDA?

  3. How does soft order compare with or complement LLaDA’s recommended block-wise decoding with low-confidence remasking under matched model-call budgets?

11 comments

Thank you again for your time and your thoughtful review. We would be grateful to hear whether our response has addressed your concerns, and we are happy to provide any further clarification.

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts to address my concerns. The added experiments largely address W2. As the authors acknowledged, with a fixed β\beta, the proposed selection method is mathematically equivalent to MaskGIT's approach, so the major contribution would be the systematic empirical study. Therefore, my focus when reading this paper is: does the empirical evaluation show and explain when and why soft confidence-based position selection impacts performance? The newly added experiments in the rebuttal strengthen the decoding behavior analysis and also support the major findings. However, the experiments cover only two model families, and most detailed analyses are conducted on the 124M model. Although the experiments on decoding configurations are systematic, the limited model coverage means that W1 is only partially addressed. This leaves me uncertain whether the major findings can be generalized to other model families. Therefore, I will maintain my original score.

AuthorsOpenReview ↗

Response2 to Reviewer Qyzg (1/8)

Thank you for the follow-up and for the careful reading of our response. We agree with the remaining concern: with two model families, and with most of the detailed analyses on the 124M model, it is fair to ask whether the main findings generalize. Over the past few days we have therefore run a substantially larger set of experiments: 14 further public checkpoints in addition to LLaDA-8B-Base, evaluated with perplexity and with four task benchmarks (GSM8K, HumanEval, MBPP, and MATH).

1. Perplexity. Unconditional generation of 256 tokens, T=1.0, decode seed 777, 1,024 documents per decoder, GPT-2-large scorer. The first row repeats the LLaDA-8B-Base result of our earlier response; the second row is a rerun of the same setting on the cluster used for the new experiments.

ModelModel callsPPL: ancestral / β=0 / β=1 / hard rankingβ=1 vs. ancestral (%)β=1 vs. β=0 (%)β=1 vs. hard ranking (%)
LLaDA-8B-Base (earlier response)3297.64 / 97.34 / 86.01 / 181.4411.91 [1.28, 21.43]11.63 [1.12, 20.77]52.59 [46.74, 57.81]
LLaDA-8B-Base (rerun)3299.54 / 96.94 / 89.22 / 203.2110.37 [−0.80, 20.24]7.96 [1.51, 13.67]56.09 [50.43, 61.14]
LLaDA-8B-Instruct3295.41 / 101.39 / 84.92 / 305.4810.99 [4.05, 17.37]16.24 [11.83, 20.48]72.20 [68.41, 75.73]
LLaDA-8B-Instruct16140.31 / 139.35 / 130.75 / 384.276.81 [−0.10, 13.16]6.17 [2.40, 9.65]65.97 [62.27, 69.35]
LLaDA-1.53288.17 / 91.22 / 76.99 / 357.6112.68 [5.48, 19.21]15.60 [10.55, 20.50]78.47 [75.59, 81.05]
LLaDA-1.516137.74 / 135.27 / 123.32 / 438.8510.47 [3.56, 16.79]8.84 [5.17, 12.46]71.90 [68.94, 74.59]
iLLaDA-8B-Base3283.80 / 79.18 / 70.70 / 273.4015.63 [8.17, 22.38]10.71 [6.56, 14.68]74.14 [71.46, 76.60]
iLLaDA-8B-Base16148.54 / 141.97 / 135.16 / 340.829.01 [1.46, 16.10]4.80 [0.80, 8.69]60.34 [56.98, 63.50]
iLLaDA-8B-Instruct32124.31 / 118.27 / 103.40 / 638.0216.82 [10.83, 22.31]12.57 [8.38, 16.50]83.79 [82.47, 84.99]
iLLaDA-8B-Instruct16228.94 / 229.70 / 205.81 / 895.1210.10 [3.83, 15.80]10.40 [6.60, 14.05]77.01 [75.40, 78.51]
LLaDA-MoE-7B-A1B-Base3275.31 / 70.61 / 65.91 / 198.4512.47 [4.95, 19.53]6.66 [2.38, 10.73]66.79 [62.90, 70.32]
LLaDA-MoE-7B-A1B-Base16127.60 / 124.29 / 115.69 / 212.759.33 [2.04, 15.93]6.92 [3.42, 10.16]45.62 [40.25, 50.43]
LLaDA-MoE-7B-A1B-Instruct3273.08 / 71.52 / 64.56 / 165.0911.65 [5.07, 17.52]9.73 [6.21, 13.05]60.89 [55.96, 65.23]
LLaDA-MoE-7B-A1B-Instruct1695.64 / 99.21 / 92.77 / 172.293.00 [−3.57, 9.12]6.49 [3.31, 9.50]46.16 [40.85, 51.04]
Dream-v0-Base-7B32274.43 / 243.46 / 219.11 / 960.9120.16 [15.07, 24.87]10.00 [5.63, 14.17]77.20 [75.74, 78.56]
Dream-v0-Base-7B16517.76 / 496.15 / 437.52 / 1186.7615.50 [10.51, 20.10]11.82 [7.91, 15.54]63.13 [61.01, 65.17]
Dream-Coder-v0-Base-7B32325.64 / 289.86 / 255.27 / 1296.5821.61 [11.65, 30.38]11.93 [4.11, 19.11]80.31 [77.75, 82.58]
Dream-Coder-v0-Base-7B16561.62 / 496.76 / 458.85 / 1259.0218.30 [8.64, 26.99]7.63 [1.19, 13.71]63.56 [59.44, 67.19]
DiffuCoder-7B-Base32105.60 / 90.64 / 78.31 / 323.3125.84 [17.51, 33.27]13.60 [7.50, 19.34]75.78 [72.96, 78.28]
DiffuCoder-7B-Base16159.84 / 154.06 / 142.74 / 348.2710.70 [1.45, 18.98]7.35 [1.59, 12.68]59.02 [54.54, 63.01]
MDLM (mdlm-owt)32175.89 / 168.26 / 135.15 / 1022.8023.16 [19.70, 26.47]19.68 [16.84, 22.35]86.79 [85.59, 87.86]
MDLM (mdlm-owt)16283.35 / 274.10 / 229.43 / 2142.9319.03 [15.62, 22.33]16.30 [13.75, 18.80]89.29 [88.80, 89.76]
SEDD-small32156.84 / 151.91 / 129.31 / 1066.3117.55 [14.05, 20.93]14.87 [12.21, 17.45]87.87 [87.07, 88.61]
SEDD-small16248.90 / 238.53 / 213.76 / 2330.7014.12 [10.68, 17.39]10.38 [7.91, 12.70]90.83 [90.46, 91.18]
SEDD-medium32121.92 / 119.97 / 104.79 / 916.1014.05 [10.47, 17.53]12.66 [9.80, 15.42]88.56 [87.90, 89.17]
SEDD-medium16205.81 / 200.36 / 174.58 / 1647.4815.17 [11.66, 18.52]12.87 [10.45, 15.21]89.40 [88.94, 89.84]
Reviewer_coSgOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks which masked positions a diffusion language model should commit at each forward call. It places existing practice on a one-parameter family, from ignoring confidence to hard most-confident-first, and compares the settings at a fixed number of model calls against both the standard sampler and a confidence-free control, so only the confidence weight varies. On two small checkpoints and LLaDA-8B, an intermediate soft weighting gives the lowest GPT-2 perplexity, hard ranking roughly doubles it, and the gain carries over to the larger model. Revision and confidence-profiled time grids are measured as well.

Critical strengths
  1. The matched-call protocol is unusually careful. Four decode seeds, two checkpoints, three training seeds, two sequence lengths, three scorer models and a document bootstrap, plus a diversity-matched comparison, is more control than this literature usually applies, and the paper reports the case where its own method loses: a nucleus baseline calibrated to the same diversity reaches better perplexity and MAUVE than one-shot revision.
  2. The finding that hard most-confident-first ranking is harmful, roughly doubling perplexity, is a useful negative result for a practice that is widespread, and the authors separate it from the count schedule rather than attributing the whole gap to confidence.
Critical weaknesses
  1. Quality is measured by the perplexity that GPT-2-large assigns to generated text, which falls when text becomes more generic, and diversity falls along with it: corpus distinct-2 drops from 0.557 to 0.397 along the revision ladder. The matched-diversity check only matches per-document bigram diversity, so the deeper revision gains are largely bought by diversity loss, as the authors' own calibrated nucleus baseline shows. No task-based or human evaluation is reported, so it is unclear whether the lower perplexity corresponds to better text rather than safer text.
  2. The strength grid is coarse. With only zero, one half, one, two and hard as measured points, the claim that one is the best setting is bracketed only between one half and two, and at temperature 0.7 the preference moves to two while on the 8B model at 64 calls the intervals for one and two overlap. The paper states "soft order is best" more firmly than the measurement resolution supports.
Questions for authors
  1. Can you add one task-based evaluation, or a small human judgement, on the 8B model? Perplexity under a small scorer may reward generic text.
  2. For LLaDA, how much of the reported interval is seed variance rather than document variance? Please report two or three decode seeds with an interval that includes them.
  3. Is there a reason beta of one should be special, rather than the best point of a coarse sweep? A finer grid between one half and two would settle it.
1 comment

Thank you again for your time and your thoughtful review. We would be grateful to hear whether our response has addressed your concerns, and we are happy to provide any further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.