acceptodds
Under review as a conference paper at ICLR 2027

Deterministic Joint-Risk Sampling for Efficient Discrete Diffusion

Abstract

Masked diffusion language models repeatedly refine masked tokens, often spending forward passes on predictions that have already stabilized. Fixed transfer budgets do not adapt the number of commitments to observed token stability, creating a trade-off between unnecessary refinement and premature commitment. We introduce Deterministic Joint-Risk Sampling (DJRS), which ranks masked tokens using confidence, margin, Jensen–Shannon divergence, and prediction change, then admits additional low-risk commitments under a progress-dependent budget. On LLaDA-8B-Instruct, DJRS reduces the number of forward evaluations (NFE) by 29% on GSM8K at 64 steps (64.0 to 45.6) with a 0.6-percentage-point accuracy decrease, and raises MATH-500 accuracy from 27.8% to 31.4% while reducing NFE by 22%. In a separate Dream-v0-Instruct-7B HumanEval evaluation, DJRS reaches 38.4% pass@1 at 39.8 NFE, compared with 32.9% at 64.0 NFE for Default. Trace analysis shows that margin most consistently separates stable from flip-prone masked positions, while JSD shows less monotonic associations with flips. We present DJRS as a practical adaptive-decoding heuristic rather than a calibrated risk guarantee.

Primary area:
optimization
Keywords:
Masked diffusion language models, LLaDA, Early stopping / token commitment, Decoding acceleration, Joint-risk sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3ZaFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DJRS, a training free sampler for masked diffusion language models. It scores each masked token with confidence, margin, and JSD, and top-1 change flag. It then commits extra low risk tokens so the blocks finish early. The paper evaluates the model on several models and benchmarks.

Critical strengths

NFE savings at 128 steps are fairly large for a training free method, which shows the effectiveness of the method.

Also the authors are upfront that the method is a heuristic and also show the unfavorable results such as 128-step MATH-500, LATCH being better at 128 steps, the GSM8K ablation.

Critical weaknesses
  • The main claim doesn't hold up against the paper's ablations. The performance increase is very small or worse - on GSM8K, confidence only is 65.0% with 39.0 NFE, which beats full DJRS (63.8% at 45.6) on both accuracy and NFE.

  • On MATH-500 30.6% at 45.3 NFE / 31.4% at 49.7, small gap under standard error, and adding JSD drops it to 29.0%. So the savings likely come from extra confident tokens and early blocks finish. Maybe not from margin / JSD / prediction change.

  • Also the paper does not explain about weights and thresholds.

  • Also more empirical evidence is needed. Everything is a single greedy run with no CIs.

Questions for authors
  1. Table 6 shows confidence only results are better than the full DJRS. Does margin, JSD or prediction change actually help, and can you give confidence only numbers for MATH-500 128 steps / GSM8K 128 steps and HumanEval?

  2. How does DJRS compare at matched NFE with Default using fewer steps and with Fast-dLLM's confidence-threshold decoding?

  3. Can you add results with multiple seeds, and also provide details about the choice of weights and thresholds.?

2 comments

Does the evidence establish that all four signals help?

The evidence supports adaptive commitment more strongly than it supports the necessity of all four signals. The confidence-only variant uses the same cumulative-budget selection and block-completion mechanism, and improves both accuracy and NFE on GSM8K relative to the full score. We therefore distinguish two conclusions: the explicit adaptive construction achieves useful measured trade-offs against the fixed schedule, while the full four-signal operating point is not consistently preferable to simpler settings. The MATH-500 differences do not resolve that second question because accuracy and compute change together.

The completed record supplies a GSM8K confidence-only result at nominal budget 128: 68.61% at 45.955 mean NFE. The MATH-500 confidence-only near-compute comparison was run at nominal budget 80, reaching 33.0% at 51.784 NFE; it is not the requested 128-step MATH-500 result. Confidence-only HumanEval results are unavailable. We state those coverage limits explicitly rather than presenting another configuration as the requested experiment.

Default using fewer steps and confidence-threshold decoding

The completed near-compute comparison directly addresses the fixed-schedule question. On GSM8K, full DJRS at nominal budget 128 reaches 68.54% at 57.021 mean NFE; Default at 56 measured NFE reaches 60.05%. The paired difference is +8.49 points, with 95% bootstrap CI +5.91 to +11.07 over the same 1,319 examples. With eight blocks, the scheduler's fixed-work settings occur in multiples of eight; 56 is the nearest legal point, so this is a near-match rather than exact compute equality.

On MATH-500, DJRS reaches 33.6% at 69.572 NFE versus Default at 30.8% and 72 NFE. The difference is +2.8 points, with 95% CI −1.2 to +6.8; this is inconclusive. Default has 2.428 more calls per example, which should remain visible in interpreting the comparison.

These results support DJRS relative to that fixed Default schedule on GSM8K. They do not establish superiority over simpler adaptive confidence policies: confidence-only and the completed local threshold sweep give strong counterexamples, as reported in our response to Reviewer nMBX. We report both the favorable fixed-schedule comparison and the unfavorable simpler-policy comparisons. No latency advantage follows solely from these NFE measurements.

Confidence intervals and repeated seeds

The completed paired analysis uses the saved deterministic outputs, pairs the same problems, and records matching document, prompt, and target hashes. Bootstrap intervals use 20,000 problem-level resamples with seed zero. For the near-compute GSM8K comparison, DJRS is correct on 904 examples and Default on 792, with 212 DJRS-only wins and 100 Default-only wins. This yields the +8.49-point difference and interval above. For MATH-500, the corresponding counts are 168 versus 154 correct and 60 versus 46 discordant wins; the interval includes zero.

Hyperparameter and threshold selection

The exact numerical defaults are specified in our response to Reviewer nMBX. The available record does not establish an original development split, search space, or selection criterion. We therefore make no held-out-selection or robustness claim for those defaults. Weight sensitivity, budget-shape sensitivity, and independent filter/union ablations are not available. The completed threshold sweep is explicitly a test-set diagnostic curve, not a validation-selected operating point. These are limits on how broadly the existing results can be interpreted.

Thank you for the rebuttal. My main concerns are not fully addressed so I keep my score.

Reviewer_v2J7OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes deterministic joint-risk sampling (DJRS), which combines confidence, margin, distributional change, and temporal prediction change into a token-level stability score. It commits low-risk tokens earlier under a progress-dependent budget to reduce denoising iterations. Experiments on mathematical reasoning and code generation evaluate the resulting quality–compute trade-offs.

Critical strengths

The reported results show that DJRS achieves a better trade-off between accuracy and NFE than the evaluated baselines in several settings.

Critical weaknesses
  1. The proposed method relies heavily on heuristic design choices.

    1. Signal weights. Although Tables 5 and 6 provide signal ablations, the paper does not examine sensitivity to the four weights (wcw_c, wmw_m, wjw_j, and wpw_p). Consequently, the relative importance of the signals and the robustness of their weighting remain unclear.

    2. Budget allocation. The budget endpoints (bmin⁡b_{\min} and bmax⁡b_{\max}) and quadratic growth schedule are heuristic choices without corresponding ablations. This leaves the contribution of risk-based token selection entangled with that of budget allocation, making their individual effects difficult to assess.

    3. Complex token selection procedure in LLaDA. The LLaDA implementation preserves a standard confidence-based selection path and adds a separate DJRS path. Although confidence and margin already contribute to the risk score, risk-selected positions are further filtered by hard confidence and margin thresholds before the two sets are merged. This introduces additional heuristic rules and complexity, while the necessity of these extra filters beyond risk-based selection remains unclear.

  2. Limited evidence for the complementary value of the four risk signals.

    Although Section 5.3 examines associations between individual signals and subsequent flips, it does not sufficiently characterize redundancy among the four signals. Moreover, Tables 5 and 6 do not consistently demonstrate a better accuracy–NFE trade-off than simpler variants. In Table 6, confidence-only selection achieves both higher accuracy and lower NFE than full DJRS (65.0% versus 63.8%, and 39.0 versus 45.6 NFE). In Table 5, the full method improves accuracy by only 0.8 percentage points while increasing NFE from 45.3 to 49.7. These results leave the incremental value of the additional signals unclear.

Questions for authors

Please refer to Weaknesses 1 and 2.

1 comment

Why these signals and this budget?

The signals measure different observable quantities: confidence measures mass on the leading token, margin measures separation from the next candidate, JSD measures distributional change, and the prediction-change indicator records a change of the top token. A distribution can change while its argmax remains the same, and an argmax can change after a small probability shift near a tie. This motivates considering both temporal signals; it does not prove that combining them improves final-answer accuracy.

The budget controls the aggregate score of the additional risk-selected prefix. With fixed endpoints, quadratic growth allows less extra budget at intermediate progress than a linear schedule, expressing the heuristic preference to be more conservative early. The budget resets at each iteration. Because the standard scheduled set is retained separately, the budget is not a calibrated bound on the total union's error or score. We do not claim optimality of the quadratic form, endpoints, or coefficients, and the existing ablations do not establish their sensitivity.

Why retain the standard path and use additional filters?

The standard path preserves scheduled progress: if the risk prefix is empty or its candidates fail eligibility checks, decoding still proceeds using the ordinary quota. The confidence/margin filters restrict additional commitments even when a position ranks favorably under the weighted score. This makes the policy a conservative addition to the schedule by construction; “conservative” here describes eligibility, not a proven accuracy guarantee.

In the implementation, ranking and cumulative-budget selection occur before filtering. Overlapping scheduled positions and subsequently rejected candidates can consume that prefix budget, and there is no refill after filtering. This is the precise implemented rule. The filters and union have not been independently ablated, so the rationale explains what each rule does without establishing that each is necessary or that this ordering is best.

Complementarity and the ablation evidence

We agree that signal-composition ablations do not consistently establish complementarity. On GSM8K, confidence-only gives the better observed operating point. On MATH-500, the original 64-step table reports 30.6% at 45.3 NFE for confidence-only and 31.4% at 49.7 NFE for all four signals: the additional 0.8 points come with greater compute. Adding JSD in the intermediate variant also reduces accuracy in that table. These should be read as measured operating points, not isolated causal benefits of each signal.

The historical MATH-500 run record does not persist the exact effective ablation weights. The recorded code chronology is compatible with uncommitted changes and does not establish that incorrect weights were used. Nevertheless, that provenance gap makes a strong per-signal conclusion inappropriate. The completed near-compute comparison also remains inconclusive: full DJRS obtains 31.4% at 49.728 NFE versus confidence-only at 33.0% and 51.784 NFE, with a recorded DJRS-minus-confidence-only interval of −5.6 to +2.4 points.

The flip diagnostic provides additional evidence that the signals can distinguish instability within confidence strata. For block size 64, the lowest-confidence bin has future-flip rates of 39.71% and 16.21% in the lowest and highest margin quintiles, respectively. This supports examining margin beyond confidence alone, although predictive information need not translate into a better decoding policy under the chosen coefficients. The diagnostic offers a separate motivation. It compares future-flip frequencies within confidence strata, using a separated calibration/evaluation split. Its signals have lagged timing and its population excludes positions committed during the measured transition. Thus, it does not directly validate the decoder's current-step decisions, calibrate the combined score, or establish the final-answer benefit of joint ranking.

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

The context of this work is inference strategies for Masked Diffusion Language Models. Concretely it proposes a heuristic to decide which masked positions to commit early, on top of the standard low-confidence schedule, based on a weighted sum of signals that are readily available in decoding, these are

  • Confidence: the top-1 probability
  • Margin: the gap between the top-1 and top-2 probabilities
  • Jensen-Shannon divergence between the token's predictive distributions at consecutive steps
  • Prediction change: whether the top-1 prediction changed relative to the previous step

The combination of signals is used to rank positions by risk, select those that fit a cumulative budget that grows quadratically within each block, and filter based on confidence and margin to produce a decision. The method is evaluated with LLaDA-8B-Instruct on GSM8K and MATH-500 and with Dream-v0-Instruct-7B on HumanEval.

The method is simple and reports NFE reductions at comparable accuracy against related approaches, but there seem to be flaws in the experimental setup and missing information that needs to be clarified.

Critical strengths
  • It proposes a simple method using a weighted combination of standard signals available at test time to decide on token unmasking

  • It compares favorably on a battery of related baselines (LATCH, Prophet, SureLock), particularly in Coding (Human Eval)

  • With 128 steps, DJRS is more accurate than Default with 64 steps on all three tasks at similar or lower NFE (e.g. HumanEval 52.4% at 49.6 NFE vs 32.9% at 64 NFE)

Critical weaknesses

The experimental setup has several deficiencies.

  • The paper does not seem to state how the weights for different signals are determined and on which data. This is a core element of the proposed approach.

  • The canvas size L is also missing and is necessary to evaluate the results against the number of steps setup. L/T determines how many tokens Default commits per step, and thus how much accuracy it loses to parallel commitment.

  • Some baselines look odd. The reported baseline results seem lower than in the original papers and some comparisons seem not correct:

    • Fast-dLLM (arXiv:2505.22618, Table 1) reports LLaDA-8B-Instruct at 79.3% on GSM8K (L=256, 5-shot), vs 68.8% / 64.4% for Default here at 128 / 64 steps (-10.5 / -14.9 pt). For MATH it reports 33.5% (L=256, 4-shot), vs 27.8% here at 64 steps (-5.7 pt).
    • Prophet (arXiv:2508.19982, Table 3a) reports 76.2% for LLaDA on GSM8K with L=256 and a 128-step budget, vs 68.8% for Default here (-7.4 pt).
    • Prophet (arXiv:2508.19982, Table 1) reports no accuracy loss with 1.63x fewer steps on LLaDA GSM8K (77.1% -> 77.9%). Here Prophet gives no NFE reduction on GSM8K (64.0 and 128.0 NFE) and collapses on MATH-500 (13.4% / 10.6%, i.e. -14.4 / -24.8 pt vs Default). The thresholds used here (7.5, 5.0, 2.5) are Prophet's general-task values; Prophet uses (8.0, 5.0, 3.5) for GSM8K and (7.5, 5.0, 4.5) for code.
    • SureLock (arXiv:2602.06412, App. J Table 6) reports HumanEval pass@1 going from 18.6% to 20.3% at 0.47x FLOPs with LLaDA-8B-Instruct, i.e. no quality loss. Here SureLock loses 14.6 / 25.8 pt on GSM8K and 15.4 pt on MATH-500 relative to Default.

    This can be related to odd choices of L or shots discussed above, or even implementation deficiencies.

  • SureLock efficiency comes from reduced FLOPS not lower NFE, so using this for comparison is not meaningful

  • The ablations do not support the multi-signal score. On GSM8K (Table 6) confidence alone is better than all four signals on both accuracy and NFE (65.0% at 39.0 NFE vs 63.8% at 45.6 NFE). On MATH-500 (Table 5) the differences between variants are not that big.

  • Fast-dLLM confidence-threshold parallel decoding is not included as a baseline, although the experiments use the Fast-dLLM pipeline.

  • The HumanEval results use a different model and a different DJRS integration (App. F), so they say little about the LLaDA method.

Questions for authors
  • What generation length L was used for each benchmark?

  • What few-shot setting and prompt format were used?

  • How were the weights, budget and thresholds chosen, and on which data?

2 comments

Generation length, shots, prompts, and evaluation setup

For LLaDA on GSM8K and MATH-500, the generation canvas contains 256 output positions, excluding the variable-length prompt. Decoding uses 32-token blocks, temperature zero, zero-shot prompting with the model chat template, and batch size one. The recorded task templates are Question: {question}\nAnswer: for GSM8K and Problem:\n{problem}\n\nSolution: for MATH-500, followed by model chat-template application. Scoring uses flexible numeric extraction for GSM8K and math-verify for MATH-500. We tried reproducing the other paper and arrived at this conclusion, we don't claim to exactly reproduce what the previous papers claim

Weights, budget, and threshold choices

The default weights for confidence, margin, JSD, and prediction change are respectively 0.35, 0.35, 0.20, and 0.10. The block-local budget grows quadratically from 0.5 to 4.0; additional selected positions must satisfy confidence at least 0.60 and margin at least 0.10. The decoder computes JSD using the full vocabulary.

The coefficients and schedule define a reproducible heuristic operating point; the method does not rely on interpreting the score as a calibrated probability. Their design rationale is the separation of current uncertainty from temporal instability and a budget that permits fewer extra commitments early.

Simpler confidence-based alternatives

The ablation usefully identifies a simpler, stronger setting within the same budget-based commitment construction. Confidence-only retains the adaptive prefix budget and early block completion while setting the other signal weights to zero; its performance therefore supports the usefulness of confidence-driven adaptive commitment, while challenging the need for the full score. The original GSM8K 64-step ablation favors confidence-only: 65.0% at 39.0 NFE versus full DJRS at 63.8% and 45.6 NFE. In the completed follow-up record, confidence-only at nominal budget 128 reaches 68.61% at 45.955 NFE, versus full DJRS-64 at 63.76% and 45.561 NFE. Its paired advantage is 4.85 points, with 95% CI +2.43 to +7.35. Against DJRS-128, confidence-only gives 68.61% versus 68.54% at lower NFE, with an inconclusive paired accuracy difference.

The completed local confidence-threshold sweep reaches 70.89%, 70.74%, and 72.02% at thresholds 0.7, 0.8, and 0.9, respectively, with 48.766, 60.651, and 77.489 mean NFE. The 0.7 point has higher observed accuracy and lower NFE than DJRS-128. The sweep is exploratory because thresholds were examined on test data, and its local implementation is not externally validated. These qualifications do not justify dismissing its stronger performance. We do not claim a general advantage of joint ranking over confidence-based alternatives.

The authors have only partially answered the questions. Knowing L=256 is useful but the differences with reported baselines remain unexplained. More importantly it is still unclear how the weights were computed (concrete procedure, which data). For the rest the author provided no additional information that changes my view of the manuscript. I will keep current score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response

We thank the reviewers for recognizing the practical value of reducing diffusion decoding cost. DJRS introduces an explicit mechanism for adaptive token commitment: it preserves scheduled transfers, ranks additional candidates using observable stability signals, and admits them under a progress-dependent budget. This completes blocks earlier without changing backbone weights. At 128 nominal steps on GSM8K, DJRS achieves 68.54% accuracy at 57.02 forward evaluations, compared with 68.8% at 128 evaluations for the submitted Default configuration. The completed comparison against fewer-step Default also shows that the benefit at this operating point is not explained simply by requesting fewer fixed steps: DJRS reaches 68.54% at 57.02 NFE versus 60.05% at 56 NFE, with a paired difference of +8.49 percentage points (95% CI +5.91 to +11.07).

These results support the adaptive commitment mechanism relative to the specified fixed schedule. The signal ablations and stronger confidence-based comparators define its comparative limits: they do not establish that the four-signal configuration is consistently best. We address both the positive evidence and those limits below, while clarifying the setup, design rationale, and distinction between NFE and runtime.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.