acceptodds
Under review as a conference paper at ICLR 2027

When to Stop and Which to Return: Trajectory-Level Minimum Bayes Risk Decoding for Diffusion Language Models

Abstract

Masked diffusion language models can produce high-quality intermediate responses before denoising finishes, yet continued decoding may incur unnecessary computation or overwrite better responses. Existing approaches address these challenges separately. Early-commitment methods reduce computation but cannot recover earlier responses. Temporal voting exploits intermediate responses to improve quality but retains the cost of full decoding. We propose *Trajectory-MBR*, a training-free method that jointly decides when to stop and which current or earlier response to return. It reuses existing denoiser outputs to construct complete response candidates and performs trajectory-score-weighted minimum Bayes risk selection within a sliding window. Once the minimum weighted risk falls to or below a threshold, decoding stops and returns the selected response. We show that accurate trajectory-based utility estimates yield near-optimal selection within the window and provide an offline condition for improvement over full decoding. Across language and multimodal tasks, Trajectory-MBR maintains or improves average response quality with average denoiser-evaluation speedups of up to , while its token-level extension preserves response quality at a speedup on HumanEval.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Masked Diffusion Language Models, Minimum Bayes Risk Decoding, Early Stopping

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ngZ8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of decoding masked diffusion language models, aiming to find a better balance between premature stopping that leads to immature answers and excessive decoding that may degrade performance. The authors propose a decoding method called Trajectory-MBR, which combines a sliding window with trajectory-likelihood-weighted minimum Bayes risk. The theoretical analysis centers on trajectory score approximation error and conditional regret bounds for response selection. Experiments on language and multimodal benchmarks provide empirical support for the method, showing improvements in performance and efficiency.

Critical strengths

The algorithm brings clear improvements in computational efficiency and provides additional design flexibility for future work by allowing the output to be selected from historical checkpoints rather than always returning the latest one before early stopping. In the reported settings, the method also achieves competitive answer quality, though its advantage over baselines such as Sparse-dLLM appears relatively small in some settings.

Critical weaknesses

One major weakness of the paper is that it does not provide sufficient insights into the underlying mechanism. The theory part helps little here. Theorem 1 mainly characterizes the approximation error of the trajectory score. While Theorem 2 and Proposition 1 rely only on general assumptions about the utility function, their bounds do not explain what kind of utility or weighting plays a key role in the algorithm. Although the examples in Section 6.2 illustrate when temporal agreement helps, this phenomenon is already well established rather than being the main contribution of this paper. This may leave readers unclear about why this particular design works well, especially since the trajectory scoring scheme is, in my opinion, one of the unique contributions of the paper. Also, the experiments do not clearly separate the gains from different components of the algorithm. I think a comparison with simpler weighting schemes under the same stopping framework would help clarify the contribution of the trajectory score.

Questions for authors
  1. Could the authors provide more insights into the mechanism behind the algorithm? In particular, why should the proposed trajectory-likelihood weighting work better than simpler weighting schemes? I think this would help clarify the unique contribution of the method.

  2. Could the authors compare the proposed method with a fixed-budget early-exit baseline under the same inference budget? For example, stopping the base decoder after a fixed number of steps and returning its current prediction. I think this would help clarify how much of the efficiency and quality gains come from adaptive stopping rather than simply reducing the number of decoding steps.

  3. How often does the method select a response different from the final response of full decoding? Under what circumstances does this selection improve or degrade answer quality? I think such an analysis would help readers better understand the situations where the algorithm is most effective and its limitations.

10 comments

We sincerely thank the reviewer for the careful reading and the constructive feedback. We address each point below.

W1-1. Temporal agreement is already well established

We appreciate this comment. Masked diffusion language models can overwrite a good response that was already generated earlier in decoding, and work on temporal voting also studies this phenomenon. The distinguishing point of our method is that it performs this recovery without paying the cost of full decoding. Existing methods obtain only one of the two benefits. Early-commitment methods reduce computation but cannot recover an earlier response, while temporal voting recovers earlier responses but still pays the full decoding cost.

Trajectory-MBR jointly decides when to stop and which response to return, so it obtains both benefits at once. When the terminal state is poor, it recovers a better earlier response and improves quality. When the terminal state is already good, it preserves quality while reducing computation through early stopping. This is consistent with the experiments and the theoretical analysis in the paper.

W1-2, Q1. Why trajectory-likelihood weighting works better than simpler weighting

We agree that the existing theoretical analysis alone did not sufficiently explain the role of the weights. We summarize the rationale for the weighting design in three points, and we added ablations that compare against simpler weightings under the same stopping framework to Table 4 and Table 5 in Section 6.2 of the revised manuscript.

Rationale for the weighting design

First, the trajectory score approximates the likelihood of a candidate. Theorem 1 and Equation 9 in Section 4.2 show that the window weights approximate the likelihood weights under the stochastic ancestral counterpart. In the controlled experiment of Figure 2, the mean TV between the two sets of weights was below 0.10, and replacing the trajectory-score weights with exact likelihood weights selected the same response in 81.6% to 90.0% of the windows that contain distinct responses. The weighted risk of Equation 5 therefore places more weight on candidates that the model considers more plausible, which is the same principle by which MBMBR (Jinnai et al., 2024) improves over uniform MBR.

Second, it compares candidates from different checkpoints on a common scale. The trajectory score of Equation 2 multiplies the token probabilities at all LL positions for every candidate. If only the probabilities of the current completion were multiplied, later candidates with fewer masked positions would have fewer factors and would be favored. Completion confidence avoids this by averaging probabilities, but it discards the history of the tokens that are already revealed, and uniform weights do not reflect probability differences at all.

Third, the weights also determine when to stop. The stopping condition of Equation 7 uses the weighted risk of Equation 5 directly. With exact match and K=4K = 4, as on LSHT, the uniform risk changes only in steps of 0.25, so satisfying ϵ=0.05\epsilon = 0.05 requires all four candidates to be identical. With trajectory weights, a single low-scoring candidate that gives a different answer raises the risk only slightly, so stopping is barely delayed.

Comparison under the same stopping framework

In Table 4 and Table 5 in Section 6.2 of the revised manuscript, we fix the UltraLLaDA settings of Table 1 and change only the weighting and the selection rule. The evaluation uses 200 LSHT examples and 500 LCC examples. Default uses the same KK, utility, and ϵ\epsilon, and Tuned adjusts only ϵ\epsilon so that the speedup approaches that of Trajectory-MBR.

MethodLSHT ScoreLSHT SpeedupLCC ScoreLCC Speedup
Trajectory-MBR42.5012.94×68.301.21×
Uniform weights, Default41.505.13×68.071.15×
Uniform weights, Tuned42.509.57×68.031.21×
Completion confidence, Default41.505.13×68.071.15×
Completion confidence, Tuned41.0012.80×68.031.21×
Current candidate, same stopping point42.0012.94×68.181.21×
Deduplicated Trajectory-MBR42.5012.92×68.281.21×
Best simple rule at the same speedup, Table 541.5012.94×68.081.21×

Trajectory weights reach the same quality much faster. Under the same settings, switching to a simpler weighting reduces the LSHT speedup from 12.94× to 5.13×, so Trajectory-MBR is 2.5 times faster. Trajectory-MBR also scores higher on both tasks. These results appear in the Uniform weights and Completion confidence columns of Table 4.

Simpler weightings do not catch up even at a matched speed. When ϵ\epsilon is adjusted to match the speedup, both simpler weightings reach 68.03 on LCC, which is 0.27 points below Trajectory-MBR, and completion confidence is 1.50 points lower on LSHT. Uniform weights need 35% more denoiser evaluations to reach the same 42.50 on LSHT.

Selecting earlier candidates raises the score at no extra cost. Stopping at the same point and returning the current candidate gives 42.00 and 68.18, whereas also selecting among the earlier candidates in the window with trajectory weights gives 42.50 and 68.30 with the same computation. This is the Current candidate column of Table 4.

The effect does not come from repeated support for identical responses. Keeping only one copy of each identical string and recomputing changes the scores by at most 0.02 points and leaves the speed essentially unchanged. This is the Deduplicated column of Table 4.

No simple rule at the same speed matches Trajectory-MBR. In Table 5, confidence stopping, majority voting, and confidence argmax stay at or below 41.50 on LSHT and 68.08 on LCC even when the speedup is matched.

In summary, Theorem 1 supports that the trajectory score approximates the likelihood, and Table 4 and Table 5 show that this weighting in practice stops earlier while selecting better candidates.

Q2. Fixed-budget early exit at the same inference budget

Following the reviewer's suggestion, we compared against a baseline that stops the base sampler at a fixed budget and returns the current candidate at that point. We report both a comparison with the budget matched exactly to Trajectory-MBR and a comparison across all 13 tasks at several budgets.

Comparison at the same budget. The Matched budget columns of Table 5 in Section 6.2 of the revised manuscript stop after 5 evaluations on LSHT and 53 on LCC, matching the average cost of Trajectory-MBR. Returning the current candidate at the same budget gives 41.00 on LSHT and 68.05 on LCC, and selecting by majority voting or confidence argmax does not exceed 41.00 and 68.05. Trajectory-MBR obtains 42.50 and 68.30 at the same budget.

Comparison across all 13 tasks. The table below compares fixed budgets that correspond to 2.00×, 2.92×, and 4.00× speedups on the 13 LongBench tasks for both backbones.

Table R4. Comparison of fixed-budget early exit and Trajectory-MBR on UltraLLaDA and Dream-7B-Instruct. † denotes results reported by Long et al. (2026). Fixed-budget early exit has a lower average than Trajectory-MBR even with similar or more computation, which shows that the gain does not come from simply reducing the number of steps.

UltraLLaDAQasperMF-enHotpotQA2WikiMQAMuSiQueGovReportQMSumTRECTriviaQALSHTPReLCCRB-PAvg. ScoreAvg. Speedup
Full decoding†19.1425.8716.2718.0012.0832.8322.4880.0091.5841.0096.7568.2359.5044.901.00×
Fixed budget, 2.00×20.8630.0716.8217.7112.5029.9822.3180.0091.5841.0093.2868.0259.2344.872.00×
Fixed budget, 2.92×23.3734.2818.6519.6914.1427.3722.5380.0091.5841.0093.1667.9858.9245.592.92×
Fixed budget, 4.00×24.5836.2625.5825.1618.6925.3922.6580.0090.9341.0085.1267.8058.3546.274.00×
Trajectory-MBR23.4134.2528.0327.7620.3032.8822.9180.0091.5842.5097.3368.3059.5048.373.76×
Dream-7B-InstructQasperMF-enHotpotQA2WikiMQAMuSiQueGovReportQMSumTRECTriviaQALSHTPReLCCRB-PAvg. ScoreAvg. Speedup
Full decoding†35.5840.4941.5942.1023.3623.5119.6671.5087.3415.7532.5063.7962.2343.031.00×
Fixed budget, 2.00×34.6738.9834.8533.8415.0619.9219.7972.6686.7619.4428.7663.3150.6539.902.00×
Fixed budget, 2.92×33.6838.5433.2232.7615.3519.1619.6672.6686.8019.4429.6562.8449.7039.502.92×
Fixed budget, 4.00×32.1637.7327.7530.5913.5618.4719.5572.6685.2719.4429.6262.2947.5838.214.00×
Trajectory-MBR38.3343.3641.8242.4623.9021.8720.0572.5089.3118.0040.5063.8663.3844.561.94×

Adaptive stopping scores higher at a similar amount of computation. On Dream-7B-Instruct, fixed-budget early exit averages 39.90 at 2.00×, which is 3.13 points below full decoding. Trajectory-MBR reaches 44.56 at a nearly identical 1.94×, which is 1.53 points above full decoding. On UltraLLaDA, fixed-budget early exit reaches only 45.59 even at 2.92×, where it uses more computation than Trajectory-MBR, and this is below the 48.37 of Trajectory-MBR at 3.76×.

A fixed budget collapses on certain tasks, whereas Trajectory-MBR does not. Fixed-budget early exit drops from 96.75 to 85.12 on PassageRetrieval-en with UltraLLaDA and from 62.23 to 50.65 on RepoBench-P with Dream-7B-Instruct, because different examples need different numbers of steps. In contrast, Trajectory-MBR matches or exceeds full decoding on all 13 tasks with UltraLLaDA and on 12 of the 13 tasks with Dream-7B-Instruct. Because it stops only when the candidates agree sufficiently, it saves computation on examples where early stopping is safe and continues decoding on the others.

Exploiting early responses requires deciding the stopping point for each example. On UltraLLaDA, fixed-budget early exit alone already raises the QA scores. On this backbone, responses at early checkpoints are often better than the terminal response, which is consistent with the per-example analysis in Q3. However, a fixed budget falls short of Trajectory-MBR on average because of the losses on other tasks. This benefit is obtained without such losses only when the stopping point is decided for each example.

The same-budget comparison was added to Table 5 in Section 6.2 of the revised manuscript, and the 13-task comparison was added to Table 11 in Appendix C.

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

The paper proposes Trajectory-MBR, a training-free inference method for masked diffusion LMs. At each denoising,the masks with top-1 tokens are filled, and scores that candidate with a trajectory score. Then over a sliding window of recent candidates run score-weighted MBR. The process stops early when the risk is below a threshold and minimiser is returned. They report gains on many LongBench tasks and multimodal benchmarks, and a token-level variant on top of JoT.

Critical strengths
  • The method is efficient. It reuses denoiser outputs, and divides the idea of when to stop and which checkpoint to return.
  • The theory is also upfront about its limits, saying the bounds aren't online certificates.
  • The toy experiment shows the score tracks exact likelihood in a small setting.
Critical weaknesses

Gains mainly come from per-task hyperparameters picked jointly on the evaluation data, and the sensitivity sweep is on the same 200 LSHT test examples that get reported, with the default being the best cell. There's no held-out split or seeds, and many tasks only small examples. So it is hard tell selection noise from real gains. There are also no ablations such as unweighted MBR, early stopping only, terminal-step selection, simple confidence stopping.

The full-decoding baselines look degraded. UltraLLaDA gets 16.27 on HotpotQA while Dream-7B gets 41.59 under the same protocol, and LaViDa full decoding gets 8.3 on DetailCaps when other baselines get 20 to 33. It seems like a lot of the win is just reverting a bad terminal state, which isn't the same as a general quality improvement.

Questions for authors
  1. Could you report results using one fixed setting of K, epsilon, and the utility function for each benchmark family?

  2. Could you add some ablations? E.g., MBR without weights, early stopping that returns the current candidate, choosing only from the final outputs, and simple confidence based stopping.

  3. Why is full decoding so low for UltraLLaDA on HotpotQA/2Wiki and for LaViDa on DetailCaps?? Did you rerun these baselines with the same setting?

8 comments

We sincerely thank the reviewer for the careful reading and the constructive feedback. We address each point below.

W1-1. Experimental setup

We thank the reviewer for this careful observation. All experiments follow the evaluation protocols of the original papers that report the baseline numbers. Table 1 and the related language experiments follow the protocol of Focus-dLLM (Long et al., 2026) with a fixed seed of 2025. Table 2 and the related multimodal experiments follow the protocol of MPD-PAC (Hong et al., 2026) with a fixed seed of 0, except for the step budgets and the judge model documented in Appendix B.2. The token-level experiments in Figure 3 follow the protocol of JoT (Kohut et al., 2026) with a fixed seed of 0.

As a reproducibility check, we re-ran the fixed configuration on all 200 Qasper examples with UltraLLaDA using two additional seeds, 2024 and 2026. The two runs produce identical generation trajectories and identical selected responses on all 200 examples. Compared with a stored run that uses the original seed of 2025, the final full-decoding responses are identical on all 200 examples, and the intermediate candidates are identical on 198 of them. The results in this setting are therefore stable across seeds.

We address the concerns about per-task configuration and small sample sizes in three ways.

  • A single fixed configuration. In W1-2, we apply one configuration unchanged to all 13 LongBench tasks and to all nine multimodal benchmarks. It exceeds full decoding on average on LongBench and stays at the level of full decoding on the multimodal benchmarks.
  • Larger and broader evaluations. The component ablations in Section 6.2 add LCC with 500 examples to LSHT, and Appendix C extends the fixed-budget comparison to all 13 tasks on two backbones.
  • Per-example consistency. The gains on small tasks are not driven by a few examples. On the 200 HotpotQA examples with UltraLLaDA, Trajectory-MBR scores higher than full decoding on 95 examples and lower on 18, which gives a two-sided sign test p-value below 10−1310^{-13}. On 2WikiMQA, the counts are 81 and 29, with a p-value below 10−610^{-6}.

W1-2, Q1. One fixed KK, ϵ\epsilon, and utility per benchmark family

We applied a stricter test than one fixed setting per benchmark family. A single configuration is fixed and applied to all tasks. The risk threshold is fixed to ϵ=0\epsilon = 0, the task utility to Token F1, and the window size to K=5K = 5. ϵ=0\epsilon = 0 is the most conservative choice, which stops only when the candidates that disagree with the selected one carry negligible total weight. Token F1 is a general utility that applies to every task regardless of its format, and K=5K = 5 is a moderate window that is neither too short nor too long.

Language tasks. We applied the same configuration to all 13 LongBench tasks on UltraLLaDA.

Table R1. Results with a single fixed configuration on all 13 LongBench tasks with UltraLLaDA. † denotes results reported by Long et al. (2026), and the experimental settings follow that work. The fixed setting has a higher average than the per-task setting and all baselines, which shows that the gain does not depend on per-task tuning.

MethodQasperMF-enHotpotQA2WikiMQAMuSiQueGovReportQMSumTRECTriviaQALSHTPReLCCRB-PAvg. ScoreAvg. Speedup
Full decoding†19.1425.8716.2718.0012.0832.8322.4880.0091.5841.0096.7568.2359.5044.901.00×
Fast-dLLM†18.3429.9017.0317.1113.3630.0522.8979.5091.0342.0094.7567.5058.1044.741.70×
Sparse-dLLM†18.0427.2620.5917.8813.6729.9523.5776.5091.9341.5097.1267.5057.7244.861.00×
SparseD†19.0925.8715.4518.0411.9232.6422.5079.5090.7041.5096.7968.1059.0244.701.00×
Focus-dLLM†17.0229.1122.4721.4920.2026.7521.4577.0090.7841.0095.7366.7257.1445.141.00×
Trajectory-MBR, per-task setting23.4134.2528.0327.7620.3032.8822.9180.0091.5842.5097.3368.3059.5048.373.76×
Trajectory-MBR, fixed setting22.9042.8028.6225.7622.0431.4422.9080.0091.5841.0098.4267.8159.2048.811.94×

Multimodal tasks. We applied one fixed configuration in the same way to the nine multimodal benchmarks on LLaDA-V.

Table R1-2. Results with a single fixed configuration on the nine multimodal benchmarks with LLaDA-V. †† denotes results reported by Hong et al. (2026) except for the re-scored columns, and the experimental settings follow that work. All rows other than the fixed setting are identical to Table 2 of the revised manuscript. Because gpt-4-0613 was retired on October 23, 2026, LLaVA-Bench and Ferret are re-scored for all methods with gpt-4o-2024-08-06, and Table 2 is updated accordingly. The fixed setting keeps the average at the level of full decoding with a 1.48× speedup, which shows that quality is preserved without per-task tuning.

MethodMMEMMBenchMMMURefCOCOgFerretGQALLaVA-BenchDetailCapsMIAAvg. Score w/o MMEAvg. Speedup
Full decoding††199383.148.770.785.261.661.959.866.167.141.00×
Fast-dLLM199983.049.668.247.361.635.258.653.457.111.00×
Sparse-dLLM200383.049.468.685.762.053.757.873.566.711.00×
SparseD192880.448.870.289.762.256.857.974.667.581.01×
Trajectory-MBR, per-task setting200683.249.371.088.361.661.960.072.668.491.97×
Trajectory-MBR, fixed setting199783.049.268.685.761.461.960.072.467.781.48×

Token-level extension. For the token-level experiments, instead of a single configuration, Figure 3 of the revised manuscript reports the entire speed and quality curve by varying the JoT gate threshold, which controls the speed, from 30 to 300. Because this shows the full change as the setting varies rather than one particular setting, we did not run a separate fixed-configuration experiment. On every benchmark and both backbones, the curve contains accelerated operating points at which Trajectory-Token-MBR matches or exceeds full decoding. For example, on HumanEval with Dream-7B-Instruct, it stays at or above full decoding up to about 12×. The baselines fall below full decoding earlier on most benchmarks. On GSM8K with Dream-7B-Instruct, KLASS and Prophet are below full decoding before 3×, while Trajectory-Token-MBR is still above it at 2.40×.

A single configuration preserves the average gain. On the language tasks, the fixed setting averages 48.81, which is higher than the 48.37 of the per-task setting and the 45.14 of Focus-dLLM, the strongest baseline. It matches or exceeds full decoding on 10 of the 13 tasks. On the multimodal benchmarks, the fixed setting averages 67.78, which is at the level of the 67.14 of full decoding, together with a 1.48× speedup. Therefore the average gain in Table 1 does not originate from per-task hyperparameter selection, and on the multimodal benchmarks quality is preserved without it.

The per-task setting mainly increases the speed. On the language tasks, the fixed and per-task settings have similar average scores, but the speedup rises from 1.94× to 3.76×, and on the multimodal benchmarks it rises from 1.48× to 1.97×. In other words, a single configuration already yields the quality gain, and the per-task setting obtains the same quality with less computation.

On the sensitivity analysis in Table 3. Table 3 is a post hoc sensitivity analysis that varies one factor at a time on the same 200 examples, and it is not the basis for choosing the setting. The default setting is also not an isolated peak. K=6K = 6, ϵ=0.01\epsilon = 0.01, and the BLEU utility reach the same 42.50, so performance is stable in a wide region around the default. Under the fixed setting, LSHT scores 41.00, the same as full decoding, so we agree that it is accurate to read the 42.50 on LSHT as a result of the per-task setting. However, the LSHT score does not drop under the fixed setting, and the per-task setting adds a gain on top of it.

Reviewer_k8mwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Trajectory-MBR, a training-free inference method for masked diffusion language models. Its contribution is mainly determine when DLLM decoding should stop and which previously observed candidate should be returned. At each checkpoint, the method fills the remaining masks with current top-1 predictions, keeps a sliding window of completed checkpoint candidates. Then it assigns trajectory-derived weights to those candidates and applies minimum Bayes risk selection within the window. Decoding stops when the selected candidate’s weighted risk falls below a threshold.

The proposed method is evaluated on LongBench tasks, multimodal benchmarks, and a token-level extension on GSM8K, MMLU, HellaSwag, and HumanEval. The proposed method outperforms full decoding with prominent speedup gain, showing its effectiveness and efficiency.

Critical strengths
  1. The stopping-versus-selection distinction is important and well motivated. A model may encounter its best answer before the final stopping checkpoint, so returning only the current state can discard useful trajectory information.

  2. The method is training-free and comparatively easy to attach to existing samplers. It does not require changing the denoiser or retraining the underlying model, and candidate construction can reuse already-computed checkpoint outputs under the stated implementation conditions

  3. The empirical coverage is broad. The paper evaluates long-context language tasks, multimodal understanding and generation, and a token-level extension, with two backbones in each major setting

Critical weaknesses
  1. The hyper- parameter selection seems very critical in the proposed method. The search grid shown in Table 8 in the appendix is complicated which arouse my concern on the robustness of the proposed method. I am now not convinced that this method can be easily applied without careful hyper-parameter tuning.

  2. The actual speedup is not yet clear. The authors state that step-based speedup counts completed transitions and can differ from the number of denoiser forward passes.

  3. The performance in Table2 is not yet clear. MME majorly contributed to the average score, which makes the avg score less convincing.

  4. Core component ablations are missing. The paper should compare trajectory weighting against uniform weights, current-checkpoint confidence, unweighted MBR, majority voting, deduplicated candidate windows, and retrospective selection without early stopping. Because repeated identical candidates are retained and accumulate support, a deduplication ablation is especially important.

Questions for authors
  1. Can you report a single globally tuned configuration, or configurations selected on separate development tasks, and apply them unchanged to all reported benchmarks?
  2. How does the HumanEval 12.34× step speedup translate into denoiser forward passes and wall-clock latency?
  3. How much of the gain comes from trajectory weighting, retrospective candidate selection, and early stopping individually?

See also the weakness part.

7 comments

We sincerely thank the reviewer for the careful reading and the constructive feedback. We address each point below.

W1, Q1. A single configuration applied unchanged to all benchmarks

Following the concern about tuning dependence raised by the search space in Table 8, which is Table 10 in the revised manuscript, we applied a single configuration unchanged to the language benchmarks on UltraLLaDA and the multimodal benchmarks on LLaDA-V. The risk threshold is fixed to ϵ≈0\epsilon \approx 0, the task utility to Token F1, and the window size to K=5K = 5. ϵ≈0\epsilon \approx 0 is the most conservative choice, which stops only when the candidates that disagree with the selected one carry negligible total weight. Token F1 is a general utility that applies to every task regardless of its format, and K=5K = 5 is a moderate window that is neither too short nor too long.

Language tasks. We applied the same configuration to all 13 LongBench tasks on UltraLLaDA.

Table R1. Results with a single fixed configuration on all 13 LongBench tasks with UltraLLaDA. † denotes results reported by Long et al. (2026), and the experimental settings follow that work. The fixed setting has a higher average than the per-task setting and all baselines, which shows that the gain does not depend on per-task tuning.

MethodQasperMF-enHotpotQA2WikiMQAMuSiQueGovReportQMSumTRECTriviaQALSHTPReLCCRB-PAvg. ScoreAvg. Speedup
Full decoding†19.1425.8716.2718.0012.0832.8322.4880.0091.5841.0096.7568.2359.5044.901.00×
Fast-dLLM†18.3429.9017.0317.1113.3630.0522.8979.5091.0342.0094.7567.5058.1044.741.70×
Sparse-dLLM†18.0427.2620.5917.8813.6729.9523.5776.5091.9341.5097.1267.5057.7244.861.00×
SparseD†19.0925.8715.4518.0411.9232.6422.5079.5090.7041.5096.7968.1059.0244.701.00×
Focus-dLLM†17.0229.1122.4721.4920.2026.7521.4577.0090.7841.0095.7366.7257.1445.141.00×
Trajectory-MBR, per-task setting23.4134.2528.0327.7620.3032.8822.9180.0091.5842.5097.3368.3059.5048.373.76×
Trajectory-MBR, fixed setting22.9042.8028.6225.7622.0431.4422.9080.0091.5841.0098.4267.8159.2048.811.94×

Multimodal tasks. We applied one fixed configuration in the same way to the nine multimodal benchmarks on LLaDA-V.

Table R1-2. Results with a single fixed configuration on the nine multimodal benchmarks with LLaDA-V. †† denotes results reported by Hong et al. (2026) except for the re-scored columns, and the experimental settings follow that work. All rows other than the fixed setting are identical to Table 2 of the revised manuscript. Because gpt-4-0613 was retired on October 23, 2026, LLaVA-Bench and Ferret are re-scored for all methods with gpt-4o-2024-08-06, and Table 2 is updated accordingly. The fixed setting keeps the average at the level of full decoding with a 1.48× speedup, which shows that quality is preserved without per-task tuning.

MethodMMEMMBenchMMMURefCOCOgFerretGQALLaVA-BenchDetailCapsMIAAvg. Score w/o MMEAvg. Speedup
Full decoding††199383.148.770.785.261.661.959.866.167.141.00×
Fast-dLLM199983.049.668.247.361.635.258.653.457.111.00×
Sparse-dLLM200383.049.468.685.762.053.757.873.566.711.00×
SparseD192880.448.870.289.762.256.857.974.667.581.01×
Trajectory-MBR, per-task setting200683.249.371.088.361.661.960.072.668.491.97×
Trajectory-MBR, fixed setting199783.049.268.685.761.461.960.072.467.781.48×

Token-level extension. For the token-level experiments, instead of a single configuration, Figure 3 of the revised manuscript reports the entire speed and quality curve by varying the JoT gate threshold, which controls the speed, from 30 to 300. Because this shows the full change as the setting varies rather than one particular setting, we did not run a separate fixed-configuration experiment. On every benchmark and both backbones, the curve contains accelerated operating points at which Trajectory-Token-MBR matches or exceeds full decoding. For example, on HumanEval with Dream-7B-Instruct, it stays at or above full decoding up to about 12×. The baselines fall below full decoding earlier on most benchmarks. On GSM8K with Dream-7B-Instruct, KLASS and Prophet are below full decoding before 3×, while Trajectory-Token-MBR is still above it at 2.40×.

A single configuration preserves the average gain. On the language tasks, the fixed setting averages 48.81, which is higher than the 48.37 of the per-task setting and 3.67 points above the 45.14 of Focus-dLLM, the strongest baseline. It matches or exceeds full decoding on 10 of the 13 tasks. On the multimodal benchmarks, the fixed setting averages 67.78, which is at the level of the 67.14 of full decoding, together with a 1.48× speedup. Therefore the average gain of our method on the language tasks does not come from the per-task search in Table 10, and on the multimodal benchmarks quality is preserved without it.

The per-task setting mainly increases the speed. On the language tasks, the fixed and per-task settings have similar average scores, but the speedup rises from 1.94× to 3.76×, and on the multimodal benchmarks it rises from 1.48× to 1.97×. In other words, a single configuration already yields the quality gain, and the per-task search is an optional step that obtains the same quality with less computation. To apply the method directly without any tuning, the fixed configuration above can be used as is.

The same results are also reported in our response to W1-2 and Q1 of Reviewer yZKe.

W2, Q2. HumanEval step speedup in denoiser evaluations and wall-clock latency

We thank the reviewer for raising this important point. We clarify below how the step-based speedup translates into denoiser evaluations, and we additionally report the measured wall-clock latency.

From steps to denoiser evaluations. All 164 HumanEval problems use a step budget of T=512T = 512. The step-based speedup of 12.34× therefore corresponds to an average of 41.49 completed transitions per problem. As described in Appendix B.4, full decoding uses 512 denoiser evaluations for its 512 transitions, since no evaluation is performed at the terminal state x0x_0. Trajectory-Token-MBR performs an evaluation at both the initial state and every reached state, so it requires one more evaluation than the number of completed transitions, which gives an average of 42.49 evaluations per problem. The resulting denoiser-evaluation speedup is 12.05×.

Wall-clock latency. We measured the wall-clock latency over all 164 HumanEval problems on a single NVIDIA A100 GPU, under same conditions as the LSHT measurements in Appendix C.

Table R5. Three speedup metrics of Trajectory-Token-MBR on HumanEval with Dream-7B-Instruct. Measurement settings follow the LSHT measurements in Appendix C. The step-based speedup translates into actual reductions in both denoiser evaluations and wall-clock time.

MetricSpeedup over full decoding
Step-based speedup12.34×
Denoiser-evaluation speedup12.05×
Wall-clock speedup10.77×

W3. Table 2 average without MME or with normalization

We thank the reviewer for this helpful comment. We agree that MME is a summed score with a maximum of 2,800, which puts it on a very different scale from the other benchmarks and lets it dominate the average. To address this, we now report two alternative averages. The first is the mean over the eight benchmarks excluding MME. The second is the mean over all nine benchmarks after dividing MME by 2,800 to put it on a 100-point scale.

Table R6. Averages of Table 2 recomputed in two ways. The per-benchmark numbers are identical to those in the revised Table 2, in which LLaVA-Bench and Ferret are re-scored for all methods with gpt-4o-2024-08-06 because gpt-4-0613 was retired on October 23, 2026. Under both schemes, Trajectory-MBR achieves the highest average on both backbones, which shows that our conclusion does not depend on how MME is aggregated.

MethodLLaDA-V, w/o MMELLaDA-V, normalized MMELaViDa, w/o MMELaViDa, normalized MME
Full decoding67.1467.5945.8847.49
Fast-dLLM57.1158.7041.9444.02
Sparse-dLLM66.7167.2548.3149.73
SparseD67.5867.7246.7848.17
Trajectory-MBR68.4968.8449.3550.64

Without MME, Trajectory-MBR exceeds full decoding by 1.35 and 3.47 points on LLaDA-V and LaViDa, and the strongest baseline by 0.91 and 1.04 points. It also matches or exceeds full decoding on all nine benchmarks for both backbones. We have updated the average column of Table 2 to the MME-excluded average and added a note to the caption, highlighted in blue.

W4, Q3. Component ablations, especially deduplication

We performed all the requested ablations and added them to Table 4 and Table 5 in Section 6.2 of the revised manuscript, highlighted in blue. On the 200 LSHT examples and the 500 LCC examples with UltraLLaDA, we fix the settings of Table 1 and change one component at a time. Table 4 removes or replaces components of Trajectory-MBR, and Table 5 replaces it with simpler stopping and selection rules.

Mapping between the requested items and the revised manuscript. Speedup is measured in denoiser evaluations relative to full decoding. Full decoding is re-evaluated on the same stored trajectories. Tuned and matched denote settings in which ϵ\epsilon, τ\tau, or the budget is adjusted so that the speedup approaches that of Trajectory-MBR.

Requested itemMethod and location in the revised manuscriptLSHT ScoreLSHT SpeedupLCC ScoreLCC Speedup
Full decoding41.001.00×68.081.00×
Trajectory-MBR42.5012.94×68.301.21×
Uniform weights, unweighted MBRUniform weights, Default, Table 441.505.13×68.071.15×
Uniform weights, Tuned, Table 442.509.57×68.031.21×
Current-checkpoint confidenceCompletion confidence, Default, Table 441.505.13×68.071.15×
Completion confidence, Tuned, Table 441.0012.80×68.031.21×
Confidence stopping, matched τ\tau, Table 541.5012.94×68.081.21×
Confidence argmax, matched budget, Table 537.5012.80×68.051.21×
Majority votingMajority voting, all 64 checkpoints, Table 541.001.00×68.071.00×
Majority voting, matched budget, Table 541.0012.80×68.041.21×
Deduplicated candidate windowDeduplicated, Table 442.5012.92×68.281.21×
Retrospective selection without early stoppingLast KK checkpoints, Table 441.001.00×68.081.00×
All 64 checkpoints, Table 441.001.00×68.081.00×
Early stopping onlyCurrent candidate, Table 442.0012.94×68.181.21×

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank Reviewers yZKe, k8mw, and ngZ8 for their thorough and constructive feedback. We summarize the strengths noted by the reviewers and the main new results. Individual comments are addressed in the separate responses.

1. Strengths Highlighted by the Reviewers

  1. [yZKe, k8mw, ngZ8] Well-Motivated Design. Decoupling when to stop from which checkpoint to return preserves good early answers.
  2. [yZKe, k8mw, ngZ8] Computational Efficiency. Reusing existing denoiser outputs gives clear speedups over full decoding.
  3. [k8mw] Training-Free and Easy to Integrate. The method attaches to existing samplers without retraining.
  4. [k8mw] Broad Empirical Coverage. The evaluation covers language, multimodal, and token-level settings.
  5. [yZKe, ngZ8] Trajectory Scoring and Transparent Theory. The scoring scheme is a unique contribution, and the theory is upfront about its limits.

2. Summary of New Results and Revisions

  1. [yZKe, k8mw] Robustness to hyperparameters. One fixed configuration averages 48.81 on all 13 LongBench tasks against 44.90 for full decoding, and stays at the level of full decoding on nine multimodal benchmarks with a 1.48× speedup. See Tables R1 and R1-2.
  2. [yZKe, k8mw, ngZ8] Component ablations. Trajectory weighting, retrospective selection, and early stopping are isolated against simpler weights and stopping rules. See Tables 4 and 5.
  3. [ngZ8] Mechanism of trajectory weighting. Trajectory weights approximate the candidate likelihood and stop earlier, needing 4.95 evaluations on LSHT against 12.48 with uniform weights.
  4. [ngZ8] Fixed-budget early exit. At comparable budgets it falls below Trajectory-MBR, for example 39.90 against 44.56 on Dream-7B-Instruct. See Table 11.
  5. [yZKe, ngZ8] Selection behavior analysis. Improved examples outnumber degraded ones 95 to 18 on HotpotQA and 81 to 29 on 2WikiMQA, with sign test p-values below 10−1310^{-13} and 10−610^{-6}. See Table R7.
  6. [yZKe] Baseline reproduction and setup. Re-running full decoding reproduces the low scores within 1 point, and we explain their causes. See Table R3.
  7. [k8mw] Efficiency and Table 2 averages. The HumanEval speedup is 12.34× in steps, 12.05× in denoiser evaluations, and 10.77× in wall-clock time. Trajectory-MBR has the highest Table 2 average both without MME and with normalized MME. See Tables R5 and R6.
  8. [yZKe, k8mw] Re-scored judge columns in Table 2. Because gpt-4-0613 was retired on October 23, 2026, LLaVA-Bench and Ferret are re-scored for all methods with gpt-4o-2024-08-06, and Trajectory-MBR keeps the highest average on both backbones.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.