acceptodds
Under review as a conference paper at ICLR 2027

Learning a Reusable Token-Commitment Policy for Diffusion Language Models

Abstract

Diffusion language models refine many token positions in parallel, but efficient decoding requires deciding when to commit proposed tokens. We ask whether token commitment can be learned as a reusable policy over diffusion traces, rather than calibrated separately for each decoding configuration. We introduce TraceLock, a lightweight controller for a frozen diffusion backbone. An intermediate token proposal is labeled future-stable when it matches the final token of its completed trace; a shared contextual scorer learns from these labels on variable-length trace states. A single TraceLock checkpoint per backbone is deployed across the tested window widths and generation lengths without collecting new traces or retraining. Experiments on mathematical reasoning, question answering, and code generation show an improved quality–step tradeoff over heuristic and learned baselines and greater stability than Learn2PD when the decoding window changes. We use one global acceptance cutoff across tasks and configurations, avoiding per-setting calibration; changing that cutoff provides a controllable quality–step tradeoff. These results support treating token commitment as a reusable learned policy over evolving diffusion traces.

Primary area:
generative models
Keywords:
Discrete diffusion language models, Decoding policy

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_8GbyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TRACELOCK, a learned token-commitment controller for frozen diffusion language models. Trained using agreement between intermediate proposals and final trace tokens, it combines contextual hidden-state features, a dynamic threshold, and a moving local window to determine which tokens to commit. Experiments evaluate quality–step tradeoffs and reuse across decoding configurations on LLaDA and Dream.

Critical strengths
  1. The method addresses an important problem: improving diffusion decoding without modifying the backbone.
  2. Evaluation covers mathematical reasoning, question answering, and code generation on two backbones.
  3. Cross-configuration experiments and single-token interventions provide useful evidence about controller reuse and commitment behavior.
Critical weaknesses
  1. Limited novelty over Learn2PD. Both methods learn token acceptance from agreement with final decoded tokens.
  2. Unclear efficiency objective. The paper should distinguish whether “efficiency” refers to fewer decoding steps, lower wall-clock latency, or improved quality at a fixed computational budget. These objectives are not equivalent, and the reported results do not consistently improve both quality and speed.
  3. Unclear acceleration at matched quality. TRACELOCK sometimes requires more steps and higher measured latency than existing accelerated decoders, particularly on LLaDA. Since quality also varies with decoding-window size, matched-quality latency comparisons and quality–latency curves across window sizes are needed to establish a better tradeoff.
  4. Missing relevant strong baselines such as CCD, DAPD, SlowFast.
  5. Limited generalization evidence. Generalization is demonstrated primarily across decoding configurations within the same backbone and related task families. Robustness to unseen domains and deployment-induced trajectory distribution shift remains unclear.
  6. Insufficient methodological details. Lines 216–222 do not adequately explain how high-quality training traces are obtained, including the decoding schedules and filtering criteria. Although Appendix K reports 8,000 traces per backbone, task proportions and filtering retention rates remain unspecified.

Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles

DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs

Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models

Questions for authors

NAN

6 comments
AuthorsOpenReview ↗

Response 8Gby part 1

Thank you for your review and for recognising the importance of the problem and the breadth of our experiments. We have conducted additional experiments, including comparisons with the baselines you suggested, and address your concerns below.

Novelty over Learn2PD

Although both methods learn from agreement with final decoded tokens, they learn different commitment policies. Learn2PD uses a fixed-width, confidence-based filter. TraceLock instead uses contextual hidden-state features and a dynamic threshold conditioned on the current decoding state. This allows both token scores and the acceptance boundary to adapt throughout decoding, drawing on information beyond scalar token confidence. Figure 8 illustrates how TraceLock’s scores differ from token confidence.

The practical consequence is a reusable controller: TraceLock accommodates changes in decoding window and generation length using a single checkpoint per backbone. Under Learn2PD’s standard training protocol, changing the filter’s block size requires collecting new traces and retraining. Table 2 evaluates reuse without retraining when the decoding window is enlarged, comparing TraceLock with the original Learn2PD filter applied independently to smaller subwindows. Figure 5 further demonstrates reuse across generation lengths and window sizes on HumanEval.

Appendix Table 12 evaluates L2P-Hidden, which retains the Learn2PD-style blockwise interface but replaces confidence inputs with concatenated hidden-state features. On HumanEval with generation length 128 and block size 64, this substitution improves Pass@1 from 32.9% to 46.9% on Dream and from 24.4% to 28.1% on LLaDA. The full TraceLock system achieves 51.2% and 32.32%, respectively. These results support the value of hidden-state information, while showing that feature substitution alone does not match the full TraceLock system in these settings.

Efficiency definition and quality–latency tradeoff

Our objective is to improve the quality–latency Pareto frontier, with particular emphasis on achieving high task accuracy while controlling generation time. We therefore assess task accuracy alongside measured wall-clock latency, with decoding steps reported as a complementary measure.

We have added Figure 3 to the manuscript, comparing TraceLock, Learn2PD, and Fast-dLLM across acceptance thresholds and window sizes 32 and 64 on GSM8K, with generation length 256.

On Dream, TraceLock achieves the highest observed accuracy, reaching 80.80% in 7.73 seconds per sample. At window 64, every evaluated TraceLock configuration achieves higher accuracy and lower latency than every evaluated Learn2PD configuration. TraceLock also achieves 79.00% accuracy in 6.42 seconds, compared with Fast-dLLM’s highest observed accuracy of 76.42% in 6.56 seconds.

On LLaDA, TraceLock at window 32 achieves 79.33% accuracy in 7.39 seconds, exceeding the highest observed accuracy of both Fast-dLLM (78.70% in 8.00 seconds) and Learn2PD (78.30% in 7.44 seconds) across the two windows, at lower measured latency. At window 64, TraceLock further increases accuracy to 80.00% in 8.09 seconds.

These results demonstrate that TraceLock improves the observed quality–latency tradeoff in the higher-accuracy region on both backbones, surpassing the best accuracy achieved by Learn2PD and Fast-dLLM at lower measured latency, while also reaching the highest overall accuracy among the evaluated configurations.

Additional baselines

We added comparisons with the suggested baselines and additional methods: DAPD-Direct, CCD, CCD-DS, SlowFast, and Apple’s learned unmasking policy. All comparisons use Dream on GSM8K, with generation length 256 and window size 64. We report accuracy, average decoding steps, and generation time per sample below. Results are compared within each experimental setup.

MethodAccuracy (%) ↑Average Decoding Steps ↓Generation Time (s/sample) ↓
TraceLock80.80106.109.76
Fast-dLLM79.90197.4014.74
DAPD-Direct77.9490.988.55
CCD79.38256.0018.29
CCD-DS64.59101.008.17
SlowFast77.7182.737.83
Learned unmasking policy (Apple)69.9862.8410.43

TraceLock achieves the highest observed accuracy (80.80%), with 33.8% and 46.6% lower latency than Fast-dLLM and CCD, respectively. DAPD-Direct, CCD-DS, and SlowFast are faster but less accurate. Apple’s learned unmasking policy uses the fewest decoding steps, but has lower accuracy and higher measured latency than TraceLock.

AuthorsOpenReview ↗

Response 8Gby part 2

Robustness to commitment-induced trajectory shift

Appendix C.1 directly tests commitment-induced trajectory shifts through counterfactual interventions. From the same decoding state, we compare native generation with a continuation that additionally commits one TraceLock-selected token. This intervention changes a correct answer to an incorrect one in only 2/367 cases on LLaDA and 2/323 on Dream. Together with our full-rollout evaluations, these results support robustness to commitment-induced shifts under the evaluated conditions.

Training-trace construction

For each frozen backbone, we collect 8,000 retained complete decoding trajectories using its native confidence-based decoder with generation-length/window-size pairs of (256, 64), (192, 32), and (128, 32). We apply rejection-sampling-style filtering to the completed trajectories. For coding, we retain only trajectories whose generated programs pass the supplied test cases, rejecting 38% of candidates. For QA, we use LanguageTool-assisted grammatical and formatting checks to filter malformed or repetitive outputs, rejecting 9% of candidates. For math, we reject trajectories with repetitive outputs or without an extractable final answer in \boxed{} format. The final retained training set contains coding, math and QA trajectories in a 3:1:6 ratio.


We hope these additional experiments and clarifications address your concerns and clarify TraceLock’s contribution. We would be happy to answer any further questions or discuss any points that remain unclear.

Dear Authors,

Thank you for the response. The added baselines, the efficiency objective, the quality–latency results, and the training-trace details address my earlier questions.

My remaining concern is the method itself. TraceLock trains a small controller to decide whether to keep the frozen decoder’s proposed token, using whether that token survives to the end of the decoder’s own trace. This needs extra trace collection and training for each backbone. I did not see generalization beyond decoding settings on the same backbone, so maintaining such a controller in real use is less convincing if that generalization is weak. The accept-or-reject rule is also close to speculative decoding, and only deciding whether to keep the current token may not be the best policy. Better ways to choose the committed token are still worth studying. So I remain my scores.

Reviewer_8TESOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper formulates fast decoding for frozen DLLMs as a token-commitment problem: deciding at each denoising step which active token proposals are sufficiently stable to lock in irreversibly. Rather than relying on hand-designed confidence thresholds or block-size-specific MLP filters. The authors introduce TraceLock, a lightweight, reusable Transformer-based controller. Experiments show that there is improved quality-step tradeoff on multiple datasets.

Critical strengths
  1. Reusability across different decoding regimes over evolving diffusion traces is important. Learn2PD is tied to a fixed block size and from Table 2 and Figure 4 Learn2PD drops a lot when the window doubles but TraceLock remains stable.
Critical weaknesses
  1. In Section 3.2, the training objective is defined via self-supervised future-stability labels collected from frozen base model. This supervision is purely observational, and during TraceLock training, the controller learns from trajectories generated under a fixed, fine-grained block structure. When deploying, locking a subset of tokens alters the bidirectional conditioning for all later denoising steps. So this distribution shift is not really theoretically bound.
  2. The efficiency gains are highly backbone-dependent. On Dream, TraceLock reduces the number of denoising steps while improving accuracy. On LLaDA, it seems that TraceLock requires more denoising steps than both fast-dllm and Learn2PD.
Questions for authors
  1. Please check the weakness above.
3 comments
AuthorsOpenReview ↗

Response to 8TES

Thank you for your valuable review and for recognising the clarity and soundness of our work, as well as the importance of TraceLock’s reusability across decoding regimes compared with previous methods. We address your concerns about commitment-induced distribution shift and backbone-dependent efficiency below.

Future-stability and distribution shift

We agree that observational future-stability labels do not provide a theoretical guarantee under the controller-induced state distribution. We therefore evaluate the effect of commitment empirically. Appendix C.1 examines the effect of commitment through counterfactual interventions. From the same decoding state, we compare native generation with a continuation that additionally commits one TraceLock-selected token. Among cases where the native continuation answers correctly, this intervention changes the answer to incorrect only in 2/367 cases on LLaDA and 2/323 on Dream.

We also analyse full TraceLock-controlled trajectories. Across 1,300 prompts, the proportion of future-stable proposals among still-revisable positions changes relative to native decoding by just +4.04/+3.49/−2.41 percentage points for LLaDA and +1.87/+2.87/−2.50 points for Dream at 25/50/75% decoding progress. These complementary analyses provide empirical evidence that individual commitments rarely turn correct answers into incorrect ones in the tested cases, and that future-stable proposals remain similarly prevalent after repeated commitments.

Backbone-dependent quality–latency tradeoff

Our objective is to improve the quality–latency tradeoff, with particular emphasis on achieving high task accuracy while controlling generation time. We therefore assess task accuracy alongside measured wall-clock latency, with decoding steps as a complementary measure. In the revised manuscript, Figure 3 now compares TraceLock, Learn2PD, and Fast-dLLM across acceptance thresholds and window sizes 32 and 64 on GSM8K, with generation length 256.

On Dream, TraceLock achieves the highest observed accuracy, reaching 80.80% in 7.73 seconds per sample. At window 64, every evaluated TraceLock configuration achieves higher accuracy and lower latency than every evaluated Learn2PD configuration. TraceLock also achieves 79.00% accuracy in 6.42 seconds, compared with Fast-dLLM’s highest observed accuracy of 76.42% in 6.56 seconds.

On LLaDA, TraceLock at window 32 achieves 79.33% accuracy in 7.39 seconds, exceeding the highest observed accuracy of both Fast-dLLM (78.70% in 8.00 seconds) and Learn2PD (78.30% in 7.44 seconds) across the two windows, at lower measured latency. At window 64, TraceLock further increases accuracy to 80.00% in 8.09 seconds.

These results show TraceLock’s advantage in the higher-accuracy region on both backbones, while Fast-dLLM offers faster configurations at lower accuracy. They also demonstrate why decoding steps alone are insufficient to assess practical acceleration.


We hope these analyses address your concerns about commitment-induced distribution shift and clarify TraceLock’s quality–latency tradeoffs across backbones. We would be happy to discuss any remaining questions.

I appreciate the author's response. I am keeping my score.

Thank you for your time and thoughtful review.

Reviewer_6nBPOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method for adaptive decoding in masked diffusion. The idea is that for a given masked token, we look at the prediction for it, whichever value has the maximum probability is considered the candidate token, a model is trained to predict when a candidate token is already the token we will see at the end of generation. By doing so, one is able to unmask tokens when there is already high certainty for that particular token. Notably, this method is self-supervised which makes the training easy to implement. The authors demonstrate their method on diverse tasks proving its generability.

Critical strengths
  • The paper is easy to read and well motivated
  • The paper tests on important reasoning tasks including coding and math
  • The paper performs ablations and demonstrates that the design is necessary to achieve all the benefits
Critical weaknesses
  • The paper only considers as "candidate tokens" the argmax. Although the greedy decoding approach may seem reasonable, it is known that this has many issues and does not recover the data distribution. There are also other sampling strategies like sampling, top-k, etc that yield good results. The problem formulation is restrictive and does not fully capture the dynamics of mask diffusion used in practice. It seems like this might be easy to fix, for instance, by making the candidate token follow from a different distribution.
  • The paper misses important baselines for this task for example [1],[2],[3] among others

[1] Kim, Jaeyeon, et al. "Any-order flexible length masked diffusion." International Conference on Learning Representations. Vol. 2026. 2026. [2] Jazbec, M., Olausson, T.X., Béthune, L., Ablin, P., Kirchhof, M., Monteiro, J., Turrisi Da Costa, V.G., Ramapuram, J. & Cuturi, M.. (2026). Learning Unmasking Policies for Diffusion Language Models. Proceedings of the 43rd International Conference on Machine Learning [3] Chen, Shirui, et al. "dultra: Ultra-fast diffusion language models via reinforcement learning." arXiv preprint arXiv:2512.21446 (2025).

Questions for authors
  • What would happen if we change the distribution of the candidate tokens away from greedy decoding?
  • For the training objective, when setting W(m,y), how was k = 4 chosen? What would happen if we vary this quantity?
4 comments
AuthorsOpenReview ↗

Response 6nBP

Thank you for your review and for recognising the clarity and motivation of our work, the relevance of the reasoning tasks evaluated, and the ablation evidence supporting our design choices. We have conducted additional experiments to address your questions and discuss the results below.

Beyond greedy decoding

TraceLock’s commitment mechanism can also be used with stochastic token proposals. We can generalize the candidate definition to xt,i^∼qt,i(⋅∣xt)\hat{x_{t,i}}\sim q_{t,i}(\cdot\mid x_t) while retaining the same commitment rule and definition of future stability.

To test this, we replaced greedy proposals with categorical sampling at temperature 1.0 on GSM8K using LLaDA, reusing the existing TraceLock checkpoint without retraining.

MethodGreedy accuracy (%)Greedy avg. stepsSampling accuracy (%)Sampling avg. steps
Fast-dLLM79.4079.4079.5278.01
TraceLock80.0085.7880.5984.13

Under categorical sampling, TraceLock maintains its accuracy and decoding-step profile, achieving 80.59% accuracy with 84.13 average steps, compared with 80.00% and 85.78 under greedy decoding. These results demonstrate that the trained controller remains effective under the tested stochastic proposal rule. We will incorporate the generalized formulation and this experiment into the manuscript.

Training objective and sensitivity to k

The weighting factor kk is a training hyperparameter, set empirically to four in our experiments. To assess its influence, we evaluated k∈{1,2,4,8}k\in\{1,2,4,8\} on both backbones while keeping the prompt-disjoint data splits, architecture, initialization, and optimizer settings fixed. Checkpoints were selected by ROC-AUC over active validation positions. We report held-out test results below, with precision and recall measured at the same commitment threshold of 0.95. We will add this ablation to the manuscript.

BackbonekROC-AUCPrecisionRecall
LLaDA10.854295.90%17.21%
LLaDA20.840498.45%8.59%
LLaDA40.854399.83%9.18%
LLaDA80.828834.48%0.14%
Dream10.665063.81%10.00%
Dream20.687328.45%24.65%
Dream40.7439100.00%10.41%
Dream80.6274—0.00%

We particularly focus on precision because an incorrect commitment is irreversible and can affect subsequent denoising, whereas an unaccepted stable token remains available for reconsideration at later steps.

At k=4k=4, the controller achieves the highest observed precision on both backbones, with recall of 9.18% on LLaDA and 10.41% on Dream. On LLaDA, precision increases from 95.90% at k=1k=1 to 99.83%, while ROC-AUC remains nearly identical (0.8542 versus 0.8543). On Dream, k=4k=4 also achieves the highest ROC-AUC among the tested values. Increasing kk to eight sharply reduces recall on both backbones; Dream makes no non-EOT commitments at this threshold, leaving precision undefined.

These results support k=4k=4 as a favourable setting for high-precision future-stability prediction among the values tested.

Additional baselines

Thank you for these suggestions. Our comparisons isolate the contribution of the decoding controller by keeping the pretrained backbone fixed. Among the suggested methods, Learning Unmasking Policies for Diffusion Language Models (Jazbec et al.) follows this setting, and we have added it as a baseline on Dream/GSM8K with generation length 256 and window size 64. TraceLock achieves 80.80% accuracy versus 69.98% for this baseline.

FlexMDM fine-tunes the backbone to support mask insertion and variable-length generation, while dUltra jointly trains the backbone and unmasking planner. Comparing against these modified backbones would mix improvements from backbone training with those from the decoding controller. We will therefore discuss these methods as related work and focus our empirical comparisons on methods operating on frozen backbones.

We additionally evaluated DAPD-Direct, CCD, CCD-DS, and SlowFast. Full results, including accuracy, decoding steps, and latency, are provided in our response to Reviewer 8Gby.


We hope these additional experiments and clarifications address your concerns and demonstrate TraceLock’s applicability beyond greedy decoding. We appreciate your constructive suggestions and would be happy to answer any further questions.

Dear Authors,

Thanks for sharing these ablations and improved formulation which I believe improve the paper.

Regarding the baselines, I accidentally shared the FlexMDM paper instead of [1] where they do a an analysis of different samplers for masked diffusion. I believe the methods presented there should also be considered as a baseline, since it doesn't require additional training.

[1] Kim, Jaeyeon, et al. "Train for the worst, plan for the best: Understanding token ordering in masked diffusions." arXiv preprint arXiv:2502.06768 (2025).

I will keep my score leaning towards acceptance

Thank you for the follow-up and for clarifying the reference, we appreciate your positive assessment. Following your suggestion, we have now evaluated the probability-margin baseline from Kim et al. and included it in the expanded comparison below.

MethodAccuracy (%) ↑Average Decoding Steps ↓Generation Time (s/sample) ↓
TraceLock80.80106.109.76
Fast-dLLM79.90197.4014.74
DAPD-Direct77.9490.988.55
CCD79.38256.0018.29
CCD-DS64.59101.008.17
SlowFast77.7182.737.83
Learned unmasking policy (Apple)69.9862.8410.43
Probability margin (Kim et al., 2025)72.71256.0018.32

Thanks for the feedback, this is nice to see. I keep my score towards acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of reviews

Dear Area Chair,

We sincerely thank you and the reviewers for the thoughtful discussion. We appreciated the reviewers’ recognition of the importance of reusing a controller across decoding regimes (8TES). The reviews also highlighted the importance of improving decoding with a frozen backbone, the breadth of the evaluation across math, question answering, and coding on two backbones, and the evidence for controller reuse and commitment behaviour from the window experiments and single-token interventions (8Gby). We were also pleased that the paper was found “easy to read and well motivated,” and that the coding and math tasks and the ablations supporting our design choices were identified as strengths (6nBP).

We addressed the questions raised by 6nBP and 8TES. For 6nBP, we demonstrated that the same checkpoint works with stochastic token proposals, evaluated sensitivity to the training weight, and added the requested baselines. For 8TES, we quantified the effects of commitment through both individual interventions and full decoding trajectories. We also added Figure 3, which compares measured accuracy and latency for TraceLock, Learn2PD, and Fast-dLLM across two window sizes on both backbones.

8Gby explicitly confirmed that the additional baselines, efficiency definition, quality–latency results, and training-trace details addressed their earlier questions. Their remaining concern centred on whether a controller trained separately for each backbone generalises enough to justify its cost. We therefore ran a new experiment excluding the entire evaluation task family from controller training. Across coding and math on LLaDA and Dream, the held-out controllers scored within 0.22–0.83 percentage points of controllers trained on all three task families in the same ablation. This provides evidence of reuse beyond changes in decoding settings. We also reported the one-time cost per backbone: approximately two GPU-hours for trace collection and 20 minutes for controller training on one A40. We also clarified the differences from Learn2PD, supported by the L2P-Hidden ablation, and distinguished TraceLock’s learned commitment policy from standard speculative decoding.

We thank all three reviewers again for the time they devoted to the paper and for the suggestions that improved it. We also thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.