acceptodds
Withdrawn from ICLR 2027

Fast-to-Slow Thinking Comes Naturally to Multimodal Diffusion Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion language models (DLLMs) enable parallel generation by denoising multiple tokens at arbitrary masked positions, with existing studies primarily exploiting this property to accelerate inference. Beyond efficiency, the flexibility to generate different numbers of tokens at each step naturally supports different thinking speeds, yet its potential for fast-to-slow thinking remains unexplored. We exploit this inherent property of DLLMs for fast-to-slow thinking, without specially designed training or additional denoising steps compared with standard decoding. We first obtain a complete thought through fast denoising, then remask uncertain tokens for slow refinement while retaining confident predictions as context. We further extend this idea to multimodal DLLMs, where fast and slow thinking favor different visual conditions: rich visual details can distract aggressive fast denoising but benefit slow refinement. We thus introduce EdgeDrop, which suppresses regions with complex edge information during fast thinking and restores the whole image for slow refinement. Extensive experiments across four multimodal benchmarks demonstrate consistent performance improvements with comparable overall inference time. Notably, the fast stage alone achieves a speedup while outperforming standard decoding on SQA-IMG and MMBench, with subsequent slow refinement yielding further gains on M3CoT and V*.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, multimodal large language models

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_79E8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes F2S-DLLM, a fast-to-slow thinking framework for multimodal diffusion language models. The method first generates a complete response by unmasking multiple tokens in parallel, then selectively remasks low-confidence tokens for further refinement. The authors also introduce EdgeDrop, which applies noise to regions with complex edge information during the fast stage, while restoring the original image during the slow stage.

The method is evaluated on three multimodal diffusion models—LaViDa, MMaDA, and LLaDA-V—across four multimodal benchmarks. The authors report performance improvements over standard decoding.

Critical strengths
  1. A practical training-free approach with a fixed denoising budget.

The proposed method redistributes the existing denoising budget between initial generation and selective refinement without additional training or external modules. The ability to refine generated responses without increasing the number of denoising steps is a potentially useful direction for efficient inference-time decoding.

  1. An interesting perspective on visual conditioning across decoding stages.

Through EdgeDrop, the paper explores the idea that different types of visual information may be beneficial during initial generation and subsequent refinement. Considering the role of visual complexity at different decoding stages, rather than merely modifying the decoding schedule, is an interesting research direction. However, the validity of this hypothesis requires more convincing experimental evidence.

Critical weaknesses
  1. Insufficient comparisons with existing decoding and refinement methods make it difficult to assess the advantages of the proposed approach.

My primary concern is the very limited comparison with existing methods. Table 1, which presents the main results, compares F2S-DLLM only against standard decoding and its own variant without EdgeDrop. Although Figure 4 includes a comparison with ReMDM, this evaluation is restricted to a single backbone, LaViDa. No direct comparisons with existing decoding or refinement methods are provided for the other two backbones.

In particular, the proposed approach closely resembles existing confidence-based remasking and draft-then-refine strategies. The paper discusses relevant methods, including Speculative Correction, but does not experimentally compare against these directly related approaches. Consequently, it is difficult to determine what meaningful advantages F2S-DLLM provides over existing methods.

Furthermore, the performance improvements on MMaDA and LLaDA-V are relatively modest. EdgeDrop also fails to consistently improve performance across all backbones; on LLaDA-V, the variant without EdgeDrop achieves a higher micro-average score than the complete method.

For a paper proposing a new decoding strategy, relying primarily on comparisons with standard decoding, without sufficient evaluation against strong and directly relevant baselines, is a significant experimental weakness. The current results demonstrate improvements over standard decoding in certain settings but do not convincingly establish the superiority or distinct contribution of the proposed approach over existing methods.

  1. The claimed fast-to-slow thinking mechanism is not sufficiently distinguished from conventional token refinement, and its effectiveness is not convincingly established.

The authors claim to introduce fast-to-slow thinking, but the proposed method essentially generates a complete response quickly and then remasks low-confidence tokens for refinement. It is unclear how this process fundamentally differs from existing uncertainty-based refinement methods. Moreover, the method applies the same fixed fast/slow schedule to every input, without adaptively adjusting the thinking mode or computational budget according to problem difficulty.

The experiments are also limited to multiple-choice tasks with a maximum generation length of 64 tokens. Many qualitative examples primarily demonstrate corrections to final answer tokens rather than improvements in the underlying reasoning process. Thus, while the results suggest that selective refinement can improve answer accuracy, they do not sufficiently support the stronger claim that the method enables fast-to-slow thinking or improves reasoning capabilities.

There is also a concern regarding a key implementation detail. The authors compute token confidence using the logits from the final fast step. However, at this point, most tokens have already been committed and are included in the model input. It is unclear whether confidence scores for already committed tokens are directly comparable to those for positions that remain masked. This issue could directly affect the validity of the uncertainty-based remasking strategy, which is central to the proposed method.

The motivation behind EdgeDrop is also insufficiently validated. The authors argue that complex visual information can distract fast thinking while benefiting slow thinking. However, the experiments primarily demonstrate differences in final accuracy rather than directly establishing this mechanism. Furthermore, in Figure 2, increasing the remasking budget K also increases the number of denoising steps, making it difficult to distinguish the benefits of selective refinement from those of additional computation.

As a minor presentation issue, the terminology is somewhat inconsistent. Although the paper focuses on multimodal diffusion language models, the proposed method is named F2S-DLLM, and Figure 1 also uses the term DLLM. While DLLM can broadly encompass multimodal models, the distinction between general DLLMs and multimodal DLLMs (dMLLMs) should be made clearer and more consistent, especially since visual conditioning is a central contribution. This is not a major technical flaw, but it makes the scope of the proposed method less precise.

Questions for authors

1.What fundamentally distinguishes F2S-DLLM from existing confidence-based remasking and draft-then-refine approaches? Can the authors provide comparisons with directly related methods, particularly Speculative Correction, under identical model configurations and inference budgets?

2.Why is the comparison with ReMDM restricted to LaViDa? Can the authors provide direct comparisons with established decoding and refinement methods on MMaDA and LLaDA-V as well?

3.How do the authors justify describing the proposed method as fast-to-slow thinking rather than conventional token refinement? Can they provide evidence demonstrating improvements in the actual reasoning process?

4.Are confidence scores for already committed tokens and still-masked tokens directly comparable when computed using the logits from the final fast step? Please clarify how token uncertainty is measured and used for remasking.

5.Why does EdgeDrop provide only limited improvements or even degrade performance on certain backbones? Additionally, can the authors analyze the effect of the remasking budget K under a fixed denoising-step budget?

4 comments

W1/Q1/Q2:Insufficient comparisons

ModelMethodM3cotV*SQA-IMGMMBenchMirco Avg
LavidaBaseline34.8645.5562.2265.0556.08
ReMDM37.6242.9365.4968.2859.07
SC34.4342.4163.9165.6356.57
F2S(w/o ED)36.3247.1266.1469.7259.67
F2S37.4948.6966.7369.7260.15
MMaDABaseline26.6232.4646.1639.4537.47
SC25.5035.6045.4639.7137.21
ReMDM27.0135.0844.2739.4337.19
F2S(w/o ED)24.8133.5146.2640.1937.4
F2S26.136.6547.2541.0338.36
LLaDA-VBaseline47.841.3667.6774.5265.25
SC47.5037.7068.1273.6264.75
ReMDM48.2341.8869.2674.2965.62
F2S(w/o ED)49.439.2768.9175.166.19
F2S49.0143.9868.7774.9166.06

We thank the reviewer for pointing out the insufficient baseline comparisons and method distinctions. We add comparisons against ReMDM and Speculative Correction (SC) on three backbones and four benchmarks; results are shown in the table above. Under identical evaluation settings, the full F2S achieves Micro Averages of 60.15, 38.36, and 66.06 on LaViDA, MMaDA, and LLaDA-V, respectively, in each case above ReMDM (59.07, 37.19, 65.62) and SC (56.57, 37.21, 64.75). We also report F2S w/o EdgeDrop so that the contributions of the fast-to-slow decoding schedule and of the visual perturbation can be separated.

W2/Q3:Special Contrubition

We call the process "fast-to-slow thinking" because it is inspired by the human cognitive pattern of shifting from rapid judgment to deliberate reasoning: the model first forms a complete preliminary answer quickly and keeps the high-confidence, relatively easy portions, then allocates more fine-grained decoding to the low-confidence portions that may require further reasoning. Our method thus performs a transition from rapid drafting to slow deliberation within a single sample, rather than applying the same decoding granularity throughout generation. Naturally, slow thinking, at the operational level, uses uncertainty-based token refinement, but this is merely a technical means of connecting the fast and slow decoding modes; it is not the entirety of what we propose as fast-to-slow thinking.

Even though all three methods use uncertainty-based token refinement, we differ from ReMDM and SC in both the research problem and the solution.

  1. Research problem. All three methods concern error accumulation in diffusion language model generation, but the settings differ. ReMDM concerns the inconsistency that may arise between tokens already generated and content generated later in the decoding process, and corrects it through continual re-masking. SC concerns how errors in one block propagate to subsequent blocks in block-autoregressive decoding. We concern the high parallel commit rate of the fast thinking stage: the model may prematurely commit to content that still requires careful judgment, so that this content survives into the complete draft and affects subsequent refinement.
  2. Solution. ReMDM repeatedly introduces re-masking during decoding, giving already-generated tokens a chance to be updated in later steps. SC forms a complete draft and then applies a threshold-based correction. Ours completes the fast draft and then applies concentrated re-masking, followed by slow refinement at 2 tokens per step. Compared with ReMDM, we decide which positions to refine only after the complete draft is formed; compared with SC, we deliberately organize the slow stage as lower-parallelism refinement, so that the slow stage is itself less likely to introduce the errors associated with parallel generation.

W2/Q4:Confidence Comparability and Remasking Procedure

We compute token uncertainty directly from the logits of the final fast step. For each position ii, we obtain the predictive distribution pi=Softmax⁡(ℓi)p_i=\operatorname{Softmax}(\ell_i) and define confidence as ci=max⁡vpi(v)c_i=\max_v p_i(v). Uncertainty is then ui=1−ciu_i=1-c_i; positions with lower confidence are prioritized for remasking.

To assess whether confidence scores for committed and still-masked positions are directly comparable, we analyzed 100 LaViDA samples and grouped token positions by their state at the start of the final fast step. Of the 64 positions per sample, 60 had been committed and four remained masked. Results are as follows.

State at last fast step# positionsMedian final-step confidenceBottom-32 entry rate
Committed6,0000.999546.88%
Still masked4000.678196.75%
All6,4000.999350.00%

In the current method, we compute the confidence of each position from the predictive distribution of the last fast step and re-mask the K=32K=32 lowest-confidence positions. The results above show that the score distributions of the two kinds of positions indeed differ: still-masked positions enter the slow stage more readily.

To further examine the validity of the current confidence ranking, we compare three selection strategies on the same set of LaViDA samples. Leave-one-out (LOO) masks one position at a time within the complete draft and runs a forward pass to recompute that position's confidence; "re-forward once" performs one additional forward pass after the complete draft is obtained and recomputes the confidence of all positions. All three strategies select the 32 lowest positions for slow refinement.

Selection strategyAccuracyOverlap w/ LOOOverlap w/ re-forwardOverlap w/ Final-fastNFE
LOO40.0%—76.81%69.56%32+64
Re-forward once39.0%76.81%—78.63%33
Final-fast (ours)42.0%69.56%78.63%—32

The bottom-32 positions selected by the three methods overlap by 69.56%–78.63%. On these 100 samples, the Final-fast strategy requires no additional forward pass and reaches 42.0% accuracy, whereas LOO and re-forward once reach 40.0% and 39.0%, respectively. This indicates that although the confidences of committed and still-masked positions at the last step are not strictly scores from the same condition, the current ranking still selects useful refinement positions at lower computational cost.

We further examine the four positions that remain masked at the last fast step. Under the current Final-fast logits, 387 of the 400 such positions enter the bottom-32 (96.75%); after the draft is complete, an additional forward pass that recomputes confidence under the full-text condition still selects 316 of them (79.00%).

How bottom-32 is computedEntering bottom-32Token entry probabilityMean entries per sampleAll four entering
Final-fast logits387/40096.75%3.87/490%
Full-draft re-forward316/40079.00%3.16/449%

This shows that the masked state at the last step does raise the probability that these positions enter the bottom-32; yet even when they are re-evaluated uniformly on the complete draft, 79.00% are still selected, so their high entry rate cannot be explained by the masked state alone. We also tabulate the entry rate under the Final-fast strategy by commit time: positions committed at steps 1–4, 5–8, 9–12, and 13–15 enter at 37.50%, 45.44%, 49.69%, and 57.58%, respectively, showing a trend in which later-committed positions are more likely to be re-evaluated. Together with the strategy comparison above, we regard Final-fast confidence as an effective but state-heterogeneous practical ranking signal.

We emphasize that how token uncertainty is computed is not the core contribution of this paper. The current method uses confidence to select re-masked positions; in principle, margin, entropy, and other measures could also serve as candidate uncertainty metrics, but we have not yet compared these selection strategies systematically.

W2/Q3:Reasoning Process

Regarding improvement in the actual reasoning process, Appendix Table 9 shows the complete change from the Fast Thinking draft to the Slow Thinking output on the same samples. These cases change not only the final answer token but also correct the reasoning content preceding the answer. In the tow-truck example, for instance, the fast draft gives contradictory explanations of the direction of force and selects the wrong answer, whereas the slow stage re-assesses the towing relationship and corrects both the reasoning statement and the final answer; other cases similarly show corrections to judgments about vehicle distance, object orientation, and visual attributes. The improvement therefore does not occur only at the answer option, but manifests in the visual judgments and reasoning statements that support the answer.

Table 7 provides quantitative evidence: even under the Original Image condition without EdgeDrop, the slow stage corrects 475 erroneous fast predictions across the four benchmarks while introducing only 161 regressions, a net correction of +314. This indicates that the slow stage brings a substantially positive correction to the fast draft overall. Together, these results show that fast-to-slow decoding changes not only the generated text but also the reasoning performance on the multimodal tasks we evaluate.

To directly test whether slow refinement improves the reasoning text, we sampled 100 examples stratified by domain from the 2,318 M³CoT samples and used GPT-6 Sol to score, along six dimensions, the fast draft and the slow output produced for the same example.

StageReasoningCoherenceGroundingCompletenessFormatLanguage
Fast draft3.173.343.084.113.512.59
Refined output3.684.073.374.324.144.12
Δ+0.51+0.73+0.29+0.21+0.63+1.53

After slow refinement, the reasoning-correctness score rises from 3.17 to 3.68, logical coherence from 3.34 to 4.07, and evidence grounding from 3.08 to 3.37, while the six-dimension total increases from 19.80 to 23.70. Among the 100 paired comparisons, the refined output is judged overall better in 78 cases and the draft in only 9, with the remaining 13 tied; the gain is significant in all six dimensions under a paired sign test (p<0.001). Together with the case study in Table 9, these results indicate that the benefit of the slow stage is reflected not only in the final answer but also in the quality of the reasoning text.

W2/Q5: EdgeDrop Robustness

To test the robustness of the EdgeDrop increment relative to this ablation variant, we run three independent runs on each of the three backbones.

BackboneF2S (w/o ED)Complete F2S-DLLMMean (Δ)95% CI of (Δ)
LaViDA59.6760.10 ± 0.06+0.43[+0.29, +0.56]
MMaDA37.4038.65 ± 0.26+1.25[+0.60, +1.89]
LLaDA-V66.1965.94 ± 0.14−0.25[−0.59, +0.09]

For LaViDA and MMaDA the confidence intervals lie strictly above zero, showing that EdgeDrop, as the visual scheduling component of the complete F2S-DLLM, delivers a stable incremental gain on these two backbones. For LLaDA-V the interval contains zero, so the result indicates that the incremental effect of EdgeDrop is not significant on this backbone; it does not establish a stable degradation.

Existing ablations also show that the effect of EdgeDrop is not caused by generic visual noise. On LaViDA, Table 3 indicates that the gain of the complete F2S-DLLM depends on which visual region is perturbed in the fast stage, and does not merely come from adding noise. Table 7 further shows that the visual condition affects how refinable the draft is. The effect of EdgeDrop therefore need not appear as higher fast-draft accuracy; rather, it makes the state produced by the fast stage more amenable to subsequent slow refinement.

In some samples, temporarily suppressing high-detail regions can reduce the premature influence of local evidence on aggressive parallel decoding, so that the affected content can be re-judged in the slow stage once the original image is restored; if, however, a suppressed region contains evidence needed for the current question, EdgeDrop can also lower the quality of the fast draft. Since different backbones use different visual encoders while we apply the same pixel-level EdgeDrop parameters to all models, the balance between this benefit and the information loss may vary across models. We will expand the discussion of this point in a revised version.

W2/Q5:Fixed-Budget KK Ablation

We further add an ablation on KK under a fixed denoising-step budget. The experiment uses LaViDA and evaluates on M³CoT, V*, ScienceQA, and MMBench; the output length is fixed at 64 tokens and EdgeDrop is disabled. All settings share the same 16-step fast draft and perform the same 16-step slow refinement, so the total budget is exactly 16+16=3216+16=32 NFE. The only difference across settings is the number of re-masked positions, K∈{8,16,24,32}K\in\{8,16,24,32\}.

KKM3CoTV*SQA-IMGMMBenchMicro Avg.
834.5143.4663.6665.9356.70
1635.2943.4664.8568.3858.37
2436.5442.9365.6969.6259.49
3236.3247.1266.1469.7259.67

Relative to the shared fast draft, K={8,16,24,32}K=\{8,16,24,32\} bring micro-average gains of +0.68, +2.36, +3.47, and +3.65 points, respectively. Although individual benchmarks are not strictly monotonic, both Macro and Micro Averages rise consistently with KK, and K=32K=32 performs best within the range evaluated.

Increasing KK does not increase the number of model forward passes: all settings use exactly 32 NFE, and the average per-sample runtime is comparable (1.83–1.91 s). The performance difference therefore cannot be attributed to more denoising steps. The results show that, under the same forward budget, widening the set of positions re-evaluated in the slow stage improves final performance.

W2:Input-Adaptive

We agree that the current experiments validate a fixed, within-sample fast-to-slow decoding strategy rather than input-adaptive compute allocation. The contribution of this paper is that even without dynamic routing by input difficulty, this within-sample reallocation from rapid parallel generation to selective slow refinement already outperforms standard decoding under the same budget. The current experiments nevertheless cannot show that the model adaptively decides whether to enter the slow stage, or dynamically allocates the slow-stage budget, according to question difficulty. We will therefore state this boundary explicitly in the abstract and introduction. At the same time, deciding dynamically, based on input difficulty or uncertainty, whether to trigger slow refinement and how much refinement budget to allocate is a valuable direction for follow-up work, which we will discuss as future work.

W2: Terminology

We thank the reviewer for pointing out this terminology inconsistency. We fully agree, and will unify the wording of Figure 1 and the main text in the revision, clearly distinguishing general DLLMs from the multimodal dMLLMs studied in this paper.

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

The paper studies whether the multimodal diffusion language model can use different denoising granularities for fast-to-slow inference. It proposes F2S-DLLM, an inference-time fast-to-slow decoding strategy for multimodal diffusion language models. The method first forms a complete fast-stage prediction, then remasks the 32 lowest-confidence tokens and refines them with finer-grained denoising. For multimodal inputs, EdgeDrop perturbs high-edge regions during the fast stage and restores the original image during slow refinement.

As for the contributions, the main contribution is an inference-only decoding schedule that combines coarse parallel drafting with uncertainty-guided remasking. Besides, the paper also introduces EdgeDrop, which provides different visual conditions to the fast and slow stages. The method does not require additional training or an external visual verifier.

For about the validation, the paper evaluates three multimodal diffusion language models on four multiple-choice benchmarks. Under LaViDa, full F2S-DLLM improves the micro average from 56.08 to 60.15 under the same 32-step budget. The micro-average gains are 4.07 points for LaViDa, 0.89 points for MMaDA, and 0.81 points for LLaDA-V. Table 2 shows that lowest-confidence selection outperforms random and highest-confidence selection on all four reported benchmarks. Table 10 also shows that the 4/2 schedule achieves 60.15 compared with 56.08 for the matched standard baseline. However, the improvement is not uniform: on MMaDA M3CoT, F2S-DLLM obtains 26.10, below the standard decoding result of 26.62.

In conclusion, I lean toward weak rejection. The decoding idea is promising and the empirical results are encouraging, The core two-stage decoding schedule is directly compared with standard decoding through the w/o-EdgeDrop control. However, the incremental effect of EdgeDrop is small and backbone-dependent, and the τ=0 result in Figure 3 does not match the corresponding w/o-EdgeDrop result in Table 1. The claimed 2× speedup refers to halving the number of denoising steps in the fast-only stage; The paper does not report a fast-only wall-clock latency measurement, so the claimed 2× fast-stage speedup is not directly validated in wall-clock terms.

Critical strengths
  1. The paper introduces a clear and simple inference-time mechanism for using the parallel nature of DLLMs beyond acceleration. Completing a coarse complete thought and selectively remasking uncertain tokens is a meaningful design direction because it gives DLLMs an explicit way to reallocate computation within a fixed denoising budget.

  2. Within the tested protocol, Table 3 shows that suppressing high-edge regions gives the strongest final accuracy among the tested perturbation locations, Table 4 shows that restoring the original image benefits slow refinement, and Table 7 reports the highest total Net Fix for EdgeDrop. Therefore, these results provide empirical evidence that the visual condition used during fast decoding affects subsequent refinement.

Critical weaknesses
  1. The paper motivates fast-to-slow thinking partly in terms of allocating reasoning effort according to input difficulty or uncertainty. However, the main experiments apply the same 16/16 schedule and K = 32 remasking budget to every input, and every input undergoes slow refinement. Thus, I am concerned that the experiments establish a fixed two-stage coarse-to-fine schedule, rather than input-adaptive allocation of reasoning computation. If input-level routing is outside the intended scope, the abstract and Introduction should state this boundary explicitly.

  2. Table 1 provides a matched comparison between F2S-DLLM with and without EdgeDrop, but the incremental effect of EdgeDrop is modest and model-dependent: it changes the micro average by +0.48 points on LaViDa, +0.96 on MMaDA, and −0.13 on LLaDA-V. The gains are also not uniformly positive: on MMaDA M3CoT, the EdgeDrop result remains below standard decoding (26.10 vs. 26.62). Finally, Figure 3 reports 36.58 for M3CoT at τ = 0, however, Table 1 reports 36.32 for F2S-DLLM without EdgeDrop. The authors should reconcile this discrepancy and clarify whether the relevant ablations use identical backbones, prompts, preprocessing, and evaluation settings.

Questions for authors

Clarification Requests:

  1. The main experiments use a fixed 16/16 schedule and K = 32 for every input. Is “fast-to-slow thinking” intended to denote a deterministic two-stage schedule, or input-adaptive allocation of computation? If the latter is not addressed, please narrow the claims in the abstract and Introduction.

  2. Table 7 reports aggregate Fix and Regress counts. What fraction of examples changes its final answer after slow refinement for each benchmark, and what backbone is used for this analysis?

  3. Please clarify the interpretation of the w/o-EdgeDrop row in Table 1 and report whether the EdgeDrop gains are robust under repeated runs or paired bootstrap confidence intervals.

  4. How were K, the 4/2 decoding schedule, and the EdgeDrop parameters (τ, ρ, λ) selected? Were they fixed before evaluation, or selected after inspecting the reported benchmark examples? Please also state which backbone is used for each ablation table.

  5. How should the M3CoT value at τ = 0 in Figure 3 (approximately 36.58) be reconciled with the 36.32 result for F2S-DLLM without EdgeDrop in Table 1? Please clarify whether these results use identical evaluation examples, prompts, preprocessing, and backbone settings.

Suggestion for future work:

As future work, it would be useful to evaluate input-adaptive slow-stage allocation on open-ended multimodal tasks under a fixed latency budget.

2 comments

W1/Q1:Overclaim about input-adaptive allocation

We thank the reviewer for the accurate summary. We agree that the current experiments validate a fixed, within-sample fast-to-slow decoding strategy rather than input-adaptive compute allocation. The current experiments, however, cannot show that the model adaptively decides whether to enter the slow stage, or dynamically allocates the slow-stage budget, according to question difficulty. We will therefore state this boundary explicitly in the abstract and introduction.

W2/Q5: EdgeDrop Robustness and Result Consistency

We thank the reviewer for pointing out that the EdgeDrop gain is task- and backbone-dependent. Relative to F2S w/o EdgeDrop, EdgeDrop improves V* by 1.57, 3.14, and 4.71 points on LaViDA, MMaDA, and LLaDA-V, respectively; yet its micro-average gain is modest, and on LLaDA-V it is slightly negative. We also acknowledge that on M3CoT with MMaDA, adding EdgeDrop still leaves the result below standard decoding (26.10 vs. 26.62). This shows that the effect of EdgeDrop is task- and backbone-dependent.

In Figure 3, M3CoT is 36.58 at τ=0\tau=0, whereas Table 1 reports 36.32 for F2S-DLLM without EdgeDrop. We acknowledge that τ=0\tau=0 corresponds to applying no EdgeDrop and should therefore match the F2S w/o EdgeDrop result under the same setting. On checking, we found that Figure 3 and Table 1 were originally produced by separate runs under different computational environments. We have re-verified τ=0\tau=0 using the backbone, samples, prompt, preprocessing, and decoding settings of Table 1 and obtained 36.32, consistent with Table 1. We will correct Figure 3 in the revision and state the experimental setting of the relevant ablation explicitly.

Q2:Answer-Change Fraction and Analysis Backbone

We thank the reviewer for the question. The analysis in Table 7 uses the LaViDA backbone. We compare, sample by sample, the final answers extracted after fast thinking and after slow thinking; answer changes include Fix, Regress, and wrong-to-wrong transitions (Wrong→Wrong). Results per benchmark are as follows.

visual conditionM³CoT (N=2,318)MMB (N=4,329)SQA (N=2,017)V* (N=191)
Original Image247 (10.66%)363 (8.39%)110 (5.45%)31 (16.23%)
Random Noise273 (11.78%)423 (9.77%)132 (6.54%)33 (17.28%)
Low-Detail Noise277 (11.95%)439 (10.14%)130 (6.45%)34 (17.80%)
EdgeDrop246 (10.61%)412 (9.52%)143 (7.09%)37 (19.37%)

These results show that the frequency of changes made in the slow thinking stage is not equivalent to correction performance. For example, EdgeDrop yields fewer confirmed answer changes than Random Noise and Low-Detail Noise, yet Table 7 shows that it attains the largest total net correction (+380, versus +354 and +332, respectively).

Q3: w/o-EdgeDrop Interpretation and Statistical Robustness

The "w/o EdgeDrop" row of Table 1 retains the complete fast-to-slow decoding process — 16 fast steps, K=32K=32 re-masking, and 16 slow steps — while using the original, complete image in both stages with no visual-region perturbation. This setting therefore isolates the contribution of the fast-to-slow schedule itself. Because w/o EdgeDrop uses greedy decoding and involves no stochastic visual perturbation, its result is not seed-dependent. We further repeat EdgeDrop with three different seeds.

BackboneStandardF2S (w/o ED)F2S-DLLMMean Δ vs95% CI of Δ
LaViDA56.0859.6760.10 ± 0.06+0.43[+0.29, +0.56]
MMaDA37.4737.4038.65 ± 0.26+1.25[+0.60, +1.89]
LLaDA-V65.2566.1965.94 ± 0.14−0.25[−0.59, +0.09]

The Micro Averages of LaViDA, MMaDA, and LLaDA-V are 60.10 ± 0.06, 38.65 ± 0.26, and 65.94 ± 0.14, respectively. The 95% t-confidence intervals computed over the three runs show that, relative to the deterministic w/o EdgeDrop, the mean increment on LaViDA is +0.43 with interval [+0.29, +0.56], and on MMaDA +1.25 with interval [+0.60, +1.89]. Both intervals lie strictly above zero, indicating that the gain is stable across repeated runs. For LLaDA-V, the mean increment is −0.25 with interval [−0.59, +0.09], indicating that the incremental effect of EdgeDrop is not significant on this backbone. These results therefore support a stable EdgeDrop gain on LaViDA and MMaDA, while also showing that its effect is partly backbone-dependent.

Q4: Hyperparameter Selection and Ablation Settings

We thank the reviewer for noting that the parameter-selection procedure is not sufficiently described in the main text. The main experiments use an output length of 64 tokens with a total decoding budget fixed at 32 NFE. We first generate the complete draft in 16 fast steps, then re-mask half of the positions (K=32K=32) and refine them with 16 slow steps. The fast and slow stages therefore commit 4 and 2 tokens per step, respectively; these two decoding speeds are determined by the output length, the number of steps per stage, and KK, and are not independently tuned hyperparameters. Our experiments are designed to test whether, under the same computational resources, this compute allocation can improve on standard decoding. We subsequently ran an ablation on the choice of KK; performance generally improves as KK grows and is best at K=32K=32. We emphasize that this ablation serves as a post-hoc verification of the originally fixed structural setting, not as the basis on which that setting was initially chosen.

Regarding the EdgeDrop parameters (τ\tau, ρ\rho, λ\lambda), we acknowledge that τ=50\tau=50 was chosen from a sensitivity experiment on M3CoT: we selected a subset of the M3CoT data, measured accuracy at different τ\tau magnitudes, and settled on τ=50\tau=50 as our final parameter. Once this choice was made, we fixed τ\tau and applied it to all remaining benchmarks and backbones, without tuning it separately for different datasets or models. By contrast, the region quantile ρ=0.5\rho=0.5 and the complexity exponent λ=2.5\lambda=2.5 are fixed heuristic defaults, and we did not sweep these two parameters separately.

τ\tau050100150200
M3CoT.36.32%37.49%36.28%36.50%36.32%

Q6: Suggestion for future work

We are very grateful to the reviewer for this insightful and encouraging suggestion. We share the reviewer's enthusiasm for this direction, as it closely aligns with the question we are currently exploring: how to jointly exploit visual and textual signals to adapt decoding speed to input difficulty and to allocate the compute budget of the slow-thinking stage dynamically. The current method uses a fixed fast/slow schedule; realizing input-adaptive compute allocation under a fixed latency budget is a natural and meaningful extension, particularly for open-ended multimodal scenarios with widely varying task complexity. We are glad that the reviewer recognizes the potential of this direction, and we will further emphasize and explore this question in future work.

Reviewer_R2RqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes F2S-DLLM, a training-free decoding framework for MDLLM. The main idea is to first generate a complete response through aggressive parallel denoising, and then selectively remask low-confidence tokens for slower refinement while retaining high-confidence predictions as context. The authors further introduce EdgeDrop, which perturbs visually complex regions during fast decoding and restores the original image during refinement.

The method is evaluated on three models across four benchmarks. The results show improvements over standard decoding under a matched denoising-step budget. Additional ablations examine remasking budgets, token selection strategies, visual perturbation choices, and decoding granularities.

Critical strengths
  1. A simple and practical inference-time framework.

The proposed method is easy to implement and does not require additional training, auxiliary models, or extra denoising steps compared with standard decoding. Reallocating a fixed decoding budget between fast drafting and targeted refinement is a practical approach.

  1. Interesting observations on stage-dependent visual conditioning.

EdgeDrop introduces a simple mechanism for controlling visual complexity, and the ablations provide evidence that the choice of visual conditioning affects subsequent refinement.

Critical weaknesses
  1. Limited baseline comparisons

The paper mainly compares against standard diffusion decoding, which does not isolate the benefit of the proposed fast-to-slow schedule. Although ReMDM is included, the improvement without EdgeDrop is relatively small. Stronger comparisons with methods(e.g. Speculative Correction. ) and iterative remasking strategies are needed to establish the novelty of the method.

  1. Insufficient understanding of the improvements.

The method remasks a fixed number of low-confidence tokens, but low confidence does not necessarily indicate an error. It remains unclear how many remasked tokens actually change and contribute to better predictions. Furthermore, the benefits of EdgeDrop are inconsistent across models, and the paper does not sufficiently explain why suppressing complex visual regions improves subsequent refinement.

Questions for authors
  1. Clarification on confidence estimation and remasking effectiveness.

More importantly, could the authors report the fraction of remasked tokens that actually change after refinement, and analyze how often these changes correct an incorrect prediction versus corrupting a correct one? This would help establish whether the proposed remasking strategy is effectively targeting errors of visual information

  1. Is the explicit fast-to-slow schedule necessary?

Could the authors provide a more direct comparison with existing draft-and-refine approaches, particularly Speculative Correction, under matched computational budgets?

It would also be valuable to compare the proposed fixed two-stage schedule against alternative strategies, such as repeated remasking after an aggressive initial decode (e.g., decoding eight tokens per step followed by multiple refinement rounds). Such comparisons would help clarify whether the improvement comes from the more general ability to revise completed predictions.

2 comments

W1. Limited baseline comparisons

We agree that direct comparison against related methods is essential. We therefore add comparisons between F2S-DLLM, ReMDM, and Speculation Correction(SC) across multiple backbones and multiple benchmarks under the same inference budget.

ModelMethodM3CoTV*SQA-IMGMMBenchMicro Avg
LaViDABaseline34.8645.5562.2265.0556.08
ReMDM37.6242.9365.4968.2859.07
SC34.4342.4163.9165.6356.57
F2S (w/o ED)36.3247.1266.1469.7259.67
MMaDABaseline26.6232.4646.1639.4537.47
SC25.5035.6045.4639.7137.21
ReMDM27.0135.0844.2739.4337.19
F2S (w/o ED)24.8133.5146.2640.1937.40
LLaDA-VBaseline47.8041.3667.6774.5265.25
SC47.5037.7068.1273.6264.75
ReMDM48.2341.8869.2674.2965.62
F2S (w/o ED)49.4039.2768.9175.1066.19

As the table shows, under a matched inference budget, F2S (w/o ED) attains a higher micro average than both ReMDM and SC on all three backbones: it improves over ReMDM by 0.60, 0.21, and 0.57 points, and over SC by 3.10, 0.19, and 1.44 points, respectively. F2S-DLLM(w/o ED) therefore does not simply replicate existing remasking methods. Rather, it offers a different compute allocation: the complete draft marks the boundary between the fast and slow stages, and subsequent computation is selectively assigned to the positions that need re-evaluation.

W2:Insufficient understanding of the improvements.

We agree with the reviewer that low confidence is not equivalent to a prediction error. We therefore conduct an analysis on LaViDA over the full evaluation sets of all four benchmarks.

BenchmarkChange rateFixRegressNet fix
M3CoT25.96%12660+66
V*18.47%1610+6
SQA24.07%8123+58
MMBench30.01%26047+213

Only 18.47%–30.01% of the 32 low-confidence tokens change after refinement, which indicates that confidence is used to identify "positions worth re-evaluating" rather than to declare them incorrect. Low confidence is thus not an error label, but it serves as an effective ranking signal for concentrating the limited slow-refinement budget on the positions that most need re-evaluation.

To test EdgeDrop's robustness, we run three independent runs on each of the three backbones:

BackboneF2S w/o EdgeDropF2S-DLLMMean (Δ)95% CI of (Δ)
LaViDA59.6760.10 ± 0.06+0.43[+0.29, +0.56]
MMaDA37.4038.65 ± 0.26+1.25[+0.60, +1.89]
LLaDA-V66.1965.94 ± 0.14−0.25[−0.59, +0.09]

For LaViDA and MMaDA, the EdgeDrop gain reproduces stably; for LLaDA-V the incremental effect is not significant on this backbone.

Table 3 of our paper provides further mechanistic evidence: low-edge and high-edge perturbations produce comparable fast-draft accuracy, yet after completing slow refinement the final accuracies are 36.10 and 37.49, respectively. This indicates that EdgeDrop mainly changes how refinable the draft is. Table 7 tells a consistent story: drafting with the original image yields a net correction of +314, which rises to +380 with EdgeDrop. These results suggest that high-detail regions influence how candidate tokens are formed in the fast stage. Temporarily suppressing such regions allows the model to re-evaluate the relevant positions under global textual context. However, if a suppressed region contains evidence critical to the task, EdgeDrop can also be harmful. Moreover, different visual encoders may respond differently to the same pixel-level perturbation, so the benefit of EdgeDrop varies with task and backbone.

Q1:Confidence Estimation and Remasking Effectiveness

Token confidence is taken directly from the logits of the last fast-step forward. For position ii, we define the maximum predicted probability as ci=max⁡vSoftmax⁡(ℓi)vc_i=\max_v\operatorname{Softmax}(\ell_i)_v. After the fast draft is complete, we sort all 64 positions by cic_i in ascending order, re-mask the K=32 lowest-confidence positions, and then perform slow refinement conditioned on the full image and the retained draft context. This procedure requires no additional model forward pass.

We further measure token changes and answer transitions on LaViDA across the four benchmarks. The results are shown in W2. These results indicate that low confidence effectively identifies the positions worth re-evaluating and yields a positive overall correction gain. We note that this analysis measures correction of errors in the multimodal prediction as a whole; it does not attribute every individual change to visual information. Table 3 in our paper further shows that suppressing high-edge regions yields higher final accuracy than low-edge and random-region perturbations, which indicates that the gain does not stem from generic visual noise. Table 7 shows that EdgeDrop raises the net correction of slow refinement from +314 to +380, indicating that it produces drafts that are more amenable to subsequent correction.

Q2:Matched-Budget Schedule and Iterative Refinement

We add a direct comparison with existing draft-and-refine methods in W1.

We also evaluate more aggressive fast thinking combined with multi-round slow thinking under the same 32-NFE budget. Specifically, we compare the default 4 tokens/step setting with 8, 16, and 32 tokens/step fast thinking. More aggressive decoding completes the initial draft in fewer steps and thus leaves more of the compute budget for subsequent refinement rounds.

MethodDraft speedM3CoTV*SQA-IMGMMBenchMicro Avg
Standard—34.8645.5562.2265.0556.08
F2S (w/o ED)4 token/step36.3247.1266.1469.7259.67
Aggressive 88 token/step36.6746.0766.9868.6359.40
Aggressive 1616 token/step36.1545.0367.6367.2258.70
Aggressive 3232 token/step36.4141.3667.4867.7558.92

All F2S variants achieve a higher micro average than standard decoding, so different fast/slow configurations all deliver an overall gain. Yet more aggressive fast thinking, despite leaving more budget for subsequent slow thinking, does not surpass the default 4-tokens/step configuration overall. This indicates that the improvement does not come solely from adding more slow rounds; the granularity at which the initial draft is generated, and the allocation of compute between the fast and slow stages, matter equally.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author Responses and Key Revisions

Dear Reviewers and Area Chair,

Thank you for your valuable feedback. We have completed point-by-point responses and uploaded the revised manuscript.

1. Insufficient comparisons and control experiments (Reviewers R2Rq and 79E8).

We added comparisons between F2S-DLLM, ReMDM, and Speculative Correction. F2S-DLLM without EdgeDrop achieves higher Micro Averages than both methods across all three backbones. We also added controls combining aggressive drafting with repeated refinement, as well as a fixed-32-NFE ablation of KK. The results demonstrate the importance of both fast-draft quality and compute allocation between the fast and slow stages.

2. Effectiveness of remasking and EdgeDrop (Reviewers R2Rq and Kmuo).

Only 18.47%--30.01% of the remasked tokens change after refinement. This indicates that low confidence is not treated as an error label, but instead serves as a ranking signal for identifying positions worth re-evaluating. EdgeDrop provides reproducible improvements on LaViDA and MMaDA, while its incremental effect on LLaDA-V is not significant, indicating that its benefit is task- and backbone-dependent.

3. Methodological contribution, reasoning evidence, and scope (Reviewers Kmuo and 79E8).

We further clarified that F2S-DLLM performs an intra-sample transition from coarse-grained parallel drafting to lower-parallelism slow refinement, using the completed draft as context for subsequent refinement. This distinguishes it from continuous remasking within an incomplete denoising trajectory. We also clarify that the current method uses a fixed intra-sample fast-to-slow schedule rather than input-adaptive compute routing.

In addition, we clarified the confidence computation, answer-change fractions, and hyperparameter-selection procedure; standardized the relevant terminology; and corrected the numerical inconsistency at (τ=0)(\tau=0), among other presentation issues.

We hope these revisions address the reviewers’ main concerns. If any questions remain, we would be pleased to provide further clarification.

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

Related papers

Discussion (0)

Sign in to comment.