Two Clocks in Diffusion MLLMs: When Answers Stabilize Before Rationales Unfold

Keuntae Kim; Yong Suk Choi

cs.CVOct 1, 2026

Two Clocks in Diffusion MLLMs: When Answers Stabilize Before Rationales Unfold

Authors: Keuntae Kim, Yong Suk Choi

Organizations: Department of Computer Science, Hanyang University, Seoul, Korea

Abstract

An answer candidate in a masked diffusion MLLM can stabilize while its rationale is still unfolding. We distinguish retrospective stabilization of the logged candidate from token commitment, and examine these two clocks relative to rationale generation. Analyzing our results across three visual question-answering benchmarks, we find that 89.4-98.1% of the rationale-side canvas remains unwritten at stabilization in single-block, EOS-suppressed LaViDa runs. On VBench, reducing block length from 128 to 8 changes this fraction from 89.4% to 1.7%, together with answer coverage and the eligible observation window. Under EOS-enabled prompting, direct instructions improve Nemotron's overall accuracy by 15.0 and 19.5 percentage points on M3CoT and ScienceQA, but reduce LaViDa/VBench accuracy by 11.0 points. A symmetric decomposition associates the larger absolute component of each change with coverage rather than conditional accuracy. Matched-canvas image ablations measure visual sensitivity alongside answer stabilization, separating the two temporal readouts. Together, these measurements distinguish answer stabilization, rationale unfolding, and visual sensitivity, and identify coverage as the larger component of the prompting differences.

Figures & tables

Figure 1: Overview of the measurement protocol. (a) Hold the model and input fixed, vary the block schedule, and record the decoding trace. (b) Distinguish stabilization of the logged answer candidate, $T_{\mathrm{stab}}$ , from permanent token commitment, $T_{\mathrm{commit}}$ ; the image-ablation peak $T_{\mathrm{PDM}}$ is a separate diagnostic. (c-i) Measure the unwritten fraction $U$ of the rationale-side region $\mathcal{R}$ at $T_{\mathrm{stab}}$ . (c-ii) Compare image-present and image-removed token distributions on the same pre-step canvas and mask pattern. The PDM readout region is defined in Section 3.3 . The trace and curve are schematic illustrations. After commitment, a logged candidate may be copied from the canvas. Here $\mathrm{PDM}_{t}=D_{t}$ in Equation 3 .

Data	$N$	$r_{\mathrm{pos}}$	$T_{\mathrm{stab}}$	$T_{\mathrm{commit}}$	$U$ (%)	$C$ (%)	$A$ (%)
V*Bench	191	0.15	6.41	23.65	89.4	100.0	41.9
M 3 CoT	400	0.09	0.93	7.60	98.1	99.8	71.5
ScienceQA	400	0.19	1.92	12.14	96.2	97.5	62.5

Table 1: LaViDa timing under single-block, EOS-suppressed decoding.

N

is the evaluation size; timing means are conditional on successful trace localization.

U

measures the rationale-side canvas, not a semantic rationale annotation. Coverage and accuracy use all evaluated outputs.

Figure 2: Recorded answer clocks and the block control. Markers are mean recorded times, and segments are differences of means. All rows are LaViDa with EOS suppressed; the short-block condition also changes answer-slot eligibility.

Data	Correct: mean $U$ (%)	Incorrect: mean $U$ (%)	$p$
M 3 CoT	98.5	97.1	0.68
ScienceQA	96.8	95.1	0.038

Table 2: Correctness-stratified LaViDa results. Mean

U

is the rationale-side fraction still unwritten at stabilization, expressed as a percentage. The last column gives the two-sided Mann–Whitney

p

-value for the correct/incorrect comparison.

Block	$r_{\mathrm{pos}}$	$T_{\mathrm{stab}}$	$T_{\mathrm{commit}}$	$U$ (%)	$C$ (%)	$A$ (%)	$Q$ (%)
128	0.15	6.41	23.65	89.4	100.0	41.9	41.9
8	1.00	53.82	56.01	1.7	91.6	39.8	43.4

Table 3: Block-length control on LaViDa/V*Bench. Both rows evaluate 191 examples with the same configured canvas and step budget.

Q

is accuracy conditional on a valid extracted answer.

Figure 3: Visual-sensitivity timing. Markers are mean individual stabilization and PDM peak steps.

N_{\rm rec}

denotes the recorded subset size. Each row gives the two means for one model–dataset configuration.

Model	Data	$C$	$A$	$Q$
Nemotron	V*Bench	$90.6\to 99.5$	$47.6\to 52.4$	$52.6\to 52.6$
Nemotron	M 3 CoT	$73.5\to 88.8$	$54.5\to 69.5$	$74.1\to 78.3$
Nemotron	ScienceQA	$69.2\to 90.2$	$57.0\to 76.5$	$82.3\to 84.8$
DiffusionGemma	V*Bench	$84.3\to 91.1$	$53.9\to 58.6$	$64.0\to 64.4$
LaViDa	V*Bench	$84.8\to 66.5$	$39.8\to 28.8$	$46.9\to 43.3$

Table 4: EOS-enabled prompting comparisons. Entries are reasoning-requested

\rightarrow

direct (percent), with a shared scorer within each pair.

C

: valid-answer coverage;

A

: overall accuracy;

Q

: conditional accuracy.

N=191

for V*Bench and

400

for the other rows. DiffusionGemma retains its own sampler.

Model	Data	$\Delta A$	Coverage term	Conditional term
Nemotron	V*Bench	$+4.8$	$+4.68$	$0.00$
Nemotron	M 3 CoT	$+15.0$	$+11.66$	$+3.41$
Nemotron	ScienceQA	$+19.5$	$+17.55$	$+1.99$
DiffusionGemma	V*Bench	$+4.7$	$+4.37$	$+0.35$
LaViDa	V*Bench	$-11.0$	$-8.25$	$-2.72$

Table 5: Decomposition from the reported percentages. Values are percentage points. The two terms are computed from rounded

C,Q

and can sum slightly differently from the displayed

\Delta A

; the largest residual is 0.12 points.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Model	Data	$N_{\rm rec}$	Mean $T_{\mathrm{stab}}$	Mean $T_{\mathrm{PDM}}$	Final/peak (%)
LaViDa	V*Bench	191	6.3	3.6	25
LaViDa	M 3 CoT	146	0.7	0.8	13
LaViDa	ScienceQA	146	2.1	0.7	12
Nemotron	V*Bench	64	10.8	25.5	28
Nemotron	M 3 CoT	119	15.0	30.9	47
Nemotron	ScienceQA	118	14.0	30.5	45

Appendix

Table 6: PDM timing and relative final sensitivity.

N_{\rm rec}

is the recorded subset size. The final/peak column is

\rho

from Equation 7 . Stabilization means belong to these PDM evaluations, not the primary timing conditions in Table 1 .

Explore similar work

Routing by Reasoning Need: Trajectory-Aware Decoding Control for Diffusion Vision-Language Models
Sep 12, 2026Yixiang Liu, Zhongxing Xu, Zhonghua Wang +1Diffusion Language Models Inference-Time Steering
Southern University of Science and Technology · Monash University
Answer First, Reason Later: Commitment Order in Diffusion LLMs
Aug 6, 2026Jewon Yeom, Jaewon Sok, Seonghyeon Park +3Diffusion Language Models LLM Reasoning Strategies
Graduate School of Data Science, Seoul National University · Department of Rural Systems Engineering, Seoul National University · Department of Aerospace Engineering, Seoul National University
ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding
Jul 20, 2026Keuntae Kim, Beomseok Lee, Hyunwoo Kim +1Diffusion Language Models Multimodal Reasoning Benchmarks
Hanyang University · LG Electronics · KT Corporation

Sep 12, 2026, cs.AIJ/K move · Enter open · S save

Routing by Reasoning Need: Trajectory-Aware Decoding Control for Diffusion Vision-Language Models

Yixiang Liu, Zhongxing Xu, Zhonghua Wang, Xiaoying Tang

Southern University of Science and Technology · Monash University

Diffusion Language Models Inference-Time Steering Reasoning Skills Decoding Commitment

Open paper AskPDF arXiv

Diffusion vision-language models generate answers through iterative refinement, exposing intermediate answer trajectories that can be inspected and controlled at inference time. However, this controllability creates a reasoning-need mismatch, where a universal generation length is applied to questions with different reasoning demands. Visually closed questions may be harmed by continued refinement after a stable answer has formed, whereas reasoning-sensitive questions may be harmed by premature commitment. We formulate this problem as reasoning-budget mismatch and study it in LLaDA-V. Rather than choosing a universal generation length, our training-free controller routes each example to early commitment, baseline preservation, or reasoning-supportive decoding using trajectory signals from answer closure, commitment evidence, and representation revision pressure, without using ground-truth answers. Across answer-focused, mixed-reasoning, and CoT-sensitive benchmarks, routed control improves robustness over fixed long decoding, pure short decoding, and single-rule interventions. The gains are not explained by shorter outputs alone. Answer-closed examples often benefit from commitment, whereas CoT-sensitive examples require preserving or supporting intermediate reasoning. Taken together, these results suggest diffusion VLM decoding should route inference-time control by the state suggested by the observed trajectory instead of relying on a universal decoding length.

Abstract

Explore similar work

Routing by Reasoning Need: Trajectory-Aware Decoding Control for Diffusion Vision-Language Models

Answer First, Reason Later: Commitment Order in Diffusion LLMs

ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding