Source-linked AI summary
Two Clocks in Diffusion MLLMs: When Answers Stabilize Before Rationales Unfold
Keuntae Kim, Yong Suk Choi
TL;DR
A rationale's displayed order need not faithfully reveal how an answer was produced, and masked diffusion makes prediction and permanent commitment distinct events. The paper measures answer stabilization, rationale commitment, and visual sensitivity across benchmarks, finding that coverage is the larger absolute component of all five prompting comparisons.
Problem
Explanations can improve task performance, yet their order and content need not faithfully reveal the computation behind an answer.
Source: §1 · p. 1Method
The paper compares retrospective stabilization of logged answer candidates with permanent commitment and rationale-side commitment, and uses matched-canvas image ablations and coverage measures.
Source: §1 · p. 2Results
Takeaways & Limitations
Reading rationale text alongside commitment traces distinguishes the order in which explanation tokens are presented from when output tokens become permanent.
Source: §8 · p. 8In the V*Bench block control, lowering U accompanies lower answer coverage and overall accuracy.
Source: §7 · p. 8
Abstract
from arXiv · showhide
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 V*Bench, 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/V*Bench 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.
1 Introduction
The paper examines how answer stability relates to rationale unfolding in masked diffusion generation, where displayed order need not match prediction or commitment order. It relates answer timing to visual sensitivity and answer coverage under explicit schedules and prompting conditions.
Introduction
- Chain-of-thought prompting can improve task performance, but an explanation's order and content need not faithfully reveal the computation behind an answer.In masked diffusion, a token displayed later may be predicted or permanently written before tokens displayed earlier.Source: §1 · p. 1
- The study distinguishes retrospective stabilization of a logged answer candidate from its permanent commitment, and examines both relative to rationale unfolding and visual sensitivity.Rationale unfolding refers to commitment of rationale-side tokens to the canvas.Source: §1 · p. 1
- In the LaViDa block-control analysis, the unwritten fraction falls from 89.4% to 1.7%, linked to changed answer-slot eligibility and lower answer coverage under the evaluated decoding policies.Source: §1 · p. 2
- In LaViDa, the block control links a fall in unwritten fraction from 89.4% to 1.7% with changed answer-slot eligibility and lower answer coverage under evaluated decoding policies.Source: §1 · p. 2
2 Related work
Prior work studies early answer convergence, answer-first commitment, and visual dependence in language and multimodal models. This paper examines these phenomena alongside rationale unfolding, matched-canvas visual sensitivity, and valid-answer coverage.
Early convergence and commitment order
- Li et al. [5] study early answer convergence and introduce Prophet; this paper instead tracks a stable suffix matching the output's own final token on eligible logged steps.The earlier main analysis tracks the first ground-truth-token match on finally correct outputs, with appendix analysis of incorrect outputs.Source: §2 · p. 2
- Yeom et al. [6] study answer-first commitment, finite-canvas backfill, and positional restrictions, while this paper adds visual-sensitivity and valid-answer-coverage comparisons.Their work also compares delaying a reserved answer position with a timing-matched reasoning-token delay.Source: §2 · p. 2
Visual grounding and rationale faithfulness
- Favero et al. [7] introduce visual PDM for autoregressive VLMs, and Thinking Diffusion [8] applies it to diffusion MLLMs; this paper compares matched-canvas image ablations with answer stabilization.Thinking Diffusion also reports early answer generation and weak early visual dependence, and proposes PSP and VRG.Source: §2 · p. 2
- ST-Veto [9] uses confidence residuals and image-attention mass to veto unmasking candidates, unlike this paper's retrospective final-token stability and image-ablation distribution comparison.Turpin et al. [2] show explanations can omit input biases, while Lanham et al. [3] intervene on CoT content to measure answer dependence.Source: §2 · pp. 2–3
Multimodal diffusion systems
- The analysis covers answer timing in LaViDa, prompting and visual sensitivity in LaViDa and Nemotron, and an additional prompting comparison in DiffusionGemma.These systems provide different decoding implementations.Source: §2 · p. 3
3 Measurements
The measurements separate logged answer candidates from permanent writes, quantify rationale-side commitment at stabilization, compare matched-canvas visual sensitivity, and decompose accuracy into coverage and conditional accuracy.
3.1 Observable candidates and the two clocks
- Tstab is the first eligible step after which the logged answer candidate remains the final answer, while Tcommit is when that answer position becomes permanent.The stabilization scan is retrospective and excludes steps where the answer position is not observable.Source: §3.1 · p. 3
3.2 Rationale unfolding under a commitment schedule
- U measures the mean per-example fraction of rationale-side positions still unwritten at Tstab; same-step commitments count as written.The operational region includes formatting tokens and positions retained under EOS suppression, and U is derived from the rationale-side commitment distribution.Source: §3.2 · p. 3
3.3 Matched-canvas visual sensitivity
- Matched-canvas PDM-H compares image-present and image-removed token distributions on the same pre-step canvas and mask pattern.Peak times are means of individual first-maximum readouts; image-derived text or visible-token copying can reduce sensitivity without removing visual information.Source: §3.3 · pp. 3–4
3.4 Valid-answer coverage and accuracy
- Valid-answer coverage C is the fraction of outputs yielding an option accepted by the shared scoring rule, and overall accuracy satisfies A = CQ.Missing answers and format or extraction failures count as uncovered.Source: §3.4 · p. 4
4 Experimental setup
The experiments use three visual question-answering benchmarks and distinct timing and prompting configurations across the evaluated models. Timing statistics, scoring, and configuration controls are specified separately.
Benchmarks
- The evaluation uses V*Bench (191 examples), M3CoT (400 examples), and an image-containing ScienceQA subset (400 examples).The passages characterize them respectively as fine-grained visual search, multimodal multi-step reasoning, and an image-containing ScienceQA subset.Source: §4 · p. 4
Decoding configurations
- LaViDa timing uses a 128-token canvas, 64 denoising steps, one block, greedy candidates, and EOS suppression; its V*Bench control uses block length 8.Nemotron timing disables its confidence early-exit shortcut and suppresses EOS; approximate dLLM-Cache is disabled in traced settings.Source: §4 · p. 4
Prompting configurations
- Prompting comparisons allow EOS and match configured budgets within each model, but direct prompts change both explanation instructions and response-format constraints.LaViDa and Nemotron use 128 tokens and 64 steps, while DiffusionGemma retains its own sampler.Source: §4 · p. 4
Scoring and analysis
- Timing means are conditional on valid answer localization, whereas coverage and overall accuracy use the full evaluated set.Appendices specify extraction, aggregation, prompt, dataset-interface, and decoding-control rules.Source: §4 · p. 4
5 Results
Across LaViDa's timing analyses, answer stabilization often precedes commitment of rationale-side canvas tokens, while schedule eligibility changes that gap. EOS-enabled prompting effects vary by model, and coverage is the larger arithmetic component of those differences.
5.1 Answers stabilize while most rationale-side tokens remain unwritten
- 89.4%, 98.1%, and 96.2% of LaViDa’s rationale-side canvas remains unwritten at answer stabilization on V*Bench, M3CoT, and ScienceQA.Timing means are conditional on successful trace localization; coverage and accuracy use all evaluated outputs.Source: §5.1 · p. 4, Table 1 · p. 5
5.2 Correctness-stratified statistical analysis
- Both correctness groups have mean U above 95% on M3CoT and ScienceQA; ScienceQA’s group means differ by 1.7 points across 398 analyzable examples.The comparison uses nominal, per-benchmark two-sided Mann–Whitney p-values.Source: Table 2 · p. 5, §5.2 · p. 5
5.3 The block control changes measurement and output coverage
- Reducing LaViDa/V*Bench block length shifts stabilization from 6.41 to 53.82 and commitment from 23.65 to 56.01; unwritten fraction falls 87.7 points as coverage drops 8.4.Conditional accuracy rises 1.5 points; the short-block schedule also changes answer-slot eligibility.Source: Table 3 · p. 6, §5.3 · p. 6
5.4 Visual-sensitivity summaries
- Nemotron’s mean PDM peaks occur at 25.5, 30.9, and 30.5 steps versus LaViDa’s 3.6, 0.8, and 0.7 across V*Bench, M3CoT, and ScienceQA.Final sensitivity relative to mean individual peak sensitivity is 25%, 13%, and 12% for LaViDa, versus 28%, 47%, and 45% for Nemotron; matched-canvas comparisons measure image influence given existing text.Source: Figure 3 · p. 6, §5.4 · p. 6
5.5 Direct prompting has model-dependent effects on answer coverage
- Direct prompting raises Nemotron’s overall accuracy by 15.0 points on M3CoT and 19.5 on ScienceQA, but lowers LaViDa/V*Bench accuracy by 11.0 points.For LaViDa/V*Bench, coverage falls 18.3 points and conditional accuracy falls 3.6.Source: Table 4 · p. 7, §5.5 · p. 7
6 Discussion
The discussion separates answer-candidate stabilization from permanent commitment and rationale progress, while treating visual sensitivity and answer coverage as distinct readouts. Prompting changes in accuracy are associated more with coverage than conditional accuracy.
Answer stability and rationale unfolding
- 89.4–98.1% of the rationale-side canvas remains unwritten at stabilization in single-block LaViDa decoding, while mean stabilization precedes mean commitment on all three benchmarks.The stable answer candidate persists through a substantial portion of subsequent rationale generation.Source: §6 · p. 8
- The clock-mean gap is smaller on M3CoT than ScienceQA (6.67 versus 10.22 steps), yet M3CoT has the larger unwritten fraction (98.1% versus 96.2%).The clock gap separates two answer-position events; the unwritten fraction summarizes rationale-side commitments at stabilization.Source: §6 · p. 8
Scheduling and visual sensitivity
- The short-block control reduces the unwritten fraction to 1.7%; PDM peak timing is a separate readout from answer stabilization.PDM peaks follow stabilization on all three Nemotron benchmarks, but on LaViDa precede it on V*Bench and ScienceQA and are close on M3CoT.Source: §6 · p. 8
Coverage and prompting
- Coverage is the larger absolute component of all prompting-related accuracy changes, including LaViDa/V*Bench's accuracy loss.Nemotron and DiffusionGemma accuracy rises under direct prompting, while LaViDa/V*Bench accuracy falls.Source: §6 · p. 8
7 Limitations
The block control links greater positional commitment with lower answer coverage and overall accuracy, while direct prompting has opposing accuracy effects across models.
7 Limitations
- Reducing block length lowers the unwritten fraction from 89.4% to 1.7%, while coverage falls from 100.0% to 91.6% and overall accuracy from 41.9% to 39.8%.Direct prompting raises accuracy on Nemotron and DiffusionGemma but reduces LaViDa/V*Bench accuracy by 11.0 points.Source: §7 · p. 8
8 Conclusion
The conclusion distinguishes answer stability, rationale commitment, and visual sensitivity, alongside answer extraction and the measurement choices used to characterize them.
8 Conclusion
- In single-block, EOS-suppressed LaViDa decoding, answers stabilize before most rationale-side tokens are committed.The conclusion distinguishes the order in which explanations are presented from the order in which output tokens become permanent.Source: §8 · p. 8
A.1 Answer extraction and token localization
- Extraction failures count as incorrect for overall accuracy and uncovered for valid-answer coverage; token localization can change with tokenization, whitespace, or punctuation without changing the extracted option.Scoring and trace localization are distinct operations.Source: §A.1 · p. 10
A.2 Clock computation and rationale-side region
- The unwritten fraction U is the fraction of rationale-side commitment times later than Tstab, so answer clocks alone do not determine writing progress.The timing region varies by format, and reported rationale-side fractions are averaged per example rather than over pooled tokens.Source: §A.2 · p. 10
A.3 Visual-sensitivity readout and aggregation
- PDM uses a visual-position set defined by the first </think> prefix when present, otherwise by a model-specific before/after-answer rule; matched-canvas ablation measures image influence at that canvas.Visible text may already contain image-derived information.Source: §A.3 · p. 10
B Prompting and supplementary results
The appendix specifies prompting policies and supplementary measurements, including visual-sensitivity summaries and an expansion of the accuracy decomposition.
B.1 Instruction templates
- Nemotron and DiffusionGemma receive model-specific chat templates after the benchmark question and choices, with instructions either requesting tagged reasoning or answer-only output.The two prefixes change both requested explanation behavior and response-format constraints, and use a shared scorer.Source: §B.1 · p. 11
- The prefixes alter both explanation behavior and response-format constraints, so the prompting comparison is not a change to explanation behavior alone.Both policies are evaluated with a shared scorer.Source: §B.1 · p. 11
B.2 Visual-sensitivity values
- Table 6 reports PDM timing and relative final sensitivity for recorded subsets, with peak steps measured on the denoising-step axis.Its stabilization means belong to PDM evaluations rather than the primary timing conditions in Table 1.Source: Table 6 · p. 11, §B.2 · p. 11
- Table 6's stabilization means should not be read as the primary timing-condition means reported in Table 1.The table summarizes model-specific decoding paths and uses recorded-subset sizes.Source: Table 6 · p. 11, §B.2 · p. 11
B.3 Accuracy decomposition and rounding
- Nemotron/M3CoT's two decomposition terms are 11.6586 and 3.4083 percentage points, summing to 15.0669 against a displayed change of 15.0.Across five rows, the largest absolute residual from independently rounded rates is 0.119 points.Source: §B.3 · p. 11
- Across the five prompting comparisons, the coverage-associated term is the larger absolute component of the accuracy change.This includes the LaViDa/V*Bench accuracy loss.Source: §B.3 · p. 11
C.1 Data, input construction, and decoding
The appendices describe benchmark inputs and decoding setups, compute resources, broader-impact considerations, asset terms, and checklist claims about the paper’s scope.
C.1 Data, input construction, and decoding
- V*Bench uses its question text, while M3CoT and ScienceQA combine context or hints, questions, and labeled choices; ScienceQA filters for image-bearing rows.ScienceQA answer indices are mapped to option letters.Source: §C.1 · p. 12
C.2 Compute resources
- The experiments used eight NVIDIA H100 GPUs.Principal LaViDa/Nemotron comparisons use a configured 128-token canvas and 64 denoising steps.Source: §C.2 · p. 12
C.3 Broader impacts
- The analysis reports answer stability, permanent commitment, visual sensitivity, and valid-answer coverage together, and notes performance trade-offs from schedule and instruction changes.The short-block control and LaViDa direct-prompting comparison show losses.Source: §C.3 · p. 12
C.4 Asset attribution and terms
- The appendix identifies code, model, and dataset terms, including Apache-2.0 for LaViDa and DiffusionGemma and CC BY-NC-SA for M3CoT and ScienceQA.V* project code is MIT-licensed, separately from rights to its photographs.Source: §C.4 · p. 12
NeurIPS Paper Checklist
- The checklist says the abstract and introduction reflect the measurement-focused contribution and the scope of aggregate comparisons.The checklist response is [Yes].Source: NeurIPS Paper Checklist · p. 13
2. Limitations
The limitations discussion identifies performance trade-offs from decoding and prompting choices, while the checklist notes the absence of a public execution and analysis-code repository.
2. Limitations
- The paper identifies coverage and accuracy losses in the block control and model-dependent effects of direct prompting as limitations.The checklist says these effects are discussed in Section 7.Source: §2 · p. 13
- The paper does not include a public repository for its model-execution and analysis code.The checklist says aggregate-comparison calculations and numerical inputs are specified in Section 3, Appendix A, and Tables 1–6.Source: §2 · pp. 14–15
- The main timing and prompting comparisons report point estimates, while the correctness-stratified comparison includes two-sided Mann–Whitney tests.The checklist does not report error bars for the main comparisons.Source: §2 · p. 15
10. Broader impacts
- The work analyzes existing models and datasets, introduces no new datasets or pretrained models, and does not involve recruited participants or crowdsourcing.LLM assistance was limited to routine manuscript editing and formatting.Source: §10 · pp. 18–19