PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yixiang Liu, Zhongxing Xu, Zhonghua Wang, Xiaoying Tang

arXiv:2609.11315v1cs.AI

TL;DR

Diffusion VLMs face a reasoning-budget mismatch because universal decoding lengths can over-refine stable answers or prematurely truncate reasoning. This paper introduces a training-free trajectory-based controller for LLaDA-V that routes examples among early commitment, baseline preservation, and reasoning-supportive decoding. Across answer-focused, mixed-reasoning, and CoT-sensitive benchmarks, routed control is more robust than fixed-length and single-rule alternatives.

  • Problem

    Universal decoding lengths can over-refine answer-closed questions and prematurely commit on reasoning-sensitive questions, creating a reasoning-budget mismatch.

  • Method

    A training-free controller routes each LLaDA-V example to early commitment, baseline preservation, or reasoning-supportive decoding using ground-truth-free trajectory signals.

  • Results

    Across answer-focused, mixed-reasoning, and CoT-sensitive benchmarks, routed control improves robustness over fixed long decoding, pure short decoding, and single-rule interventions.

  • Takeaways & Limitations

    Diffusion VLM decoding should route inference-time control according to the observed trajectory state rather than rely on a universal generation length.

  • Takeaways & Limitations

    The controller is evaluated on LLaDA-V, uses label-free proxies rather than correctness estimators, and is not primarily an acceleration method.

Abstract

from arXiv · show

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.

1 Introduction

Diffusion VLMs expose intermediate answer trajectories, making inference-time control possible but creating a reasoning-budget mismatch under universal decoding lengths. The paper proposes routing each example among early commitment, baseline preservation, and reasoning-supportive decoding according to trajectory evidence.

  • 1 Introduction: Diffusion VLMs generate through iterative refinement, exposing observable answer revisions that can indicate stabilization and support inference-time control.Unlike left-to-right generation, diffusion decoding produces a sequence of partially specified answer states.
  • 1 Introduction: Reasoning-budget mismatch occurs when a fixed 128-step budget over-refines answer-closed examples or prematurely commits on reasoning-sensitive examples.The appropriate response depends on the uncertainty remaining after early refinement, not simply on whether outputs should be shorter or longer.
  • 1 Introduction: The proposed training-free controller routes examples to early commitment, baseline preservation, or reasoning-supportive decoding using answer closure, commitment evidence, and revision pressure without ground-truth answers.Reasoning need is treated operationally as the control action suggested by the observed diffusion trajectory.
  • 1 Introduction: The study evaluates routed control across answer-focused, mixed-reasoning, and CoT-sensitive settings, finding greater robustness than a single global length rule.The reported evaluation is designed to test routing across differing reasoning demands.

2 Related Work

Prior adaptive decoding methods often target efficiency, uncertainty reduction, or hallucination mitigation, whereas this work treats control as a trajectory-conditioned decision problem. It distinguishes among early commitment, baseline preservation, and reasoning-supportive intervention because single-rule policies help in some regimes and fail in others.

  • 2 Related Work: Existing adaptive decoding methods commonly optimize efficiency, uncertainty reduction, or hallucination mitigation, while this work studies sample-dependent control-action mismatch.The distinction is whether internal signals serve a universal objective or guide the action appropriate to the current trajectory.
  • 2 Related Work: Pure short decoding, fixed long decoding, and revision-based control each help in some regimes and fail in others.This motivates treating internal signals as evidence for routing rather than as universal hallucination or factuality detectors.
  • 2 Related Work: The paper reframes diffusion VLM decoding as a regime-conditioned decision problem over early commitment, baseline preservation, and reasoning-supportive decoding.The system selects among actions based on trajectory state rather than applying one global length rule.

3 Method

The method routes each diffusion VLM example among early commitment, baseline preservation, and reasoning-supportive decoding using trajectory evidence rather than a universal decoding rule. It combines answer closure, visual-format and commitment evidence, representation revision pressure, and fixed guards to select the output action.

  • 3.3 Routing Evidence: The controller extracts answer closure, commitment evidence, and representation revision pressure from intermediate diffusion states to guide routing.These signals are computed from the observed trajectory without gold labels or direct correctness estimation.
  • 3.6 Routed Decoding Algorithm: Algorithm 1 computes trajectory evidence, applies the routing conditions, and returns the corresponding routed output without updating model weights or training a verifier.Cached candidate traces are restricted to offline analyses and are not used for latency claims.
  • 3.3 Routing Evidence: Answer closure measures whether the parsed answer repeatedly stabilizes within a routing window, but high closure does not imply correctness.Visual engagement contributes routing evidence through parser validity, image-token attention, prefix-dominance risk, and parsed-answer flip risk.
  • 3.3 Routing Evidence: Representation revision pressure measures disagreement between early-layer and final-layer token distributions, indicating whether internal predictions remain under revision.The signal is used for routing rather than as a hallucination detector or correctness score.
  • 3.4 Routing Policy: Routing selects among early commitment, baseline preservation, and reasoning-supportive decoding according to fixed guards and trajectory conditions.Reasoning-supportive routing is evaluated before commitment because answer stability alone is insufficient for CoT-sensitive interfaces.
  • 3.5 Routing Actions: The three actions produce distinct outputs: commitment fixes the first locally stable parsed answer, preservation returns the fixed-budget output, and reasoning support applies a fixed supportive configuration.The online commitment implementation fixes the answer field to the first stable parsed answer.

4 Experiments

Across answer-focused and CoT-sensitive settings, routed decoding improves robustness by selecting trajectory-appropriate control actions rather than applying one universal length policy. The results also show that these gains are not merely the result of uniformly shorter outputs, while auxiliary rationale diagnostics remain non-definitive for faithfulness.

  • Controlled Results Across Regimes: 79.96 on MME, 49.44 on MMMU, and 50.77 on MMStar show routed control outperforming Len128 across answer-focused benchmarks.The improvements over Len128 are 6.55, 7.00, and 0.87 points, respectively.
  • Controlled Results Across Regimes: 88.60 on ScienceQA-IMG exceeds Len128 by 13.29 points and Len2 by 4.32 points, despite the benchmark providing rationale explanations.The result concerns final-answer accuracy and does not imply that the underlying task lacks reasoning structure.
  • Controlled Results Across Regimes: 52.42 on MME-CoT and 88.75 on targeted A-OKVQA show routed control outperforming fixed long decoding while avoiding unsafe pure-short decoding.On MME-CoT, Len2 scores 48.72 versus 49.57 for Len128; on A-OKVQA, routed control exceeds Len128 at 83.75 and the strongest non-routed baseline at 85.00.
  • Length-Policy Stress Test: Routed control is not equivalent to uniformly shortening outputs: it improves over both Len2 and Len128 where short decoding removes useful reasoning structure.The length-policy stress test reports accuracies and routed average decoding steps, while MME-CoT specifically shows Len2 below Len128 and routed control above both.
  • Length-Policy Stress Test: Route allocation changes by interface and trajectory: MME favors commitment, whereas ScienceQA-IMG and MME-CoT suppress premature commitment in favor of preservation or reasoning support.This pattern supports sample-wise action selection rather than a global short policy.
  • Auxiliary CoT Diagnostics on ScienceQA-IMG: BLEU-4 rises from 0.105 to 0.124, ROUGE-L from 0.268 to 0.296, and SBERT similarity from 0.697 to 0.720 over the fixed long-budget baseline.Judge scores also indicate improved answer support while remaining competitive on factual support.
  • Auxiliary CoT Diagnostics on ScienceQA-IMG: The rationale diagnostics are auxiliary and do not establish that generated explanations faithfully explain the model’s predictions.They use reference-based metrics and judge scores rather than definitive human faithfulness evidence.

5 Conclusion

Diffusion VLM decoding exposes intermediate trajectories, so inference can route examples among commitment, preservation, and reasoning-supportive actions rather than use one global length. Across answer-focused, mixed-reasoning, and CoT-sensitive settings, this routing avoids key failures of purely short or fixed-long decoding.

  • Intermediate trajectories distinguish examples that stabilize early from those requiring additional refinement or preserved reasoning structure.
  • Trajectory-aware routing selects early commitment, fixed-budget preservation, or reasoning-supportive decoding according to observed diffusion evidence.
  • Across answer-focused, mixed-reasoning, and CoT-sensitive settings, routing avoids the main failure modes of pure short and fixed long decoding.

Limitations

The study’s limitations concern model generality, proxy-based routing evidence, benchmark-specific operating choices, CoT diagnostics, and the lack of primary acceleration benefits.

  • Evaluation on LLaDA-V alone leaves model-level generality to additional diffusion VLM studies.
  • Routing signals are label-free proxies rather than correctness estimators; visual-attention evidence measures image-token engagement, not visual truth.
  • Some operating points use benchmark-specific guards or reasoning-supportive configurations, motivating more unified threshold selection and calibration.
  • CoT diagnostics combine reference-based automatic metrics with judge-based support scores that should not be interpreted as causal evidence of rationale faithfulness.
  • Routed control is not primarily an acceleration method because some routes retain long decoding budgets for reasoning support.

A.1 Available Diagnostic Comparisons

Available diagnostics compare the full routed controller with preserve-only, commit-only, separated-source, and guarded-routing variants. The results show that routing complements preservation and commitment, especially when premature commitment can remove reasoning support.

  • Table 5 reports action-level diagnostics rather than a full component ablation, with comparisons limited to settings having corresponding artifacts.
  • The full router improves over preserve-only baselines across all reported settings.
  • Commit-only control is competitive on answer-focused settings but underperforms the full router and drops on MME-CoT.
  • On MME-CoT, the guarded route improves over the unguarded route variant, indicating reduced harmful route-on cases.
  • ScienceQA-IMG CoT separated-source control matches the full routed setting in answer accuracy, so answer accuracy alone misses source-selection effects in this diagnostic protocol.

A.2 Routing Guards and Operating Points

The routing guards encode CoT sensitivity, route-on eligibility, and hard exclusions. Reported MME-CoT thresholds and configurations are disclosed operating points rather than benchmark-independent routing rules.

  • K(x) enables a CoT-sensitive guard when prompts request rationales, interfaces score reasoning support, or disclosed configurations mark inputs as CoT-sensitive.
  • Gon(x) is the fixed route-on guard for the reported operating point.
  • χcommit(x) excludes cases such as parser unavailability, operating-point constraints, and route-specific exclusion rules.
  • For MME-CoT, Route v3 denotes a disclosed operating point rather than a benchmark-independent routing rule.
  • Cached fixed-budget and route-on outputs are used for offline trace-composition analyses, not online latency claims.

A.3 Budget and Efficiency Analysis

The budget controls show that short decoding can be effective and inexpensive for answer-closed settings, but insufficient for CoT-sensitive reasoning. Routed decoding therefore allocates or structures reasoning selectively rather than uniformly shortening outputs.

  • Fixed128 / Long128 applies the same expensive refinement budget across MME, ScienceQA-IMG, MME-CoT, and MMMU.
  • Short2 reaches 79.56 on MME, 84.28 on ScienceQA-IMG, and 48.89 on MMMU with a much smaller decoding budget.These results indicate that many answer-closed examples do not require the full fixed refinement path.
  • Route v3’s MME-CoT gain is not explained by shorter outputs because its average steps and latency are comparable to the long CoT baseline.The method selectively allocates or structures reasoning instead of simply reducing generation length.
  • Intermediate budgets do not consistently resolve the mismatch: Short32 and Fixed64 are below Short2 on MME and ScienceQA-IMG and do not improve over Fixed128 on MME-CoT.

A.4 Shared Decoding Settings

The experiments use a frozen LLaDA-V with shared deterministic decoding and benchmark parsers, while fixed-budget and adaptive controls differ in their decoding rules. Routing policies are fixed before evaluation, and diagnostic analyses are separated from online efficiency claims.

  • Shared Decoding Settings: All reported rows use the frozen LLaDA-V model, benchmark image inputs and prompts, deterministic temperature-0 decoding, low-confidence remasking, and the corresponding benchmark parser.
  • Shared Decoding Settings: Fixed-budget controls keep the model, prompt, and parser unchanged while varying generation lengths matched to 2, 32, 64, or 128 denoising steps.
  • Routing Configuration: Routing rules are fixed before evaluation and do not use gold labels; answer-focused settings choose commitment or preservation, while CoT-sensitive settings can enable a reasoning route for unresolved trajectories.
  • Evaluation Scope: Cached candidate outputs support diagnostic composition only and are excluded from online latency claims, preserving the shared model, prompt, parser, and branch-level configuration.
  • Rationale Evaluation: The text-only judge scores factual and answer support from the question, choices, reference explanation, predicted answer, and rationale, without inspecting images or establishing causal rationale faithfulness.
  • Diagnostic Analyses: Illustrative ScienceQA-IMG and MME cases diagnose rationale quality or answer-token drift, but do not constitute additional benchmark results.
Loading 2609.11315v1…