acceptodds
Under review as a conference paper at ICLR 2027

EvidenceFlow: Repaying Visual Evidence Debt for Test-Time Multimodal Reasoning

Abstract

Test-time scaling can improve multimodal reasoning without updating model parameters, but longer reasoning may not provide the visual evidence required by later steps. Existing visual refocusing methods replay image tokens based on their relevance to the current reasoning state, without examining whether realized visual support matches that state. We study the resulting mismatch between the current visual-demand distribution and the realized visual-support distribution. We define this mismatch as Visual Evidence Debt, quantified as one minus the overlapping probability mass of the two distributions. Its equivalent positive-residual form localizes visual tokens whose demand share exceeds their realized-support share. We therefore propose EvidenceFlow, a training-free framework that performs marginal debt repayment under a fixed replay budget via a saturating monotone submodular objective. Experiments and controlled analyses across multiple MLLMs and benchmarks show improved reasoning accuracy over standard thinking and a more favorable accuracy–compute trade-off than the evaluated replay baselines. Our code is available in the supplementary material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multimodal large language models, multimodal reasoning, test-time scaling, visual refocusing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9mckOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces EvidenceFlow, a training-free visual replay framework for test-time multimodal reasoning. The design idea of EvidenceFlow is to distinguish between the visual evidence currently required by the reasoning process and the visual support that has actually been realized. The authors thus model multimodal reasoning as a mismatch between current visual demand and history-aware visual support. They define this mismatch as Visual Evidence Debt and uses it to guide visual token replay. Experiments on three MLLMs and five benchmarks show consistent improvements in both reasoning accuracy and visual-compute efficiency.

Critical strengths
  1. The problem formulation is clear and quite intuitive. The distinction between current visual demand and realized visual support provides a useful way to characterize why long multimodal reasoning trajectories may gradually lose visual grounding.

  2. The empirical results are strong. EvidenceFlow improves over standard thinking and other replay baselines across all three evaluated models and five benchmarks, suggesting that the method is reasonably robust across different model families and tasks. The ablations are also substantial.

Critical weaknesses
  1. The Visual Evidence Debt measure is equivalent to the total variation distance between two normalized distributions. Moreover, the (1−1/e)(1-1/e) approximation guarantee in Proposition 1 follows from the well-established result of Nemhauser et al. (1978). Thus, the novelty seems to lie primarily in the task-specific formulation and integration of components rather than in new theoretical developments.

  2. While the paper provides validation for these estimators, it is still less clear whether the same conclusions would hold under alternative definitions of QQ and PP.

  3. The experiments focus mainly on visual QA and reasoning benchmarks, also limited at 8B scale models. It would be useful to understand whether the method generalizes to larger models or broader multimodal settings.

Questions for authors
  1. How sensitive is the method to alternative choices for the visual-demand and realized-support estimators?

  2. I see that Appendix B.7 provides useful selector comparisons. Could the authors further clarify when clause-level saturation offers the greatest benefit over simpler residual-based selection?

  3. Do similar results also hold on substantially larger multimodal models, or for settings such as multi-image reasoning?

Reviewer_3FrrOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvidenceFlow replays selected visual tokens during multimodal reasoning without additional training. It estimates visual demand and history-aware support, then selects tokens to reduce a normalized mismatch called Visual Evidence Debt. The selection objective is monotone submodular, with a greedy guarantee for the fixed-budget surrogate. Experiments cover three backbones and five benchmarks, with additional masking, replay-history, and fixed-target likelihood analyses. The theoretical guarantee concerns token selection before replay; it does not guarantee the subsequent transformer response.

Critical strengths
  1. The mathematical formulation is clear, and its main identities are internally consistent. The overlap, positive-residual, and total-variation expressions agree, and the capped coverage objective supports the stated fixed-budget guarantee (Equations 9-15). I also appreciate the masking and fixed-target likelihood tests, which check effects beyond the method's own debt metric.

  2. Table 2 gives a concrete accuracy-cost trade-off for fixed and adaptive replay on Qwen3-VL-8B-Thinking. Adaptive replay improves accuracy from 65.1% to 66.6% over fixed replay, while reducing latency from 9.3 to 8.8 seconds. The replay ratio falls from 30.0% to 18.4% per triggered event, a 38.7% reduction. The reported latency reduction is about 5.4%, alongside changes in triggering and token selection. These results support the adaptive configuration, without isolating the runtime contribution of token savings alone.

Critical weaknesses
  1. The supplementary replay implementation remains difficult to check end to end. The manuscript already specifies embedding reuse, spatial order, cache retention, and continuation positions (p. 15). However, all three submitted backbone adapters raise NotImplementedError in the replay function. A working example would help verify that the implementation follows the specified protocol and reported timing. This is a reproducibility limitation, without establishing that the reported results are incorrect.

  2. The practical comparison with native VISREF remains limited, as the paper explicitly acknowledges. Table 1 uses externally reported native results, while Table 2 and Appendix B.7 provide matched selector controls, including an independent fixed schedule. Those controls support selector quality but do not establish an end-to-end advantage over the complete native policy. The reference also matters: adaptive EvidenceFlow takes 8.8 seconds at 66.6% accuracy, versus 8.1 seconds at 60.0% for Standard Thinking.

Questions for authors
  1. Could you provide one working replay example that implements the cache and position protocol already specified on p. 15? Please clarify which reported experiments are runnable with the submitted supplementary package.

  2. The paper clearly labels the matched VISREF comparison as selector-only. Are any same-environment results for the complete native policy already available? Such evidence would affect my assessment of the practical accuracy-latency advantage, without being a prerequisite for the selector-level contribution.

I will further consider my score in light of the other reviewers’ comments and whether the authors’ response adequately addresses my concerns.

1 comment

Thank you for the detailed explanations and the additional latency experiments. After reading the other reviewers’ comments, particularly those of Reviewer 9mck, I continue to have reservations about the method’s novelty. Specifically, Visual Evidence Debt appears to be mathematically equivalent to total variation distance, while the submodular objective and its approximation guarantee draw on established theoretical results. Although the formulation offers an intuitive perspective on visual replay and the additional experiments strengthen the empirical evidence, the methodological contribution beyond task-specific formulation and integration remains insufficiently clear.

Furthermore, the supplementary code does not appear to provide a complete, executable replay implementation, making it difficult to independently reproduce the reported improvements. This limits my ability to assess the method’s practical utility, while its generalizability to broader settings, such as multi-image reasoning, remains unverified.

I appreciate the authors’ efforts and the additional evidence provided during the rebuttal. However, given these remaining concerns, I will maintain my original score.

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

The paper proposes a concept of visual evidence debt that is captured as the required visual evidence support thus far in the reasoning sequence to the accumulated evidence thus far by measuring text and visual token similarity and corresponding visual token normalized attention scores. Then the paper uses this visual evidence debt to guide the token selection process to inject back into the reasoning steps for improved test-time scaling accuracy for multi-modal reasoning tasks.

Critical strengths

The method is demonstrated to be effective over the datasets shown in experiments on multi-modal reasoning. The framing of treating visual token selection as a way to alleviate the gap of desired visual token evidence or relevance to the actual accumulated ones in the reasoning process is interesting. The proposed visual token re-injection method does not introduce additional inference efficiency latency compared to other methods.

Critical weaknesses
  1. Clarity of method formulation — Several notations in the visual evidence debt formulation are insufficiently explained in the main text, with important details deferred to the appendix (e.g., the normalized attention weights in Eq. 7). The relevance-only replay baseline is also not clearly defined in the main experimental section.

  2. Justification of visual evidence debt — The proposed metric measures the discrepancy between two independently normalized distributions derived from different model signals. While the appendix provides supporting experiments, it remains unclear whether this discrepancy faithfully quantifies insufficient visual evidence. More direct validation of the metric itself, or comparisons against alternative discrepancy measures, would strengthen the motivation.

  3. Robustness of method to clause segmentation — The method relies on heuristic segmentation of reasoning traces into rounds and clauses. It is unclear how sensitive the results are to these segmentation choices, particularly given the importance of clause-level demand modeling and saturation in the proposed objective.

Questions for authors

Please address the points mentioned in the weakness section.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their constructive feedback. In response to their comments, we have made several targeted revisions to the paper:

  • Methodological Clarification and Justification: We further clarify the definition of Visual Evidence Debt, how it differs from relevance-based visual replay methods, and the role of the clause-level saturation mechanism. We also emphasize our simple, training-free design, which achieves effective visual evidence replay through demand–support modeling and budget allocation without introducing extra model complexity or additional training.
  • Additional Empirical Validation: We add new experiments that directly verify the proposed demand–support discrepancy, evaluate the robustness of both the demand and support estimators, and compare different token selection strategies under controlled replay budgets. These results further confirm the effectiveness of our evidence-debt-driven visual replay mechanism.
  • Generalization and Efficiency: We scale our experiments to a larger backbone model (Qwen3-VL-32B-Thinking) and add multi-image reasoning evaluations on the MV-MATH benchmark. In addition, we clarify the implementation details of visual replay and provide a direct accuracy-and-latency comparison against VisRef.
  • Code Availability: To protect ongoing research and follow-up work, we are not releasing the full codebase at this stage. However, we have included a runnable demo of the core visual replay mechanism in the supplementary material, which clearly demonstrates the exact execution flow of our method.

We believe these targeted revisions address the reviewers' main concerns and further strengthen both the clarity and empirical support of EvidenceFlow. We thank the reviewers again for their valuable suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.