acceptodds
Withdrawn from ICLR 2027

Distilling One Token is Enough: Rethinking On-Policy Distillation for Visual Perception in MLLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

23 views

Abstract

On-policy distillation (OPD) has recently emerged as an effective approach for improving visual perception in multimodal large language models (MLLMs) by aligning teacher and student predictive distributions over student-generated trajectories. In visual perception tasks, these trajectories typically contain descriptive tokens that verbalize visual evidence and decision tokens that produce the final answer. This raises a fundamental question: is dense token-level supervision across the entire trajectory necessary? In this work, we find substantial redundancy in OPD supervision for visual perception. Distilling only a single decision token or only the descriptive tokens on multiple-choice visual perception data achieves downstream performance comparable to full-response distillation. Moreover, replacing the teacher distribution at the decision position with a hard ground-truth label fails to reproduce the improvement, highlighting the importance of the teacher's soft predictive distribution. Motivated by these findings, we propose **D**ecision-**O**nly **D**istillation (**DOD**), a simple, rollout-free distillation strategy that uses a direct-answer prompt and matches teacher and student distributions only at the first output position. We further introduce target-vocabulary distillation, which restricts distribution matching to the candidate decision tokens (e.g., A, B, C, and D) rather than the full vocabulary. Consequently, each training example requires distillation at only one token position over only a handful of vocabulary entries, eliminating rollout and reducing distillation costs while achieving better overall visual perception performance than standard OPD. Code is available at the [anonymous link](https://anonymous.4open.science/r/dod_code-6337).

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
On-policy Distillation, MLLMs, Visual Perception

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_bQH7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that dense token-level supervision are redundant for OPD in visual. The authors show that supervising only the first decision token can retain most of the gains of full-response OPD. Based on this observation, they propose DoD, which performs rollout-free soft distillation at the first output position. Experiments show comparable or slightly better accuracy than Vision-OPD with substantially lower training cost.

Critical strengths

[+]: This paper identifies a clear and important question about the necessity of full OPD trajectories.

[+]: Provides useful comparisons among full-response, decision-only, description-only, and hard-label supervision. Reports meaningful reductions in training time, GPU requirements, and CPU memory.

[+]: The proposed DOD method is simple, computationally efficient, and easy to reproduce.

[+]: Clear writing.

Critical weaknesses

[-]: I question the real-world applicability of DoD. Tailoring MLLMs exclusively for decision tasks (i.e., select a b c or d) through specialized designs is of marginal significance; furthermore, it may inadvertently degrade the model's broader reasoning and generalization abilities.

[-]: Improvements over Vision-OPD are relatively small and inconsistent across individual benchmarks.

[-]: How can it be guaranteed that the cropped images necessarily contain additional information? The cropping strategy seems to fail in scenarios involving multiple objects or abstract concepts (e.g., color, size).

[-]: DoD is essentially an OPSD method, and its utility relies on the premise that the cropped images can provide extra information to the teacher, and the student can gradually learn to discover that region after OPSD. This raises a concern: for extremely small objects in the original image, forcing the student to discover and perceive that region via OPSD appears overly challenging, even meaningless.

[-]: How is the stability of DoD's output ensured?

[-]: The distinction from recent sparse-token and early-stopping OPD methods should be made more explicit.

Questions for authors

See weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer bQH7 (Part 1/2)

W1: Generalization Beyond Multiple-Choice Tasks and Preservation of Reasoning Ability

We thank the reviewer for raising this concern. We would like to clarify that the multiple-choice format is required only during training, while DOD does not restrict the model to multiple-choice outputs during inference. To address the concern regarding real-world applicability and potential degradation of broader capabilities, we provide evidence from two perspectives.

1. Generalization to non-MCQ tasks. Our evaluation is not limited to multiple-choice benchmarks. Specifically, 224 of the 845 ZoomBench questions and 2,085 of the 2,612 DYSM questions are non-MCQ questions. We further evaluate DOD on OCRBench, which consists of 1,000 open-ended questions requiring textual answers. The results are summarized below.

ModelDYSM non-MCQ (%)Zoombench non-MCQ (%)OCRBench (%)
Base-4B59.6636.1685.1
Vision-OPD-4B60.4844.6485.1
DOD-4B64.9945.9885.5
Base-9B64.7038.8487.2
Vision-OPD-9B69.2647.7787.5
DOD-9B70.1745.9888.1

The results demonstrate that DOD consistently improves over the base models on all three benchmarks at both model scales, despite being trained exclusively with multiple-choice supervision. These findings indicate that the benefits of DOD are not confined to selecting predefined answer options, but can generalize to open-ended visual perception tasks.

2. Preservation of visual reasoning capabilities.

We further investigate whether decision-only distillation compromises the model's broader reasoning capabilities. Although DOD primarily targets visual perception, stronger perception can also benefit downstream visual reasoning by providing more accurate visual evidence for subsequent inference.

To examine this, we additionally evaluate DOD on three visual reasoning benchmarks: MathVision, LogicVista, and WeMath, using thinking mode.

ModelMathVision (%) ↑LogicVista (%) ↑WeMath (%) ↑
Base-4B73.4571.3670.57
Vision-OPD-4B71.3869.3568.76
DOD-4B73.4971.5975.43
Base-9B77.6674.7280.86
Vision-OPD-9B77.8074.0581.24
DOD-9B78.4276.7381.71

The results show that DOD consistently improves over the base models across all three visual reasoning benchmarks at both model scales. Notably, DOD-4B achieves a 4.86 percentage-point improvement on WeMath, while DOD-9B improves LogicVista by 2.01 percentage points. In contrast, although Vision-OPD improves visual perception performance, its gains do not consistently translate into better visual reasoning performance, with performance degradation observed on several benchmarks. These results suggest that DOD can effectively translate improvements in visual perception into downstream visual reasoning gains, without compromising the model's reasoning performance.

W2: Performance Gains over Vision-OPD and Training Efficiency

We acknowledge that the overall improvements over Vision-OPD are relatively modest, with some variations across benchmarks. However, DOD achieves notable gains on several benchmarks, including HRBench 8K (+3.50 percentage points) and DYSM (+3.94 percentage points) for 4B, and V* (+6.29 percentage points) for 9B. Moreover, our additional visual reasoning experiments (W1.2) demonstrate more consistent advantages of DOD. More importantly, our primary contribution is achieving comparable or better performance at substantially lower training cost, rather than maximizing accuracy gains.

W3: Clarification of Privileged Visual Information

We thank the reviewer for raising this concern. We would like to clarify that we do not claim cropped images necessarily contain additional semantic information. Rather, following Vision-OPD, we refer to cropped images as privileged visual information, as they can provide the teacher with a clearer, higher-resolution view of certain visual details that may be difficult to perceive in the original image. The goal of this setting is to improve the student's visual perception capabilities by distilling knowledge from a teacher with enhanced access to visual details.

AuthorsOpenReview ↗

Response to Reviewer bQH7 (Part 2/2)

W4: Limitations of Privileged Visual Information

We thank the reviewer for raising this concern. We acknowledge this limitation, as extremely small objects may contain visual details that are difficult or even impossible for the student to perceive from the original image. However, the overall goal of privileged-information distillation is to improve the student's ability to extract and utilize visual details. Importantly, this limitation is shared with Vision-OPD, as DOD adopts the same privileged-information setting. Our primary contribution is to make this distillation process more efficient, rather than addressing the inherent limitations of image cropping.

W5: Output Stability and Potential Collapse

We interpret “output stability” as whether decision-only distillation causes the model to collapse into an option classifier or lose its broader generation capabilities.

DOD does not renormalize the candidate-option probabilities to sum to one. Instead, it preserves their full-vocabulary probabilities and matches the aggregate probability mass of all non-option tokens through a tail bucket. Thus, the objective does not explicitly force all probability mass onto the answer options. Moreover, our experiments demonstrate that DOD outperforms Vision-OPD in five of the six non-MCQ comparisons and on all three visual reasoning benchmarks at both model scales, providing empirical evidence that decision-only distillation does not compromise the model's broader generation and reasoning capabilities.

W6: Distinction from Sparse-Token and Early-Stopping OPD Methods

We thank the reviewer for this comment. First, existing sparse-token and early-stopping OPD methods primarily focus on LLM reasoning, while their applicability to MLLM visual perception remains relatively underexplored. More importantly, although these methods have shown that distilling only a subset of tokens can achieve performance comparable to full-response distillation, they still require generating intermediate reasoning steps (i.e., rollouts) during training. In contrast, our key finding is that, for MLLM visual perception tasks, full-response rollouts can be eliminated entirely while maintaining or even improving performance. This enables DOD to perform rollout-free distillation using only the first decision token, resulting in substantially lower training costs. The elimination of rollouts is the fundamental distinction from prior methods, extending beyond simply reducing the number of supervised tokens.

AuthorsOpenReview ↗

Have our responses adequately addressed your concerns?

Dear Reviewer bQH7,

We sincerely appreciate your time and effort in reviewing our manuscript. Your constructive feedback has been invaluable in helping us improve the clarity and quality of our work.

We have provided detailed responses to your comments and concerns, and we hope they have adequately addressed your questions. If you find our responses satisfactory, we would greatly appreciate your consideration of adjusting your score accordingly. Should you have any remaining questions or concerns, we would be happy to provide further clarification and engage in additional discussion.

Best regards, Authors

Reviewer_dxiPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Decision-Only Distillation (DOD), which eliminates rollouts by matching soft teacher predictions at the first answer position and aggregating non-candidate vocabulary entries. The problem is practically relevant, but the methodological novelty is limited, and the experimental design does not adequately support the central claims.

Critical strengths
  1. The paper addresses the practical issue of reducing training costs in visual perception distillation.
  2. The method is straightforward to implement and appears reasonably efficient under the reported settings.
Critical weaknesses
  1. The method lacks novelty. In multiple-choice tasks, soft-label distillation over answer options is a natural application of classification distillation. Given the existing literature, the authors need to explain clearly what this paper teaches us that was not previously known.

  2. I question the generalizability of the conclusions. The proposed training approach may overly concentrate the student’s probability distribution on the four answer options, potentially leading to output distribution collapse. I am concerned that the distilled models may lose reasoning capabilities and suffer substantial degradation on open-ended question-answering tasks.

  3. The methodological formulation contains a conceptual conflation. Section 3 presents off-policy distillation using forward KL and on-policy distillation using reverse KL, potentially confusing two independent design dimensions. The off-policy/on-policy distinction concerns whether the supervision prefixes originate from the current student, whereas the KL direction determines how teacher and student distributions are matched. The current presentation may mislead readers.

  4. The experimental comparisons are insufficiently controlled, and the improvements are marginal. The main issue is the lack of controlled ablations supporting the central claim that full-response supervision is redundant. Direct-answer distillation and full-response OPD simultaneously differ in prompts, supervision scope, and loss weighting. Their comparable performance therefore cannot be attributed specifically to removing descriptive supervision. Moreover, DOD improves over Vision-OPD by only 1.22 and 1.27 percentage points for the two model sizes, respectively.

  5. The presentation requires substantial improvement.

    • Sections 4 and 5 contain considerable repetition. Section 4 already introduces rollout-free direct-answer distillation and its complete objective, while Section 5.1 describes the same mechanism again.
    • The naming and relationships among baselines increase the reading burden. Original Vision-OPD, the modified answer-first Full-response OPD with 50/50 loss weighting, Direct-answer Distillation, and DOD do not constitute a straightforward sequence of component-wise ablations. Although some differences are explained, the paper lacks a consolidated table specifying prompts, response order, loss weights, matched vocabulary entries, and rollout counts, making the comparisons easy to misinterpret.
    • The contribution narrative lacks a consistent focus. The introduction emphasizes rethinking the necessity of intermediate generation in OPD, the method section emphasizes DOD/TVD, and the experiments highlight performance and efficiency gains over Vision-OPD. These could form a coherent narrative, but the paper does not adequately distinguish the contributions of the empirical findings, the simplified training recipe, and the additional technical component.
    • Different aggregation metrics create apparent inconsistencies. Section 4 reports sample-weighted overall accuracy, whereas the main experiments report the arithmetic mean across benchmarks. Both are reasonable, but the distinction should be highlighted to prevent readers from directly comparing results around 70% and 74%.
Questions for authors

Overall, the paper presents a useful training simplification, but neither its methodological novelty nor the evidence supporting its central interpretation is sufficiently established. To save reviewers’ time, I do not recommend submitting a rebuttal.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dxiP (Part 1/2)

W1: Novelty and Key Contributions of DOD

We thank the reviewer for this comment. We agree that soft-label distillation over answer options is not a new technique in itself. However, the key contribution of our work is not the introduction of a new distillation loss, but rather a new empirical finding that challenges the necessity of full-response supervision in visual perception distillation.

Specifically, while conventional OPD relies on student-generated trajectories and dense token-level supervision, we find that distilling only the first decision token can achieve comparable or even better performance on visual perception tasks. This finding is non-trivial, as it suggests that intermediate textual supervision is not always necessary for transferring visual perceptual knowledge.

More importantly, this finding enables us to eliminate autoregressive rollouts entirely, rather than merely reducing the number of supervised tokens. Based on this insight, DOD provides a simple yet effective training recipe that achieves competitive or superior performance to Vision-OPD while substantially reducing training cost.

Therefore, our primary novelty lies in identifying the redundancy of full-response distillation for visual perception and translating this finding into a more efficient distillation paradigm.

W2: Generalizability Beyond MCQ Tasks and Preservation of Reasoning Capabilities

We thank the reviewer for raising this concern. We would like to clarify that DOD uses multiple-choice supervision during training, but does not restrict the model to multiple-choice outputs during inference. We provide additional evidence of its generalizability from two perspectives.

1. Generalization to non-MCQ tasks. Our evaluations are not limited to MCQ tasks. ZoomBench contains non-MCQ questions (224/845), while the majority of questions in DYSM are non-MCQ (2,085/2,612). We further conduct additional experiments on OCRBench, which requires open-ended textual answers.

ModelDYSM non-MCQ (%)Zoombench non-MCQ (%)OCRBench (%)
Base-4B59.6636.1685.1
Vision-OPD-4B60.4844.6485.1
DOD-4B64.9945.9885.5
Base-9B64.7038.8487.2
Vision-OPD-9B69.2647.7787.5
DOD-9B70.1745.9888.1

DOD consistently outperforms both base models across all three benchmarks, with gains of 5.33/5.47 points on DYSM non-MCQ and 9.82/7.14 points on ZoomBench non-MCQ for 4B/9B, respectively. These findings demonstrate that the benefits of DOD extend beyond multiple-choice answer selection to open-ended visual perception tasks.

2. Preservation of visual reasoning capabilities.

We further investigate whether decision-only distillation compromises the model's broader reasoning capabilities. Although DOD primarily targets visual perception, stronger perception can also benefit downstream visual reasoning by providing more accurate visual evidence for subsequent inference. We evaluate DOD on three visual reasoning benchmarks: MathVision, LogicVista, and WeMath, using thinking mode.

ModelMathVision (%) ↑LogicVista (%) ↑WeMath (%) ↑
Base-4B73.4571.3670.57
Vision-OPD-4B71.3869.3568.76
DOD-4B73.4971.5975.43
Base-9B77.6674.7280.86
Vision-OPD-9B77.8074.0581.24
DOD-9B78.4276.7381.71

The results demonstrate that DOD consistently improves over the base models and Vision-OPD across all three visual reasoning benchmarks at both model scales. Notably, DOD-4B achieves a 4.86 percentage-point improvement on WeMath, while DOD-9B improves LogicVista by 2.01 percentage points. In contrast, although Vision-OPD improves visual perception performance, its gains do not consistently translate into better visual reasoning performance, with performance degradation observed on several benchmarks. These results suggest that DOD can effectively translate improvements in visual perception into downstream visual reasoning gains, without compromising the model's reasoning performance.

W3: Clarification of On-Policy/Off-Policy Distillation and KL Direction

We thank the reviewer for pointing out this ambiguity. We agree that the on-policy/off-policy distinction and KL direction are independent dimensions. Section 3 presents forward and reverse KL as common formulations of off-policy and on-policy distillation, respectively. We will clarify this distinction in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer dxiP (Part 2/2)

W4: Controlled Ablation and Performance Gains

We thank the reviewer for highlighting the need for more controlled comparisons.

1. Controlled ablation of descriptive-token supervision.

We agree that the original comparison between direct-answer distillation and full-response OPD involves differences in prompting and generation. To address this concern, we conducted an additional controlled experiment using exactly the same setting as full-response OPD, with the only change being the removal of the descriptive-token distillation loss. The results on Qwen3.5-4B are summarized below:

BenchmarkFull-response OPDDecision-only OPD (Controlled)
V*Bench91.1088.48
ZoomBench57.5157.99
HRBench 4K80.0080.00
HRBench 8K77.0078.50
MME-RealWorld EN70.7071.07
MME-RealWorld CN69.5869.87

Under this strictly controlled setting, decision-only OPD achieves the same average accuracy as full-response OPD (74.32%), outperforming it on four of the six benchmarks, matching it on HRBench 4K, and performing worse on V*Bench. These results provide more direct evidence that full-response supervision is largely redundant for visual perception distillation in this setting. We will replace the original answer-token distillation experiment in the revised paper with this more rigorously controlled decision-only OPD setting.

2. Performance gains over Vision-OPD and training efficiency.

We acknowledge that the average improvements over Vision-OPD are relatively modest. Nevertheless, DOD achieves notable gains on several benchmarks, including HRBench 8K (+3.50 percentage points) and DYSM (+3.94 percentage points) for 4B, and V*Bench (+6.29 percentage points) for 9B. Moreover, our additional visual reasoning experiments (W2.2) demonstrate more consistent advantages of DOD. More importantly, our primary contribution is achieving comparable or better performance at substantially lower training cost, rather than maximizing accuracy gains.

W5: Presentation and Organization

We thank the reviewer for the constructive suggestions. We will improve the manuscript's organization and presentation to address these concerns.

1. Repetition between Sections 4 and 5.

We agree that the original presentation contains unnecessary repetition. As discussed in our response to the experimental concerns, we have replaced the original direct-answer distillation experiment in Section 4 with a strictly controlled decision-only OPD setting, which retains the same prompt as full-response OPD while removing only the descriptive-token loss. Consequently, Section 4 now focuses on investigating the necessity of full-response supervision, whereas Section 5 introduces DOD, which further eliminates rollouts and incorporates target-vocabulary distillation. This revision establishes a clearer distinction between the empirical investigation and the proposed method.

2. Clarification of baseline settings.

To improve clarity, we provide the following consolidated comparison and will include it in the revised paper.

MethodPromptResponse OrderLoss Weight (Decision / Description)Matched VocabularyRollouts
Vision-OPDOriginalOriginal generationUniform token-level averagingTop-100 + tail8
Full-response OPDAnswer-then-explainDecision → Description0.5 / 0.5Top-100 + tail8
Decision-only OPDAnswer-then-explainDecision → Description1 / 0Top-100 + tail8
Description-only OPDAnswer-then-explainDecision → Description0 / 1Top-100 + tail8
Direct-answer distillationDirect-answerDecision only1 / 0Top-100 + tail0
DOD (Ours)Direct-answerDecision only1 / 0Answer options + tail0

3. Consistency of the contribution narrative.

We appreciate this suggestion and will revise the paper to present a more coherent progression of contributions:

  • Empirical finding (Section 4): We demonstrate that full-response supervision is largely redundant for visual perception distillation through controlled comparisons.
  • Simplified training strategy (Section 5.1): Motivated by this finding, we eliminate intermediate generation and distill only the first decision position, resulting in a rollout-free training procedure.
  • Technical refinement (Section 5.2): We introduce target-vocabulary distillation to focus distribution matching on decision-relevant tokens. The experiments in Section 6 then validate the effectiveness and computational efficiency of the resulting method.

4. Consistency of aggregation metrics.

Thanks for raising this point. We will report the arithmetic mean across the six benchmarks in both Section 4 and the main experiments.

AuthorsOpenReview ↗

Have our responses adequately addressed your concerns?

Dear Reviewer dxiP,

Thank you again for your thoughtful and constructive review. Your feedback on the generalizability, experimental controls, and presentation has been invaluable in helping us improve our manuscript.

We hope our responses and additional experiments have adequately addressed your concerns. If you have any remaining questions, we would be happy to provide further clarification. If you find our responses satisfactory, we would greatly appreciate your consideration of adjusting your score accordingly.

We sincerely look forward to your feedback.

Best regards, Authors

Reviewer_cuK4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper shows whether full-response supervision is necessary for on-policy distillation in visual perception tasks and finds that distilling only the first answer token can retain most of its benefits. Based on experiments separating decision-token and descriptive-token supervision, the authors propose DOD, which directly matches the student’s and privileged-view teacher’s soft distributions at the first output position without generating a rollout. DOD further restricts matching to the candidate answer tokens and an aggregated bucket through target-vocabulary distillation. Experiments with Qwen3.5-4B and 9B show improved average perception accuracy over baselines.

Critical strengths
  • The paper is well written, well organized, and easy to understand. Its research question is clearly motivated, and the progression from the empirical analysis to the proposed method is natural.

  • DOD is notably simple and practically useful, eliminating rollout generation while retaining performance. The authors also provide several relevant ablations, evaluations at two model scales, and sufficient implementation details for reproduction.

Critical weaknesses
  • It is already widely known that in vision-language reasoning tasks, certain types of tasks are hurt by textual reasoning, while other types are benefited from textual reasoning before answering. Thus, the paper should more clearly clarify which types of tasks benefit from this.

  • The proposed method seems to be restricted to multiple-choice setting. Training requires explicit candidate answers represented by single tokens such as A–D. However, such setting is not practical and more aligned with how benchmarks are constructed and not how users will typically interact with vision language models. The resulting conclusion should therefore be restricted to direct-recognition multiple-choice tasks rather than visual perception generally.

Questions for authors

See weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer cuK4 (Part 1/2)

W1. Clarifying the scope of the proposed method.

We thank the reviewer for raising this important point. We would like to clarify the scope and motivation of our work.

1. Visual perception vs. visual reasoning. MLLM tasks can be broadly categorized into two types: visual perception and visual reasoning. For visual perception tasks, answers can often be determined directly from visual evidence without requiring explicit multi-step textual reasoning. In such cases, generating additional reasoning steps may not necessarily improve performance and can sometimes even be detrimental. This motivates our central research question: Is full-response supervision necessary for on-policy distillation in visual perception tasks? In contrast, visual reasoning tasks often involve mathematical reasoning or multi-step logical inference grounded in visual information, where intermediate reasoning steps can be essential. This distinction is also supported by our experiments on reasoning tasks, where we find that distilling only the final answer tokens, while omitting intermediate reasoning tokens, does not improve reasoning performance. We therefore emphasize that our work focuses specifically on improving visual perception rather than general visual reasoning. Accordingly, the main benchmarks in our paper are selected to evaluate visual perception capabilities.

2. Improvements on visual reasoning tasks. Although our method primarily targets visual perception, stronger perception capabilities can also benefit downstream visual reasoning by providing more accurate visual evidence for subsequent inference. A recent study [1] has shown that a substantial proportion of errors in visual reasoning tasks stem from inaccurate extraction or interpretation of visual evidence, rather than deficiencies in reasoning alone. To investigate this, we additionally evaluate DOD on three visual reasoning benchmarks: MathVision, LogicVista, and WeMath, using thinking mode. The results are summarized below.

ModelMathVision (%) ↑LogicVista (%) ↑WeMath (%) ↑
Qwen3.5-4B73.4571.3670.57
Vision-OPD-4B71.3869.3568.76
DOD-4B73.4971.5975.43
Qwen3.5-9B77.6674.7280.86
Vision-OPD-9B77.8074.0581.24
DOD-9B78.4276.7381.71

The results show that DOD consistently improves over the base models across all three visual reasoning benchmarks at both model scales. Notably, DOD-4B achieves a 4.86 percentage-point improvement on WeMath, while DOD-9B improves LogicVista by 2.01 percentage points. In contrast, although Vision-OPD improves visual perception performance, its gains do not consistently translate into better visual reasoning performance, with performance degradation observed on several benchmarks. These results suggest that DOD can more effectively translate improvements in visual perception into downstream visual reasoning gains, without compromising the model's reasoning performance.

Reference

[1] Wu, J., et al. "From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models." ICML 2026.

AuthorsOpenReview ↗

Response to Reviewer cuK4 (Part 2/2)

W2. Generalization beyond multiple-choice tasks.

We thank the reviewer for raising this concern. We would like to clarify that the multiple-choice format is required only during training, while downstream inference is not restricted to MCQ tasks. To further clarify this distinction, we report additional results on non-MCQ questions, including 224 non-MCQ questions out of 845 ZoomBench questions, 2,085 non-MCQ questions out of 2,612 DYSM questions, and all 1,000 open-ended questions from OCRBench.

ModelDYSM non-MCQ (%)Zoombench non-MCQ (%)OCRBench (%)
Base-4B59.6636.1685.1
Vision-OPD-4B60.4844.6485.1
DOD-4B64.9945.9885.5
Base-9B64.7038.8487.2
Vision-OPD-9B69.2647.7787.5
DOD-9B70.1745.9888.1

As shown above, DOD consistently improves over the base models across all three non-MCQ evaluations at both model scales. In particular, DOD improves accuracy on the DYSM non-MCQ subset by 5.33 and 5.47 percentage points, and on the Zoombench non-MCQ subset by 9.82 and 7.14 percentage points, for 4B and 9B base model, respectively. These results provide direct empirical evidence that the benefits of MCQ-based decision-only distillation can transfer to open-ended visual perception tasks.

Moreover, MCQ-formatted training data can be automatically constructed from existing open-ended VQA datasets. For example, AutoConverter [2] can automatically transform open-ended VQA questions into challenging MCQs. This demonstrates the practical feasibility of constructing MCQ-formatted training data at scale.

Therefore, while DOD adopts an MCQ-based training objective, its effectiveness is not limited to multiple-choice inference settings. We will clarify this distinction and include the additional experimental results in the revised paper.

Reference

[2] Zhang, Y., et al. "Automated Generation of Challenging Multiple-Choice Questions for Vision Language Model Evaluation." CVPR 2025.

AuthorsOpenReview ↗

Have our responses adequately addressed your concerns?

Dear Reviewer cuK4,

Thank you again for your time and effort in reviewing our paper. Your constructive feedback has been invaluable in helping us improve our work.

We hope that our responses have adequately addressed your concerns. If you have any remaining questions or concerns, we would be happy to provide further clarification and engage in additional discussion. We would also greatly appreciate your consideration of adjusting your score in light of our responses and revisions.

We sincerely look forward to your feedback.

Best regards,

Authors

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

Related papers

Discussion (0)

Sign in to comment.