acceptodds
Withdrawn from ICLR 2027

MM-OPD: Towards One More Bottleneck Between Perception and Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Recent multimodal large language models (MLLMs) advance visual reasoning by strengthening both perception and reasoning, implicitly assuming a process that transitions seamlessly from perception to reasoning. However, we observe a counterintuitive phenomenon that challenges this assumption: holding the model, question, and decoding fixed, we replace images with their caption or code representations (symbolic views, which seems to be redundant given the clear image structures, but the performance surprisingly improves by 10.2% to 23.6% across model scales and datasets. We term this performance gap as the Symbolic Visual Gap and then take a closer look at it. Through experiments, we find that although the visual evidence can already appear in the reasoning trace for the image-input model, the symbolic-view-input model shows much higher attention to the correct evidence than the image-input model. This suggests that despite good capabilities from current works in perception and reasoning themselves, another bottleneck exists between perception and reasoning in selecting perceived visual information as appropriate evidence for subsequent reasoning. To handle this bottleneck, since the symbolic view steers attention toward correct evidence and is readily obtained at scale, it provides supervision for evidence selection without manually labeled evidence. Building on this, we introduce MM-OPD, a multimodal on-policy self-distillation framework for symbolic-to-visual correction that transfers guidance from symbolic-conditioned behavior to the image-conditioned policy through residual token-level targets, steering the model toward correct visual evidence. Experiments across benchmarks and model scales show that MM-OPD improves a broad range of multimodal abilities, with gains in visual perception, chart and document understanding, mathematical reasoning, and general VQA.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal, OPD, Visual Evidence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RU4zOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a critical failure mode in MLLMs: the Symbolic Visual Gap. The authors observe that replacing visual inputs with their underlying symbolic representations (e.g., generating code or descriptive captions) dramatically improves downstream reasoning accuracy by up to ~23% under identical model weights, prompts, and decoding configurations. Through fine-grained reasoning-trace inspections and answer-to-evidence attention analyses, the authors reveal that while models frequently verbalize the correct visual facts during perception, they fail to select them as premises during deduction, preferentially attending to distracting evidence instead. To bridge this bottleneck between perception and reasoning, the authors propose MM-OPD, an on-policy self-distillation framework. MM-OPD leverages privileged symbolic views in a frozen self-teacher to construct centered residual targets, effectively steering the student’s on-policy image-conditioned rollouts toward correct visual evidence without test-time overhead.

Critical strengths

Method wise and motivation wise this paper has merits in both.

Validated on both 4B and 9B backbones across 18 diverse benchmarks (general VQA, math, charts, OCR), MM-OPD consistently outperforms standard SFT, on-policy RL baselines, and vanilla OPSD.

The framework only requires symbolic representations during offline post-training. At test time, the model executes standard single-image question answering without requiring costs.

Critical weaknesses

Structured visuals (e.g., charts, geometric figures, Venn diagrams) map cleanly to executable code. However, for unconstrained in-the-wild natural scenes, dense captions often fail to capture complex 3D relationships, subtle textures, or high-density background interactions. The paper does not currently thoroughly discuss the upper limit of descriptive captions in representing open-domain images.

Calculating on-policy rollouts and evaluating four separate forward passes for each token across length budgets up to 8,192 tokens represents non-trivial training FLOPs compared to standard SFT or single-teacher distillation. Any comparisons?

Questions for authors

Any results other than qwen-family, like gemma or muse Glimmer?

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

This paper investigates the bottleneck between visual perception and reasoning in MLLMs. The authors observe that replacing images with their textual descriptions or code representations can improve performance by 10.2%–23.6%, a phenomenon they term the Symbolic Visual Gap. Based on this finding, they propose MM-OPD, an on-policy self-distillation framework that uses symbolic inputs to guide evidence selection and transfers this guidance to image-conditioned models through residual token-level training targets. Experiments show improvements across visual perception, chart and document understanding, mathematical reasoning, and general visual question answering. The validation includes comparisons with existing baselines and ablation studies, although further controlled comparisons are needed to isolate the contributions of individual design choices.

Critical strengths
  • The paper is well organized, with a coherent progression from empirical observations and case studies to method design.

  • The evaluation covers a diverse range of tasks, including general VQA, mathematical reasoning, fine-grained perception, and chart and document understanding. MM-OPD achieves the highest average score in Table 1 for both Qwen3.5-4B and Qwen3.5-9B, suggesting that its benefits extend across multiple tasks and model scales.

Critical weaknesses
  • The performance gap between visual and symbolic inputs has been studied in prior work, including Lost in Translation [1] and Beyond Captioning [2]. Other approaches, such as Vision-R1 [3], Vision-SR1 [4], Visionary-R1 [5], and Perception-R1 [6], have explored leveraging textual reasoning or supervision to improve visual reasoning. The paper does not sufficiently clarify what new insights or technical contributions MM-OPD offers beyond these existing approaches.

[1] Lost in Translation: When GPT-4V(ision) Can't See Eye to Eye with Text. A Vision-Language-Consistency Analysis of VLLMs and Beyond.

[2] Beyond Captioning: Task-Specific Prompting for Improved VLM Performance in Mathematical Reasoning.

[3] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models.

[4] Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition and Multi-Reward Policy Optimization.

[5] Visionary-R1: Mitigating Shortcuts in Visual Reasoning with Reinforcement Learning.

[6] Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward.

  • The comparison with Vanilla OPSD does not isolate the effects of symbolic inputs and the proposed residual targets. Moreover, the individual contributions of residual subtraction, clean-anchor reweighting, and probability-mass preservation remain unclear, making it difficult to assess whether these design choices are necessary for the reported gains.

  • The evidence annotation procedure lacks sufficient detail to assess the reliability of the attention analysis. Important implementation details of the SFT and RL baselines are also missing, making the experimental comparisons difficult to interpret.

Questions for authors
  • Can the authors compare OPD, adapted Purified OPSD, and the full MM-OPD under the same symbolic inputs and training conditions? Can they also provide ablations for residual subtraction, clean-anchor reweighting, and probability-mass preservation?

  • How are correct and distracting evidence spans annotated, and what quality-control procedures are used to ensure the reliability of the attention analysis?

  • What are the sources and quality-control procedures for the SFT reference responses? Given the performance degradation, how do the authors explain the results?

  • How are rewards implemented for the GRPO, GSPO, and DAPO baselines? Why are rule-based or hybrid verifiers not used for objectively verifiable tasks, and how sensitive are the results to the choice of judge model?

Reviewer_rAmjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the performance gap between image-conditioned and symbolic-conditioned multimodal reasoning, where symbolic views take the form of captions or image-generating code. The authors attribute this gap to a visual evidence-selection bottleneck: models may correctly verbalize relevant visual information but fail to use it in subsequent reasoning. Motivated by this observation, they propose MM-OPD, an on-policy self-distillation method that uses symbolic information as privileged teacher input and constructs residual token-level targets for an image-conditioned student.

Critical strengths
  1. The paper studies a unique failure mode: models can verbalize relevant visual information yet fail to use it correctly.

  2. The paper proposes to use symbolic views and codes to provide as privleged information, which is interesting.

Critical weaknesses
  1. The central claim is not well-established. How are correct visual evidence distinguished from incorrect distractions? While the paper argues that visual evidence selection is a seperate bottleneck from perception and reasoning, the clear definition of the problem is in question. What are the protocols/rubrics of finding such evidence and distractions and how correct the method is?

  2. The method assumes access to informative symbolic views during training, which narrows the domains MM-OPD can be applied to. It is also a fairly strong data assumption: images need to come with source code or captions that preserve the information needed for the task. It is unclear how this setup carries over to more general post-training setups, for example, is MM-OPD needed or beneficial for long-horizon agentic tasks such as visual coding agents? The current results establish gains on static multimodal benchmarks, but leave the broader applicability of the approach open.

  3. The compute–performance tradeoff is unclear. The average gains over vanilla OPSD are 2.0 points at 4B and 1.6 points at 9B, with mixed results across benchmarks, while MM-OPD requires three teachers. Does MM-OPD improve the compute–performance Pareto frontier over standard OPD/OPSD?

  4. The necesity of the training objective is unclear. The observation that symbolic inputs improve performance motivates using them for distillation, but does not explain why the particular training objective is needed. The three teacher passes, the T/RT/R correction, and tanh clipping have intuitive motivations, but their necessity is not established. Can the paper elaborate more with evidence why this particular training objective is necessary for solving the proposed bottleneck?

Questions for authors
  1. Do the gains hold under a matched FLOP budget? Three teachers do add some compute overhead compared to methods that have only one teacher.

  2. How would the method apply when the privileged information provided by symbolic views/codes is not available? Particularly how does MM-OPD fit into today's standardized post-training setups and what are the advantages of introducing it to the current practice?

  3. Can the bottleneck already be effectively addressed through data curation? Or is this bottleneck a foundamental constraint that would appear no matter what under the current paradigm? In other words, if higher-quality data were to be introduced, would a shift in objective from standard OPD/OPSD still be necessary? My concern is whether MM-OPD offers a lasting advantage as data and compute scale, or mainly compensates for deficiencies in the current model checkpoints.

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

Related papers

Discussion (0)

Sign in to comment.