Source-linked AI summary
MM-OPD: Towards One More Bottleneck Between Perception and Reasoning
Jintao Tong, Yujing Lou, Zhanming Shen, Jiaqi Gu, Lubin Fan, Ruixuan Li, Yue Wu, Jieping Ye, Yixiong Zou
TL;DR
The paper questions the assumed seamless transition from perception to reasoning, finding that symbolic views outperform image inputs and that evidence selection remains a bottleneck. It introduces MM-OPD to transfer symbolic-conditioned guidance to image-conditioned models, reporting improvements across multimodal abilities.
Problem
Image-input models may perceive correct visual information without selecting it as evidence for subsequent reasoning.
Source: §1 · pp. 1–2Method
MM-OPD uses symbolic-conditioned behavior to supervise the image-conditioned policy with residual token-level targets.
Source: §1 · p. 2Results
MM-OPD improves a broad range of multimodal abilities across benchmarks and model scales.
Source: §6 · p. 11Takeaways & Limitations
Symbolic views provide scalable guidance for evidence selection without manually labeled correct evidence.
Source: §2.3 · p. 4
Abstract
from arXiv · showhide
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.
1 Introduction
Replacing an image with its symbolic view improves performance, exposing evidence selection as a bottleneck between perception and reasoning. MM-OPD transfers symbolic-conditioned guidance to the image-conditioned model, with reported gains across multimodal tasks.
1 Introduction
- 10.2 to 23.6 percentage points: symbolic views improve accuracy over visual inputs across model scales and visual domains.The comparison holds the model, question, and decoding fixed.Source: §1 · pp. 1–2, Figure 1 · p. 2
- MM-OPD uses a symbolic-conditioned model as teacher to guide the image-conditioned policy through residual token-level targets.The framework transfers guidance along student-generated trajectories to steer evidence selection.Source: §1 · p. 2
- Correct visual evidence can appear in image-input reasoning traces, yet image and symbolic conditioning show opposite attention patterns toward correct evidence.Image-input models tend to adopt distracting evidence, while symbolic-view inputs favor correct evidence.Source: §1 · pp. 1–2
- Experiments report consistent improvements across model scales and multimodal reasoning tasks.Reported areas include visual perception, chart and document understanding, mathematical reasoning, and general multimodal reasoning.Source: §1 · p. 2
2 Delve into the Symbolic Visual Gap
The section examines whether perceived visual information is selected as appropriate evidence for reasoning. It finds that symbolic conditioning is associated with stronger attention to correct evidence, motivating symbolic views as scalable guidance for image-conditioned reasoning without manual evidence labels.
2.1 From Visual Perception to Evidence Selection
- Correct visual information can appear in the reasoning trace without being selected: the model may instead use distracting evidence as a premise.The subsequent comparison can be valid given that misleading premise, despite relevant values and their ordering being stated in the trace.Source: §2.1 · p. 3
2.2 Answer-to-Evidence Attention under Image and Symbolic Conditions
- Across observed layers, image conditioning favors distracting evidence while symbolic conditioning favors correct evidence; the reversal also holds in overall ratios and signed scores.The signed score is positive for symbolic conditioning and negative for image conditioning at every observed layer.Source: §2.2 · pp. 3–4, Figure 3 · p. 4
2 Delve into the Symbolic Visual Gap
- The analysis identifies evidence selection as a bottleneck between perception and reasoning: perceiving required information does not ensure it becomes appropriate evidence for answering.The section assesses whether relevant visual information is represented as language-accessible evidence and whether subsequent reasoning draws on it.Source: §2 · p. 2, §2.2 · p. 4, §2.3 · p. 4
2.3 Summary and Discussion
- Symbolic views can guide image-conditioned evidence selection without manual evidence labels, and generating code or descriptive captions can be obtained automatically at scale.The proposed guidance draws on the behavior elicited by symbolic conditioning, which exhibits stronger performance and an attention preference for correct evidence when the model and question are fixed.Source: §2.3 · p. 4
3 Method
MM-OPD treats symbolic views as privileged guidance for image-conditioned reasoning and uses residual token-level targets to refine predictions. Training uses student-generated trajectories and curated symbolic-image pairs, while inference removes symbolic inputs and teacher evaluations.
3 Method
- MM-OPD uses symbolic views as privileged supervision in on-policy self-distillation, with residual token-level targets refining image-conditioned predictions toward visual evidence.The frozen self-teacher supplies guidance while residual targets reduce the influence of preferences induced by symbolic views alone.Source: §3 · p. 4, Figure 4 · p. 5
3.1 Symbolic-to-Visual Correction
- MM-OPD contrasts joint-context T_t with privilege-only R_t, then uses the centered residual to reweight clean anchor B_t while preserving its selected-support and tail masses.The centered residual measures relative token preference beyond what the symbolic information induces by itself.Source: §3.1 · pp. 5–6
3.2 Training Objective
- Training minimizes symmetric Jensen–Shannon divergence between compressed student predictions and corrected targets over on-policy rollouts; inference removes symbolic inputs and teacher evaluations.The student generates the trajectories used for token-level supervision, and gradients flow only through the image-conditioned student.Source: §3.2 · pp. 6–7
3.3 Training Dataset Curation
- The 32K-example dataset validates image-symbolic pairs, normalizes existing captions and code, and filters for non-decreasing Pass@8 with symbolic input.It spans general VQA, mathematical diagrams, and charts.Source: §3.3 · p. 7
4 Experiments
Experiments compare MM-OPD with controlled post-training baselines, examine design choices, and test whether training shifts image-conditioned evidence selection toward correct evidence while narrowing the Symbolic Visual Gap.
4.2 Main Results
- MM-OPD achieves the highest average score at both scales, beating the Qwen3.5-4B and Qwen3.5-9B baselines by 1.9 and 2.1 points, respectively, across 38 of 40 model–benchmark settings.The benchmark set spans general VQA, multimodal reasoning, fine-grained perception, and document understanding.Source: Table 1 · p. 8, §4.2 · p. 8
- MM-OPD exceeds the best RLVR baseline average by 1.7 points at 4B and 2.8 points at 9B, while SFT reduces average scores by 3.4 and 3.3 points, respectively.RLVR methods show benchmark-specific gains and regressions; MM-OPD also improves over vanilla OPSD by 2.0 points at 4B and 1.6 points at 9B.Source: §4.2 · pp. 8–9
4.3 Ablation Study
- Across the tested distillation objectives, all improve over baseline on every benchmark; JSD has the highest overall average and is adopted as default.JSD and reverse KL each lead on three benchmarks.Source: §4.3 · p. 9, Table 2 · p. 9
- Both EMA and frozen teachers improve average performance over baseline, but the frozen teacher scores 1.6 points higher on average and leads on four of six benchmarks.EMA performs better on HR-Bench 4K and MMStar; the frozen teacher retains its initial weights.Source: §4.3 · p. 9, Table 3 · p. 9
- All tested response-length limits outperform baseline on average, but performance peaks at 8192 tokens and declines slightly at 16384.Different benchmarks favor different limits, so increasing the maximum length does not consistently improve every task.Source: §4.3 · p. 9, Table 4 · p. 9
- All tested rollout counts outperform baseline on every benchmark, and two rollouts per prompt yield the highest average, improving it by 0.9 points over one rollout.Four or eight rollouts produce slightly lower averages.Source: §4.3 · p. 10, Table 5 · p. 10
4.4 Effectiveness of Symbolic-to-Visual Correction
- MM-OPD increases relative attention to correct evidence across observed layers, with correct evidence receiving more attention than distracting evidence from layer 12 onward.The base model favors distracting evidence at every observed full-attention layer.Source: §4.4 · p. 10
- MM-OPD reduces the Symbolic Visual Gap on Chart, Math, and General VQA; Chart has the largest absolute reduction, while Math and General VQA close larger proportions of their initial gaps.The symbolic-conditioned accuracy reference is held fixed for each task family.Source: §4.4 · p. 10, Figure 5 · p. 10
5 Related Work
Prior multimodal work strengthens visual reasoning through perception, post-training, and evidence acquisition, while MM-OPD uses token-level self-distillation from symbolic views to guide image-conditioned answers.
Multimodal Reinforcement Learning
- Multimodal RL methods strengthen reasoning through cold-start training and RL or broader task scaling, while agentic methods acquire and manipulate visual evidence through operations such as cropping and zooming.Final-answer rewards do not explicitly specify which intermediate visual evidence should support an answer.Source: §5 · p. 10
On-Policy Self-Distillation; Distillation with Privileged Information
- Unlike self-distillation methods targeting specific capabilities and privileged-information methods using other conditioning, MM-OPD uses symbolic views and residual corrections to guide image-conditioned evidence use across abilities.The related-work discussion contrasts this with Vision-OPD for fine-grained perception, RLSD for mathematical reasoning, and reference-conditioned OPSD methods.Source: §5 · p. 11
6 Conclusion
The paper identifies visual evidence selection as a bottleneck between perception and reasoning and introduces MM-OPD to transfer symbolic-view guidance to image-conditioned reasoning. Experiments report improvements across a broad range of multimodal abilities.
6 Conclusion
- MM-OPD transfers guidance from symbolic views to the image-conditioned policy through residual token-level targets that steer it toward correct visual evidence.The paper frames evidence selection as an additional bottleneck between perception and reasoning.Source: §6 · p. 11
A Additional Case Studies
Additional cases illustrate that relevant visual information can appear in a reasoning trace yet be selected or used incorrectly as evidence.
A Additional Case Studies
- In the radial bar chart case, the response includes relevant values but associates a distracting value with the queried category, producing a wrong answer from that premise.Attention is stronger on the distracting value used in the derivation.Source: A Additional Case Studies · p. 15, Figure 6 · p. 15
- In the Venn diagram case, the response lists the shared factor but omits it from the comparison, concluding there is no common cause.The distracting pair-specific factors receive stronger attention than the omitted shared factor.Source: A Additional Case Studies · p. 15, Figure 7 · p. 15