mlsanityData to 7 Oct 2026

Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation

arXiv 2605.18740 ↗First posted 18 May 2026cs.CV cs.AI cs.CLcited by 71

On the map: Multimodal LLMs / vision-language models → MLLM training and visual instruction tuning

Abstract (arXiv)

Multimodal Large Language Models (MLLMs) still struggle with fine-grained visual understanding, where answers often depend on small but decisive evidence in the full image. We observe a regional-to-global perception gap: the same MLLM answers fine-grained questions more accurately when conditioned on evidence-centered crops than on the corresponding full images, suggesting that many failures stem from difficulty to focus on relevant evidence rather than insufficient local recognition ability. Motivated by this observation, we propose Vision-OPD (Vision On-Policy Distillation), a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy. Vision-OPD instantiates two conditional policies from the same MLLM: a crop-conditioned teacher and a full-image-conditioned student. The student generates on-policy rollouts, and Vision-OPD minimizes token-level divergence between the teacher and student next-token distributions along these rollouts. This enables the model to internalize the benefit of visual zooming without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use. Experiments on multiple fine-grained visual understanding benchmarks show that Vision-OPD models achieve competitive or superior performance against much larger open-source, closed-source, and "Thinking-with-Images" agentic models. The code is available at https://github.com/VisionOPD/Vision-OPD

Key facts. Cited by 71 papers in the map (data through 7 Oct 2026). Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 65 of its 74 references that are in the map. First posted on arXiv 18 May 2026.

Builds on

65 of 74 references.

Closest ancestors · 14

Same direction, newest first

  1. 2026Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception12 Feb
  2. 2025Grasp Any Region: Towards Precise, Contextual Pixel Understanding for Multimodal LLMs21 Oct
  3. 2025HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling28 Sep
  4. 2025MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe16 Sep
  5. 2025Zoom-Refine: Boosting High-Resolution Multimodal Understanding via Localized Zoom and Self-Refinement2 Jun
  6. 2025Evaluating and Steering Modality Preferences in Multimodal Large Language Model27 May
  7. 2025MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs24 Feb
  8. 2025SAISA: Towards Multimodal Large Language Models with Both Training and Inference Efficiency4 Feb
  9. 2024ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration25 Nov
  10. 2024Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models28 Aug
  11. 2024MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans?23 Aug
  12. 2024Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs24 Jun
2 more
  1. 2024TextCoT: Zoom In for Enhanced Multimodal Text-Rich Image Understanding15 Apr
  2. 2024Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs11 Jan

Other references · 51

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2025Qwen3 Technical Reportcited by 8,195
  3. 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,655
  4. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
  5. 2023Evaluating Object Hallucination in Large Vision-Language Modelscited by 1,696
  6. 2025OpenAI GPT-5 System Cardcited by 910
  7. 2022STaR: Bootstrapping Reasoning With Reasoningcited by 895
  8. 2024Are We on the Right Way for Evaluating Large Vision-Language Models?cited by 893
  9. 2026DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligencecited by 877
  10. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 840
  11. 2025SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-trainingcited by 655
  12. 2023Reinforced Self-Training (ReST) for Language Modelingcited by 486
38 more
  1. 2023V*: Guided Visual Search as a Core Mechanism in Multimodal LLMscited by 459
  2. 2026Kimi K2.5: Visual Agentic Intelligencecited by 449
  3. 2025GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learningcited by 439
  4. 2015Policy Distillationcited by 414
  5. 2025DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learningcited by 398
  6. 2026Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Modelscited by 386
  7. 2024ShortGPT: Layers in Large Language Models are More Redundant Than You Expectcited by 369
  8. 2026Reinforcement Learning via Self-Distillationcited by 285
  9. 2025Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learningcited by 267
  10. 2024Visual Sketchpad: Sketching as a Visual Chain of Thought for Multimodal Language Modelscited by 260
  11. 2025Thyme: Think Beyond Imagescited by 160
  12. 2025Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawingcited by 130
  13. 2025DeepEyesV2: Toward Agentic Multimodal Modelcited by 119
  14. 2025Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Searchcited by 116
  15. 2024Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Samplingcited by 85
  16. 2025ReFocus: Visual Editing as a Chain of Thought for Structured Image Understandingcited by 81
  17. 2025VGR: Visual Grounded Reasoningcited by 70
  18. 2025Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodologycited by 67
  19. 2024Chain-of-Spot: Interactive Reasoning Improves Large Vision-Language Modelscited by 65
  20. 2024Image-of-Thought Prompting for Visual Reasoning Refinement in Multimodal Large Language Modelscited by 64
  21. 2025LiveMCPBench: Can Agents Navigate an Ocean of MCP Tools?cited by 47
  22. 2026Uni-OPD: Unifying On-Policy Distillation with a Dual-Perspective Recipecited by 43
  23. 2025Advancing Multimodal Reasoning via Reinforcement Learning with Cold Startcited by 41
  24. 2024In-Context Editing: Learning Knowledge from Self-Induced Distributionscited by 33
  25. 2025SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learningcited by 31
  26. 2025MemSearcher: Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learningcited by 26
  27. 2025Skywork-R1V4: Toward Agentic Multimodal Intelligence through Interleaved Thinking with Images and DeepResearchcited by 25
  28. 2025CodeV: Code with Images for Faithful Visual Reasoning via Tool-Aware Policy Optimizationcited by 24
  29. 2026DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generationcited by 14
  30. 2025ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layerscited by 13
  31. 2025Perception in Reflectioncited by 13
  32. 2025Patch Matters: Training-free Fine-grained Image Caption Enhancement via Local Perceptioncited by 12
  33. 2025VLM-FO1: Bridging the Gap Between High-Level Reasoning and Fine-Grained Perception in VLMscited by 12
  34. 2025Visual Reasoning Tracer: Object-Level Grounded Reasoning Benchmarkcited by 10
  35. 2025VG-Refiner: Towards Tool-Refined Referring Grounded Reasoning via Agentic Reinforcement Learningcited by 8
  36. 2026Beyond Unimodal Shortcuts: MLLMs as Cross-Modal Reasoners for Grounded Named Entity Recognitioncited by 6
  37. 2026Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Tracescited by 6
  38. 2026HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoningcited by 5

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 71.

71 papers.

Sort:

By area:

  1. 2026Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoningcited by 15
  2. 2026ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillationcited by 12
  3. 2026Thinking Without Images: Internalizing Visual Manipulation with On-Policy Self-Distillationcited by 9
  4. 2026DOPD: Dual On-policy Distillationcited by 9
  5. 2026Visual Contrastive Self-Distillationcited by 9
  6. 2026H-OPD: Confidence Aware Heterogeneous Multi-Teacher Multimodal On-policy Distillationcited by 7
  7. 2026ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillationcited by 7
  8. 2026Weak-to-Strong On-Policy Distillationcited by 7
  9. 2026OPD-V: Visual On-Policy Self-Distillation with Modality Balancecited by 7
  10. 2026OmniOPSD: Rationale-Privileged On-Policy Self-Distillation for Affective Computingcited by 6
  11. 2026RP-OPSD: Resolution-Privileged On-Policy Self-Distillation for Multimodal Large Language Modelscited by 6
  12. 2026Latent On-Policy Self-Distillationcited by 6
38 more
  1. 2026Self-Supervised Visual On-Policy Distillationcited by 6
  2. 2026A Formula-Driven Survey and Research Agenda for On-Policy Distillationcited by 5
  3. 2026VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillationcited by 5
  4. 2026When Teachers Mislead: Spurious-Signal-Aware On-Policy Distillationcited by 5
  5. 2026ParaBridge: Bridging Paralinguistic Perception and Dialogue Behavior in Speech Language Modelscited by 4
  6. 2026Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillationcited by 4
  7. 2026Distill What the Student Can See: Fisher-Projected On-Policy Distillation for Vision-Language Modelscited by 4
  8. 2026Where to Look Matters: On-Policy Self-Distillation for Long-Video Understandingcited by 4
  9. 2026MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillationcited by 3
  10. 2026Contrastive On-Policy Distillationcited by 3
  11. 2026Evidence-RL: Towards Evidence-intensive Visual Reasoningcited by 3
  12. 2026SkillLens: Visual Skill Cards for Retrieval-Augmented GUI Action Prediction and On-Policy Distillationcited by 3
  13. 2026SenseNova-U1.5: Towards Native Unified Visual Intelligencecited by 3
  14. 2026Med-OPD: Improving Medical Vision-Language Models via Evidence-Aware On-Policy Distillationcited by 2
  15. 2026Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detectioncited by 2
  16. 2026DreOPD: Degraded-Reference Extrapolative On-Policy Distillation for Flow-matching Modelscited by 2
  17. 2026Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spotscited by 2
  18. 2026WAM-OPD: On-Policy Distillation for World Action Modelscited by 2
  19. 2026Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillationcited by 2
  20. 2026What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillationcited by 2
  21. 2026Learning Visual Spatial Planning from Symbolic State via Modality-Gap-Aware Self-Distillationcited by 1
  22. 2026OPOD: On-Policy Omni Distillationcited by 1
  23. 2026Test-Time Self-Evolving GUI Visual Grounding via Reflection-Guided On-Policy Self-Distillationcited by 1
  24. 2026Deep Thought Alignment: Trajectory-Level Latent Distillation for Video Reasoningcited by 1
  25. 2026CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual Predictioncited by 1
  26. 2026OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolutioncited by 1
  27. 2026IoU-PD: IoU-Aware Privileged Distillation for Visual Grounding with Multimodal Large Language Modelscited by 0
  28. 2026Correcting What You Cannot See: Credit Assignment for Perception Distillation in Multimodal Reasonerscited by 0
  29. 2026Latent Reward Registers for Diffusion Preference Alignmentcited by 0
  30. 2026Q-CueGraph: Query-Conditioned Visual Evidence Graphs for Multimodal Reasoningcited by 0
  31. 2026ADOPD: Reference-Privileged On-Policy Distillation for MLLM-Based Industrial Anomaly Detectioncited by 0
  32. 2026ArmorOCR: Grounded Adversarial Visual Perception via Observation-Transferred Self-Distillationcited by 0
  33. 2026Buried in Textual Debt: Context Pruning with Visual Evidence Preservation for MLLM Agentscited by 0
  34. 2026Self-OPD: On-Policy Distillation for Flow Matching Models without Teachercited by 0
  35. 2026Temporal Self-Distillation: Learning Visual State Tracking in Videos Without Supervisioncited by 0
  36. 2026OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillationcited by 0
  37. 2026ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Trainingcited by 0
  38. 2026Region-Level Policy Optimization for Fine-grained MLLM Perceptioncited by 0

50 most cited of 71 All 71 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2026Look Where It Counts: A Free, Label-Free Visual Evidence Signal for Fine-Grained Vision-Language Reasoning
  2. 2026LEGO-OPD: Factorized Teacher Composition for Multimodal On-Policy Distillation
  3. 2026MCD: Causal Distillation of Multimodal In-Context Learning in Large Vision-Language Models
  4. 2023LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge
  5. 2026LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models
  6. 2025Visual Funnel: Resolving Contextual Blindness in Multimodal Large Language Models
  7. 2025See, Think, Learn: A Self-Taught Multimodal Reasoner
  8. 2025Language Models Can See Better: Visual Contrastive Decoding For LLM Multimodal Reasoning