{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2608.05131v1",
  "slug": "2608-05131v1-0m0x349",
  "url": "https://feed7.dev/p/2608-05131v1-0m0x349",
  "title": "OPD-V: Visual On-Policy Self-Distillation with Modality Balance",
  "why_included": "OPD-V selects self-distillation tokens by measuring whether visual input influences generation, addressing text dominance that can blunt multimodal reasoning training.",
  "summary": "OPD-V uses a zoomed image as a positive teacher and a masked image as a negative teacher. Their logit differences define a modality-balance trust region for selecting on-policy tokens. Tests cover **6 benchmarks**, **4 MLLM backbones**, and **5 post-training methods**.",
  "practical_implication": "Builders training vision agents should inspect whether generated reasoning actually depends on the image, not merely whether privileged inputs are available. Token selection based on modality balance may make self-distillation more targeted and reduce wasted training work.",
  "agent_context": "OPD-V uses a zoomed image as a positive teacher and a masked image as a negative teacher. Their logit differences define a modality-balance trust region for selecting on-policy tokens. Tests cover **6 benchmarks**, **4 MLLM backbones**, and **5 post-training methods**.\n\nBuilders training vision agents should inspect whether generated reasoning actually depends on the image, not merely whether privileged inputs are available. Token selection based on modality balance may make self-distillation more targeted and reduce wasted training work.\n\nThe abstract reports consistent reasoning gains and lower training cost but gives no effect sizes, task-level results, or implementation details. It remains unclear how well the method transfers beyond the tested backbones and benchmarks.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2608.05131v1",
    "published_at": "2026-08-05T17:53:06.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "image"
  ],
  "topics": [
    "reasoning"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The abstract reports consistent reasoning gains and lower training cost but gives no effect sizes, task-level results, or implementation details. It remains unclear how well the method transfers beyond the tested backbones and benchmarks."
  ],
  "connected_context": {
    "meaning": "OPD-V narrows multimodal self-distillation from general teacher–student disagreement to whether each generated token is genuinely supported by vision. It complements cross-view supervision and other on-policy distillation methods with a modality-balance criterion, but the missing effect sizes leave its relative value and deployment economics unresolved.",
    "corpus_size": 368,
    "generated_at": "2026-08-06T10:06:22.363Z",
    "connections": [
      {
        "title": "MIRROR: Learning from the Other View for Multi-Modal Reasoning",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.21552v1",
        "feed7_url": "https://feed7.dev/p/2607-21552v1-1v5rc1p",
        "reason": "Both target modality inconsistency, but MIRROR transfers knowledge between text, diagram, and combined views, whereas OPD-V uses altered versions of the same image to select visually grounded on-policy tokens."
      },
      {
        "title": "DemoPSD: Disagreement-Modulated Policy Self-Distillation",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.02502v1",
        "feed7_url": "https://feed7.dev/p/2607-02502v1-0wngknx",
        "reason": "DemoPSD gates tokens by teacher–student disagreement; OPD-V supplies a specifically multimodal gating signal based on the difference between zoomed-image and masked-image teachers."
      },
      {
        "title": "$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.28582v1",
        "feed7_url": "https://feed7.dev/p/2607-28582v1-0egi1xh",
        "reason": "β-OPSD broadens control over self-distillation through tunable regularization and credit assignment, while OPD-V adds a complementary criterion for deciding which multimodal tokens should receive that training signal."
      },
      {
        "title": "X$^3$-OPD: Distilling Reasoning into Large Audio-Language Models via On-Policy Alignment",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.21550v1",
        "feed7_url": "https://feed7.dev/p/2607-21550v1-1j7d28n",
        "reason": "Both adapt on-policy distillation to a non-text modality, but X³-OPD grounds reasoning in acoustic interpretations while OPD-V tests visual dependence through privileged and degraded image views."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-08-05T17:53:06.000Z",
  "modified_at": "2026-08-05T17:53:06.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2608-05131v1-0m0x349",
    "json": "https://feed7.dev/p/2608-05131v1-0m0x349.json",
    "markdown": "https://feed7.dev/p/2608-05131v1-0m0x349.md"
  }
}