Sign InOpen Brain
arXivPaperNeeds Review

OPD-V: Visual On-Policy Self-Distillation with Modality Balance

OPD-V selects self-distillation tokens by measuring whether visual input influences generation, addressing text dominance that can blunt multimodal reasoning training.

arXiv · Aug 5, 2026
Open Source Open MarkdownOpen JSON
Source 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-Ready 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**.

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.

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 · Feed7 Judgment

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.

MIRROR: Learning from the Other View for Multi-Modal ReasoningBoth 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.DemoPSD: Disagreement-Modulated Policy Self-DistillationDemoPSD 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.$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillationβ-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.X$^3$-OPD: Distilling Reasoning into Large Audio-Language Models via On-Policy AlignmentBoth 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.
Context Map
modelimage#reasoning
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.