VAG
A gating mechanism for multimodal reinforcement learning that constructs two teacher views, evidence-present and evidence-degraded, for comparison, and admits only privileged corrections supported by visual contrast consistency, suppressing update contamination from non-visual confounding factors.
Timeline (last 90 days)
The research proposes a framework called Visually Attributable On-Policy Self-Distillation (VAG), which filters contaminated privileged corrections through visual attribution gating.
The study validated on six visual reasoning benchmarks, and VAG achieves consistent advantages over both GRPO and OPSD methods and sets a new best average performance level.
VAG uses verifier advantage (the task-level signal of RLVR) as the optimization anchor, ensuring that the overall direction of policy updates does not deviate from verifiable correctness.
For each privileged correction, VAG constructs two teacher views—evidence present and evidence degraded—and observes whether the correction direction receives directional support in the visual contrast; only corrections that are consistently supported are accepted.
VAG borrows the idea of perturbation contrast: it constructs two views—evidence present and evidence degraded—and if a correction direction remains unchanged after evidence degradation, this indicates it is driven by language priors or stylistic factors rather than visual grounding.
All Facts (5)
The research proposes a framework called Visually Attributable On-Policy Self-Distillation (VAG), which filters contaminated privileged corrections through visual attribution gating.
50%UnverifiedThe study validated on six visual reasoning benchmarks, and VAG achieves consistent advantages over both GRPO and OPSD methods and sets a new best average performance level.
50%UnverifiedVAG uses verifier advantage (the task-level signal of RLVR) as the optimization anchor, ensuring that the overall direction of policy updates does not deviate from verifiable correctness.
50%UnverifiedFor each privileged correction, VAG constructs two teacher views—evidence present and evidence degraded—and observes whether the correction direction receives directional support in the visual contrast; only corrections that are consistently supported are accepted.
50%UnverifiedVAG borrows the idea of perturbation contrast: it constructs two views—evidence present and evidence degraded—and if a correction direction remains unchanged after evidence degradation, this indicates it is driven by language priors or stylistic factors rather than visual grounding.
50%