Get Started
Home
Topics
Search
Library
Research questionHow can on-policy distillation prevent early student errors from corrupting rollouts in autoregressive vision-language models?In on-policy distillation, the student generates its own token sequence, so an early mistake can alter later states and make subsequent teacher supervision less reliable. This error propagation is especially problematic for autoregressive vision-language models performing structured visual prediction.
AI
Computer Vision
Machine Learning
Multimodal Models
Research Paper
Latest papersRecent research connected to this question, newest first.CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual PredictionThe evidence concerns multi-teacher training for GUI grounding and optical character recognition across six target benchmarks, using a Qwen3.5-0.8B student baseline. Controlled studies examine intervention placement, progressive transfer of rollout control, and supervision aligned with interventions; no evidence is provided for other tasks or deployment settings.research paper · Sep 7, 2026
Related questions
How can on-policy distillation remain stable when policy updates change future training states?How can autoregressive logit distillation preserve local next-token preferences when teacher and student rankings disagree?How can model distillation block hidden teacher-trait transfer through clean data without degrading the target task?How can RLVR reduce the cost of on-policy rollouts and reliable targets without hurting reasoning quality?