VISTA: Verifier-Informed Student-to-Teacher Adaptation for On-Policy Self-Distillation
VISTA lets the privileged teacher move toward verified student reasoning instead of treating it as a fixed target.
The method keeps the standard OPSD student update, then uses outcome-verified rollouts to adapt the teacher toward the student distribution. It limits that teacher adaptation to the top-k token positions with the largest teacher-student KL divergence. The authors say it adds no extra sampling or separate reward objective. On AIME24, AIME25, and HMMT25 with Qwen3 models from 1.7B to 8B, VISTA beats OPSD by 0.6, 0.7, and 2.1 Avg@12 points. ArXiv · AI/CL/LG's note
The method keeps the standard OPSD student update, then uses outcome-verified rollouts to adapt the teacher toward the student distribution. It limits that teacher adaptation to the top-k token positions with the largest teacher-student KL divergence. The authors say it adds no extra sampling or separate reward objective. On AIME24, AIME25, and HMMT25 with Qwen3 models from 1.7B to 8B, VISTA beats OPSD by 0.6, 0.7, and 2.1 Avg@12 points. ArXiv · AI/CL/LG's note
score 5