The authors examined states late in student-generated episodes where the untrained student disagreed most strongly with its teacher.
Teacher rollouts repair myopic token scoring in agent distillation
R2S uses successful teacher continuations to improve token-level supervision without task rewards during repair or student training.
Big Tech
Shu Zhao · Tan Yu
NVIDIA
Research Digest··3 min read
Zhao and Yu identify a failure mode in agentic on-policy distillation that cannot be explained solely by students drifting into states their teachers cannot handle.
§