Unverified50% confidenceFactExact time
On-policy self-distillation computes the distillation loss on trajectories sampled by the student itself, conditioning on the states produced by the student's current policy to let the teacher provide a better target distribution, thereby reducing error accumulation caused by distribution shift.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified自回归生成系统训练采用Teacher Forcing策略，推理阶段切换为自回归模式后产生分布偏移导致误差累积放大74% similarUnverifiedDistillation can be viewed as an implicit, data-constrained form of model merging—distilling a student using the averaged outputs of multiple teacher models is effectively performing a soft merge at the level of output distributions.73% similarUnverifiedIn on-policy distillation, the student model first generates trajectories itself, and then the teacher model provides supervision signals for these paths, thereby avoiding the distribution shift between offline data and student behavior.73% similarUnverifiedPolicy distillation typically refers to using a student policy to imitate the behavior of one or more teacher policies.72% similarUnverifiedTeacher forcing introduces the exposure bias problem: during training the real target token is fed as input, while during inference the model's own predicted token is used, and prediction errors accumulate and amplify in subsequent steps.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1033281API
curl https://kongchang.com/api/v1/knowledge/claims/1033281MCP
get_claim(id=1033281)