Unverified50% confidenceFactExact time
Traditional knowledge distillation uses an off-policy approach, having the student imitate the teacher's output distribution on a fixed dataset, which leads to exposure bias.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn-policy distillation applies teacher supervision on trajectories generated by the student itself, enabling better handling of the exposure bias problem.78% similarUnverifiedexposure bias的根源在于训练与推理的分布不匹配，训练时使用teacher forcing策略而推理时模型依赖自身生成的token71% similarUnverifiedThe arXiv paper "Imprinting: Knowledge Distillation Can Shift Student Behavior" found that even when the distillation task is unrelated to safety, the teacher model's safety tendencies still transfer to the student model; when the teacher holds an accepting attitude toward harmful requests, the distilled student shows a marked reduction in harm-avoidance capability.71% similarUnverifiedThe privileged information's benefit to critic value prediction accuracy is stronger at the end of the response, while the self-teacher guidance signal exhibits a pattern of being strong at the beginning, decaying in the middle, and rising again at the end.68% similarUnverifiedThe fundamental difference between reinforcement learning and supervised learning is that the data distribution changes with the agent's behavior, forcing it to deal with the consequences of its own actions.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130757API
curl https://kongchang.com/api/v1/knowledge/claims/1130757MCP
get_claim(id=1130757)