Unverified50% confidenceFactExact time
Even when teacher demonstrations are completely correct at the semantic level, SFT still loses previously learned knowledge.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
Unverified多对一映射导致教师token间的不确定性在投影到学生第一个token时被直接加和抹平，这是投影方式的固有属性而非实现错误69% similarUnverifiedStudents who drop out early are often trapped in "recursive conceptual questioning" and are never able to enter the actual problem-solving execution phase.67% similarUnverifiedThe teacher model outputs in distillation training rarely include self-correction examples, causing student models to form an implicit belief that "strong models don't make mistakes," leading them to repeat rather than correct when verification fails.66% similarUnverifiedLecture notes typically lack the systematic exercise solutions and index systems found in textbooks, and may also make trade-offs in depth due to course hour constraints.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975962API
curl https://kongchang.com/api/v1/knowledge/claims/975962MCP
get_claim(id=975962)