Unverified50% confidenceFactExact time
Simple token removal can only weaken the transfer effect of the teacher's safety tendencies to the student and cannot completely eliminate it; post-hoc DPO alignment can also only partially correct this imprinted behavioral tendency.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWeak-to-strong generalization does not hold universally, but depends on conditions such as target structure and training methodology; when conditions are not met, the student may fail to surpass the teacher67% similarUnverifiedEven when teacher demonstrations are completely correct at the semantic level, SFT still loses previously learned knowledge.63% similarUnverifiedThe strong-version hypothesis that 'programming makes people broadly smarter' has been largely abandoned by the academic community, though some subsequent studies have observed limited transfer effects under specific conditions.63% similarUnverifiedSynthetic data distillation prevents the student model from learning the teacher's confidence distribution, resulting in relatively limited generalization ability on ambiguous samples.63% similarUnverifiedThe paper finds that when the student's trajectory moves in the wrong direction, the teacher does not correct the error but instead continues writing along the student's error, that is, the teacher has the ability to recognize errors but lacks the ability to escape from them.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044460API
curl https://kongchang.com/api/v1/knowledge/claims/1044460MCP
get_claim(id=1044460)