Verified75% confidenceOpinionExact time
The 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.
3
Sources
75%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Adding a 'Repair Layer' to Open-Source Models: How DeepSeek Can Outperform Claude Opus
bilibiliGelai_AI7/5/2026
Related Claims
Unverified许多开源模型继承教师模型的自信偏差，导致工具调用错误时不愿自我纠正82% similarUnverifiedSynthetic data distillation prevents the student model from learning the teacher's confidence distribution, resulting in relatively limited generalization ability on ambiguous samples.79% similarUnverifiedThe reinforcement learning modifications tested in the study did not yield consistent gains, and directly training models to learn to abstain proved unstable.76% similarUnverifiedFoundation model evaluations should include simple non-learning control checks (such as handcrafted features); if a model cannot even outperform handcrafted features, its claim to have learned useful representations is untenable.73% similarUnverified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72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/114525API
curl https://kongchang.com/api/v1/knowledge/claims/114525MCP
get_claim(id=114525)