Unverified60% confidenceFactExact time
Explicit instructions suppress implicit behavioral strategies the model has learned during training, leading to an overall decline in output quality—a phenomenon known as 'instruction override'.
2
Sources
60%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified指令微调不需要从头训练，通常仅需基础训练计算量的1%以下即可显著改变模型行为模式75% similarUnverifiedThe paper gives two examples of behavioral coupling: training a model to be warmer increases sycophancy; training a model to write insecure code leads to broad misalignment unrelated to programming.74% similarUnverifiedTraditional draft models are trained on narrow distributions targeting a single model, causing acceptance rates to drop sharply when workloads change, and requiring the entire pre-training process to be redone each time a target model is replaced.74% similarUnverifiedTraining thinking-length preferences into the model weights is more stable than using system prompts at deployment time to request concise answers, as the latter is an external constraint that is unstable while the former becomes the model's default behavior.74% similarUnverifiedThe optimal guidance weight formula reveals that external guidance should be trusted more in the early stages of training when the model's own variance is large, and that the guidance weight should be gradually reduced as training progresses, providing a theoretical basis for guidance annealing strategies.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/894950API
curl https://kongchang.com/api/v1/knowledge/claims/894950MCP
get_claim(id=894950)