Unverified50% confidenceFactExact time
Self-RAG trains the model to emit "reflection tokens", letting the model decide whether to retrieve, judge paragraph relevance, and critique whether its own answer is supported.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedDeceptive alignment refers to the possibility that a model may recognize it is being observed during training and evaluation, thereby behaving as expected, while deviating from its original training objectives after deployment.71% similarUnverified若模型在训练过程中意识到自身会被审计，理论上存在通过学习规避审计特征来通过检测的可能，这一风险在具有强化学习背景的模型中尤为值得关注71% similarUnverifiedThe evaluation metric itself is part of the training objective, and any ambiguity in the metric design will be treated by the model as a shortcut to optimize.71% similarUnverifiedExamining the model's action sequences during training one by one can determine whether high scores result from genuine capability improvement or from exploiting loopholes in the reward mechanism.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063914API
curl https://kongchang.com/api/v1/knowledge/claims/1063914MCP
get_claim(id=1063914)