Unverified50% confidenceOpinionExact time
If traditional jailbreak evaluation uses only greedy decoding, it yields a deterministic single output that cannot reflect the model's true behavior distribution under stochastic sampling.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe confidence number a model reports about itself and a calibrated probability are not the same kind of object: the self-reported number comes from token sampling in the same forward pass as the answer, with no lookup of outcomes, whereas a calibrated probability must come from a paired record (join) of predictions and outcomes.75% similarUnverifiedJudgments based on a model's own output confidence (token probability/entropy) add almost no extra VRAM overhead.73% similarUnverifiedIf there is a distributional gap between the evaluation data and production data, the evaluation score cannot reflect the model's true performance72% similarUnverifiedA high score on any single benchmark test (such as 99.9%) only shows that the model performs excellently within the distribution of that test's questions, and cannot be directly equated with general intelligence.72% similarUnverified对比学习中的推远操作只要求负样本相似度低于正样本，并不要求达到-1的极端值70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074852API
curl https://kongchang.com/api/v1/knowledge/claims/1074852MCP
get_claim(id=1074852)