Verified65% confidenceFactExact time
Although LLMs do not directly output normalized probability vectors, decision confidence can be approximately estimated through various engineering methods.
3
Sources
65%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
AI Agent人在回路（HITL）设计：风险分级与最佳实践
Related Claims
UnverifiedLLM输出的置信度来自logits或token概率，是模型内部的相对偏好，并不等同于事件发生的真实频率，即模型可能过度自信。79% similarUnverifiedThe implications for LLM engineering practice include: do not directly treat the probabilities output by the model as decision confidence; in scenarios where handoff to a human is not possible, a conservative fallback strategy should be explicitly designed; threshold selection requires statistical guarantees.79% similarUnverifiedLLM可通过分析输出token的对数概率分布、思维链末尾自我评估、温度采样统计答案分布熵值等手段近似估计决策置信度79% similarUnverified当前 LLM 在代码生成时是基于概率分布生成下一个 token，而非基于工程经验做出系统性设计决策79% similarUnverifiedLLMs' probability judgments drift under the influence of irrelevant factors such as how evidence is framed and the order in which it is presented.79% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/507715API
curl https://kongchang.com/api/v1/knowledge/claims/507715MCP
get_claim(id=507715)