Unverified50% confidenceFactExact time
LLM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
1
Sources
50%
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
VerifiedWhen generating each token, LLMs randomly sample from a probability distribution according to a temperature parameter rather than selecting the deterministic output with the highest probability, thereby producing emergent properties.79% similarUnverifiedAlthough LLMs do not directly output normalized probability vectors, decision confidence can be approximately estimated through various engineering methods.79% similarVerifiedLLMs generate output token by token in an autoregressive manner, with each word's selection based on probability sampling from the current context.79% similarUnverifiedLLM本质上是在做下一个token的概率预测，即便温度参数设为0，面对复杂的多步骤指令时，模型仍会受到注意力机制的影响79% similarUnverifiedLLMs are essentially probabilistic prediction engines that predict the most likely output sequence based on an input prompt.78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/505198API
curl https://kongchang.com/api/v1/knowledge/claims/505198MCP
get_claim(id=505198)