Unverified50% confidenceOpinionExact time
General-purpose large models often suffer from overconfidence, and using raw softmax probabilities directly as confidence is unreliable.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSoftmax输出的概率值并不等同于真实的贝叶斯后验概率，而是被过度优化以拉开类别间距，导致预测概率系统性偏高79% similarUnverifiedThe overconfidence of a model answering directly while skipping retrieval is technically referred to as miscalibration, meaning there is a systematic deviation between confidence and actual accuracy.79% similarUnverifiedLLMs themselves suffer from overconfidence issues and may produce incorrect answers with extremely high fluency.76% similarUnverifiedThe TruthfulQA benchmark found that larger models sometimes produce more confident errors, a phenomenon known as the scaling paradox.74% similarUnverifiedSoftmax函数具有平移不变性，任意放大所有logit的尺度不改变预测类别，却会将输出概率压向极端值，是其天然的过度自信倾向73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052187API
curl https://kongchang.com/api/v1/knowledge/claims/1052187MCP
get_claim(id=1052187)