Unverified50% confidenceFactExact time
Softmax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
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
元认知反馈：用强化学习让大模型准确表达不确定性
hackernewshackernews7/7/2026
Related Claims
UnverifiedSoftmax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85% similarUnverifiedSoftmax函数将神经网络logits压缩为概率分布，但其输出值在大规模模型中系统性偏离真实贝叶斯后验概率81% similarUnverifiedThe softmax operation is prone to numerical overflow or underflow due to insufficient precision when handling extreme values, which is the core technical reason why early FP8 quantization performed poorly.77% similarUnverifiedWhen a model encounters out-of-distribution inputs, prediction entropy increases, the softmax output tends toward a uniform distribution, and the range of candidate tokens permitted by nucleus sampling expands, giving long-tail anomalous tokens a chance of being sampled.72% similarUnverified结构化提示（角色设定+任务背景+约束条件+输出格式）能降低模型的解码不确定性，使 softmax 概率质量集中到预期输出区间72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/279889API
curl https://kongchang.com/api/v1/knowledge/claims/279889MCP
get_claim(id=279889)