Unverified50% confidenceSolutionExact time
计算 softmax 函数时指数运算可能导致浮点数溢出，标准做法是先减去输入向量的最大值再取指数
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
UnverifiedThe 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.79% 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74% similarUnverifiedSoftmax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71% 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69% similarUnverified缩放点积将点积除以√d（d为嵌入维度），使点积方差保持在合理范围，让Softmax输出更平滑的概率分布以稳定训练68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/689952API
curl https://kongchang.com/api/v1/knowledge/claims/689952MCP
get_claim(id=689952)