Unverified50% confidenceFactExact time
自回归生成中通过softmax和argmax从连续隐藏状态选择离散token是一次硬量化操作，会不可逆地丢失大量精细的中间信息
1
Sources
50%
Confidence
Long-term
Relevance
8/11/2026
First Seen
Sources
潜空间推理：DeepSeek-V4如何让AI在隐藏层中思考
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.69% similarUnverified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67% 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67% similarUnverifiedCurrent mainstream large models excel at handling isolated, well-bounded small tasks, but expose shortcomings of incomplete understanding and loss of detail when faced with highly coupled, large-context legacy systems.67% 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67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/729943API
curl https://kongchang.com/api/v1/knowledge/claims/729943MCP
get_claim(id=729943)