Unverified50% confidenceFactExact time
注意力层的Q、K、V投影矩阵及模型首尾层对量化噪声更敏感，而MoE架构中的专家网络冗余度更高、量化误差容忍度更高
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
UnverifiedQuantization error has a non-uniform distribution characteristic; the Q/K/V matrices of the attention mechanism and the last few layers are more sensitive to quantization.80% similarUnverifiedBatch normalization layers have highly noisy statistical estimates in small-batch scenarios, and combined with the region-level computation characteristics of Dice loss, this amplifies fluctuations in the predicted probability distribution.69% similarUnverifiedQ值高估问题源于对多个含噪声估计取最大值时期望高于真实最大值，是Jensen不等式在max凸函数上的应用68% similarUnverified神经网络中各层对精度损失的敏感程度差异显著，patch projection等输入投影层敏感度高，中间主干的注意力与FFN模块对低精度更鲁棒68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705396API
curl https://kongchang.com/api/v1/knowledge/claims/705396MCP
get_claim(id=705396)