Unverified50% confidenceSolutionExact time
IKNO retains only the necessary invariant scalars and l=1 vector features, bypassing the overhead of higher-order tensors while achieving equivalent equivariance guarantees.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedK2.5的MoE架构单次推理实际激活的参数量可能只有总参数的1/8甚至更少62% similarUnverifiedThe core trade-off of MoE is large capacity at low compute: the activated parameters are only a fraction of the total parameter count, but memory requirements are not reduced accordingly.60% similarUnverifiedRMSNorm 相比传统 LayerNorm 去掉了均值中心化步骤，仅保留均方根归一化，计算更轻量59% similarUnverified哈达玛变换用于量化预处理时不会引入额外的数值不稳定性，因为其条件数恒为159% similarUnverifiedNALD and NRALD preserve the target distribution invariant without requiring computation of derivatives of the target density.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990614API
curl https://kongchang.com/api/v1/knowledge/claims/990614MCP
get_claim(id=990614)