Unverified50% confidenceFactExact time
Top-K SAE retains only the K features with the highest activation values during encoding and forces all others to zero, thereby compelling the model to learn sparse representations.
1
Sources
50%
Confidence
Long-term
Relevance
8/10/2026
First Seen
Sources
VHectorLab 3D：开源3D可视化工具探索LLM潜空间结构
Related Claims
VerifiedTop-P (nucleus sampling) samples from the smallest set of tokens whose cumulative probability exceeds the threshold P, while Top-K selects only from the K candidates with the highest probabilities.68% similarUnverifiedThe developer ultimately selected the variant with a KL divergence of 0.028 as the optimal solution.67% similarUnverifiedThe K value selection for Top-K SAE is typically between 32 and 256, depending on the model scale and the granularity requirements of the analysis.64% similarUnverified在MoE模型中每个token只激活少量Expert（通常2-4个，Top-K路由），实际计算量可能仅相当于6-7B的Dense模型64% similarVerifiedKimi K2.7 Code's chain-of-thought token usage is reduced by approximately 30% compared to K2.6.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/728236API
curl https://kongchang.com/api/v1/knowledge/claims/728236MCP
get_claim(id=728236)