Unverified50% confidenceFactExact time
The dictionary optimization problem of SAEs is non-convex and has many local optima, so different random seeds cause dictionary atoms to rotate or recombine in concept space.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDictionaries trained by SAEs differ significantly across random seeds, undermining the reliability and reproducibility of analyses.73% similarUnverified2023年的单义性研究探索了如何通过稀疏自编码器（SAE）在更高维稀疏激活空间中分离混叠特征63% similarUnverifiedWhen one component of a system already has randomness, stacking a probabilistic component on top of it exponentially amplifies debugging difficulty.63% similarUnverifiedSetting a random seed can only guarantee consistency in initialization, data order, and explicit random sampling; it cannot eliminate non-determinism caused by floating-point parallel computation during backpropagation.60% similarUnverifiedThe features disentangled by SAEs actually have response widths several times coarser than the original neurons, meaning that SAEs may create a mathematical illusion.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065917API
curl https://kongchang.com/api/v1/knowledge/claims/1065917MCP
get_claim(id=1065917)