Unverified50% confidenceOpinionExact time
The features disentangled by SAEs actually have response widths several times coarser than the original neurons, meaning that SAEs may create a mathematical illusion.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUsing SAE decomposability as a proxy for smaller, more easily reverse-engineered circuits may be misleading in low-faithfulness settings.70% similarVerifiedSparse autoencoders (SAEs) introduce sparsity regularization in addition to reconstruction loss, encoding activation vectors as overcomplete sparse representations.69% similarUnverified叠加现象指神经网络倾向于用远少于特征数量的神经元同时编码大量语义特征，与传统一个神经元一个特征的直觉相悖67% similarUnverifiedWhen ε ≈ 1, the gradient second-moment term in the denominator is diluted, adaptive scaling has almost no effect, and the optimizer behaves close to pure SGD.64% similarUnverifiedWhen directly computing an arithmetic average of two models' parameters, neurons at corresponding positions may serve different functions, and the averaged result can destroy the feature representations learned by each model, leading to a significant performance drop or even degradation to random-level performance.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1045667API
curl https://kongchang.com/api/v1/knowledge/claims/1045667MCP
get_claim(id=1045667)