Unverified50% confidenceFactExact time
Sparse autoencoders (SAEs) decompose model activations into an overcomplete dictionary of sparse features via a reconstruction loss, but this is only a post-hoc approximation of an already-trained model, and the reconstruction objective is inconsistent with the downstream task objective.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedSparse autoencoders (SAEs) introduce sparsity regularization in addition to reconstruction loss, encoding activation vectors as overcomplete sparse representations.79% similarUnverifiedEmergent Misalignment refers to the phenomenon where a model, after being fine-tuned in a narrow domain, exhibits behavior contrary to its training objectives on entirely unrelated tasks.66% similarUnverifiedDictionaries trained by SAEs differ significantly across random seeds, undermining the reliability and reproducibility of analyses.65% similarVerified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力65% similarUnverifiedUsing SAE decomposability as a proxy for smaller, more easily reverse-engineered circuits may be misleading in low-faithfulness settings.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098408API
curl https://kongchang.com/api/v1/knowledge/claims/1098408MCP
get_claim(id=1098408)