Unverified50% confidenceFactExact time
Dictionaries trained by SAEs differ significantly across random seeds, undermining the reliability and reproducibility of analyses.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe 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.73% similarUnverified即便固定所有随机种子，在不同硬件或并行度下，训练结果仍可能出现被大模型放大为显著能力差异的细微偏差67% similarUnverifiedIn the reproduction experiments, the variance across different random seeds was almost as large as any trend across training set sizes N.66% similarUnverifiedWhen randomized experiments are not feasible, causal inference methods become critical, including Difference-in-Differences (DID), Regression Discontinuity Design (RDD), Instrumental Variables (IV), and Propensity Score Matching (PSM).65% similarUnverifiedThe results of a single-task evaluation are subject to randomness, and the same model may produce noticeably different answers under different prompts and different random seeds.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065908API
curl https://kongchang.com/api/v1/knowledge/claims/1065908MCP
get_claim(id=1065908)