Training Seed
The initial state of the random number generator set before model training, which affects stochastic processes such as weight initialization and data ordering and leads to differences in training outcomes.
Timeline (last 90 days)
The sparse mapping between features can explain 0.85 of a graph's energy on its active features, but when another seed's graph retains only its own features, this proportion drops to 0.47.
SAEs trained with different seeds may learn feature dictionaries that are semantically similar but differ in specific direction or magnitude, leading to differences in the structure of attribution graphs.
The margin effect may amplify differences between seeds: features near the threshold have different activation states under different seeds, thereby affecting the structure of the graph.
All Facts (3)
The sparse mapping between features can explain 0.85 of a graph's energy on its active features, but when another seed's graph retains only its own features, this proportion drops to 0.47.
50%UnverifiedSAEs trained with different seeds may learn feature dictionaries that are semantically similar but differ in specific direction or magnitude, leading to differences in the structure of attribution graphs.
50%UnverifiedThe margin effect may amplify differences between seeds: features near the threshold have different activation states under different seeds, thereby affecting the structure of the graph.
50%