Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
SAE reconstruction converges on far less data, but one-to-one feature matches keep rising with more tokens.
Stories per day · 30 days
1 RSS
Relatedfeature-reproducibility1gemma-21interpretability1mechanistic-interpretability1sparse-autoencoders1
3 stories, newest first · one card per story
Stories tagged #data-efficiency
Does Native 3D Texture Generation Necessarily Require 3D Assets for Training?
Tex-Zero trains native 3D texture models from constructed 2D image planes, without real 3D assets.
Revisiting Complete Reasoning Traces for Post-Training
Researchers show full reasoning traces provide limited benefit in LLM post-training, with heavily truncated or endpoint-only trajectories performing comparably.