Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens
Paper • 2609.01936 • Published • 3
Sparse Readout Prism
Factorizing a model's final unembedding into a sparse feature basis for logit-lens analysis.
Factorizing a model's final unembedding into a sparse feature basis for logit-lens analysis.