Search papers, labs, and topics across Lattice.
1
0
SAE-based explanations can be certified for faithfulness, revealing that later layers of language models are significantly easier to interpret without sacrificing predictive accuracy.
