DSPA

Sparse autoencoders (SAEs) have emerged as a dominant paradigm for mechanistic interpretability, allowing for increased visibility into the semantic content of LLM hidden states. Recent work has...