WriteSAE: Sparse Autoencoders for Recurrent State
Researchers have developed WriteSAE, a novel sparse autoencoder designed to manipulate the matrix updates within recurrent language model states. This method learns rank-1 matrix atoms that directly replace the model's own matrix updates, showing a significant improvement in final token distribution accuracy. The technique has been successfully applied to models like Gated DeltaNet and Mamba-2, demonstrating its potential for steering model generation and understanding internal state dynamics. AI
IMPACT Enables direct intervention and steering of recurrent language model states, potentially leading to more controllable and understandable AI generation.