Sparse Shift Autoencoders for Identifying Concepts from Large Language Model Activations

Researchers introduce Sparse Shift Autoencoders (SSAEs), a new method for improving large language model interpretability by learning sparse representations of differences between embeddings rather than the embeddings themselves. This approach addresses the identifiability problem in current sparse autoencoder techniques, potentially enabling more precise control over specific AI behaviors without unintended side effects.

• →SSAEs solve the identifiability problem that plagues current sparse autoencoder approaches in LLM interpretability.
• →The method learns representations of differences between embeddings rather than the embeddings directly.
• →SSAEs can identify and steer single concepts with only weak supervision from paired observations.
• →The approach reduces risk of unintended interventions when steering specific LLM behaviors.
• →Empirical validation shows successful concept recovery across multiple real-world language datasets and different LLMs.