From Sparse Features to Trustworthy Proxies: Certifying SAE-Based Interpretability
Researchers have developed a new framework to certify the interpretability of Sparse Autoencoders (SAEs) when used with language models. This framework establishes an upper bound on the risk of a language model by using a sparse proxy derived from SAE reconstructions. The method has been shown to be effective on models like GPT-2 Small, Gemma-2B, and Llama-3-8B, with later layers of Llama-3-8B proving easier to certify. The approach helps distinguish genuine semantic alignment from mere statistical sparsity, offering a diagnostic tool for the reliability of SAE-based explanations. AI
IMPACT Provides a new method for understanding and verifying the internal workings of language models, potentially improving trust and debugging.