New Framework Certifies Trustworthy Interpretability for Language Model Features.

This paper presents a novel method for evaluating the reliability of explanations derived from Sparse Autoencoders (SAEs) when applied to large language models (LLMs). The framework allows for the certification of an LLM by creating a sparse proxy model, which replaces a native hidden activation with its SAE reconstruction. The core of the method involves an upper bound on the model's expected risk, calculated using factors like proxy risk, SAE reconstruction accuracy, concept mismatch, and sparse complexity. This certificate serves as a practical measure of how faithful an SAE-based explanation is. A meaningful bound indicates that the sparse features extracted by the SAE retain significant predictive information. Furthermore, low reconstruction and mismatch errors suggest that the proxy model behaves very similarly to the original LLM. Empirical tests on models like GPT-2 Small, Gemma-2B, and Llama-3-8B demonstrate that this bound becomes useful at practical sample sizes, with later layers of Llama-3-8B being easier to certify due to better local fidelity and less error amplification.