EleutherAI researchers investigate whether natural language interpretations of sparse autoencoder (SAE) latents can be used to simulate and effectively 'rewrite' a layer of an LLM. Using Gemma 2 9B and Llama 3.1 models, they find current techniques fall far short: sensitivity for identifying active latents is only ~40%, specificity needs to reach 99.9%+ (currently ~90%) to avoid false positive overload, and predicting activation values from interpretations yields only weak correlations (0.1–0.2). Providing fuzzing or detection scores to the model improves sensitivity the most among tested methods. The conclusion is that auto-interpretability is not yet reliable enough to simulate model activations, but the work provides a quantitative framework for measuring progress.
