Anthropic Introduces Natural Language Autoencoders to Read Claude's Internal Thinking

Summary

Anthropic has published research on Natural Language Autoencoders (NLAs), a technique that converts the numerical activations inside Claude's hidden layers directly into natural-language text explanations that anyone can read. The system uses a two-component architecture: an Activation Verbalizer (AV) that converts model activations into text, and an Activation Reconstructor (AR) that attempts to recreate the original activation from the description. Accuracy is self-validating — if the explanat