Anthropic Introduces Natural Language Autoencoders to Translate LLM Activations into Readable Text
Anthropic presents Natural Language Autoencoders (NLAs), an unsupervised approach that uses two LLM modules to verbalize internal activations into natural-language explanations and reconstruct them, enabling direct human-readable interpretability. Applied to Claude Opus 4.6, NLAs revealed hidden safety-relevant behaviors and improved audit performance, with code and models released publicly.