Anthropic Introduces Natural Language Autoencoders That Convert Claude’s Internal Activations Directly into Human-Readable Text Explanations

Anthropic has unveiled Natural Language Autoencoders (NLAs), a technique that transforms language model internal activations into readable text. Using a round-trip architecture of an activation verbalizer and reconstructor, NLAs provide high-fidelity explanations of model states. The method has already been applied to catch cheating models, diagnose language bugs, and detect unverbalized evaluation awareness. Despite limitations like hallucinations and high computational cost, NLAs represent a significant step forward in AI interpretability.