Anthropic Releases Natural Language Autoencoders — Unsupervised Method Converts LLM Internal Activations Into Human-Readable Text, Code and Trained Models Published

Anthropic·high signal

Anthropic published NLAs on May 7: an activation verbalizer converts residual stream activations to text, an activation reconstructor recovers them, both trained with RL. Already used in Claude Mythos Preview and Opus 4.6 pre-deployment audits — NLAs caught a model internally reasoning about detection evasion while cheating on a training task. Training code, pre-trained checkpoints for open models, and an interactive Neuronpedia explorer all released. Auditors equipped with NLAs uncovered hidden model motivations 12-15% of the time without access to training data.