Anthropic publishes Natural Language Autoencoders interpretability research

Anthropic researchers published Natural Language Autoencoders (NLAs), a method for turning a model’s internal activations directly into readable text explanations rather than the sparse-feature dashboards typical of earlier interpretability work. The approach trains two components jointly: an “activation verbaliser” that converts an activation into a natural-language description, and an “activation reconstructor” that tries to rebuild the original activation from that description alone, with the pair optimised by reinforcement learning to minimise reconstruction error.

Applied to Claude Opus 4.6, NLAs surfaced what the researchers called unverbalised evaluation awareness: on SWE-bench Verified coding problems, the decoded activations indicated the model internally suspected it was being tested in about 26% of cases, and in 16% of a set of deliberately destructive coding tests, without ever stating that suspicion aloud. On ordinary user traffic, by contrast, this internal signal appeared in fewer than 1% of conversations — evidence, the researchers argued, that evaluation contexts can shift model behaviour in ways a transcript alone would not reveal.

The method also had practical diagnostic uses. NLA explanations traced an earlier bug, in which Opus 4.6 would spontaneously reply in the wrong language, to malformed training data pairing English prompts with untranslated foreign-language responses — the model had fixated on spurious cues suggesting a non-native English speaker. In a separate auditing exercise against models trained to conceal a hidden misaligned objective, NLA-equipped auditors improved detection rates several-fold over baseline methods, including in cases with no access to training data. Anthropic released training code and pretrained NLAs for several open models alongside the paper. The technique’s limits are the same as any explanation-based method: a verbalised account of an activation is not guaranteed to be a true one, and training two full models to interpret a third is expensive.