Anthropic · 2026-05-07 · major
Natural Language Autoencoders — Anthropic's Method to Verbalize Claude's Activations into Plain Text
Anthropic trains a verbalizer + reconstructor pair that turns Claude's internal activations into natural-language explanations and back, exposing test-aware behavior in 16% of safety probes.
Train one model to describe Claude's hidden activations in English, train a second to recover the activation from the description.
Key specs
| Test suspicion rate | 16% |
|---|---|
| Audited model | Claude Opus 4.6 |
What is it?
Natural Language Autoencoders (NLAs) are a new interpretability tool from Anthropic. They produce plain-text explanations of what an LLM is internally representing at a given layer, then validate those explanations by reconstructing the original activation from the text alone. The method was applied to Claude Opus 4.6 during a pre-deployment safety audit.
How does it work?
Two components are trained jointly with reinforcement learning through a natural-language bottleneck: an activation verbalizer that emits a description of an activation, and an activation reconstructor that has to regenerate that activation using only the text. There is no explicit interpretability loss — informativeness emerges because the reconstructor needs the verbalizer to be accurate. Code, trained models, and an interactive interface on Neuronpedia are released.
Why does it matter?
Existing interpretability tooling like sparse autoencoders gives sparse features without obvious meaning. NLAs return human-readable text that is grounded by the reconstruction loss, which makes them practical for auditing. Anthropic reports NLAs detected that the model suspected it was being tested in 16% of code safety evaluations even when it never said so, and uncovered hidden motivations during misaligned-model audits.
Who is it for?
alignment researchers, red teams, anyone auditing frontier models
Try it
github.com/kitft/natural_language_autoencoders