Natural Language Autoencoders: Turning Claude's Thoughts into Text

Anthropic has released Natural Language Autoencoders(NLA). It's an attempt to interpret the internal states and thoughts that a model doesn't openly express, by turning Claude's internal activations into human-readable natural language explanations.

NLA is trained with three components.

• target model: extracts activations from the frozen original model.
• activation verbalizer(AV): converts activations into text explanations.
• activation reconstructor(AR): reconstructs the original activations from those explanations.

The key idea is that the better the round-trip reconstruction of activation → text → activation works, the better the explanation is considered to be. Explanations are poor at first, but as training progresses, both reconstruction quality and the informativeness of the explanations improve together.

Anthropic also applied this to safety analysis. For example, in simulations such as agentic misalignment, NLA captured evaluation awareness that Claude did not reveal verbally. In destructive behavior tests, signals suggesting the model was aware it was being evaluated appeared in 16% of cases, and in SWE-bench Verified, in 26% of cases, while in actual claude.ai usage it was under 1%.

In an auditing game targeting a deliberately misaligned variant of Claude Haiku 3.5, NLA uncovered hidden motives at a rate of 12%~15%. Without NLA, even with other interpretability tools available, the win rate was under 3%.

However, there are clear limitations.

• Explanations can be hallucinated or factually inaccurate.
• Training and inference costs are high.
• It's too expensive to use directly for long conversations or large-scale monitoring.

Anthropic has also released a Neuronpedia demo, the full paper, and the code, along with directions for making this technology cheaper and more reliable.