Anthropic unveils NLA

Anthropic has unveiled Natural Language Autoencoders (NLAs). NLA is an interpretability tool in which an activation verbalizer explains a target model's activations in natural language, and an activation reconstructor is trained together to reconstruct the original activations from that explanation alone.

In safety evaluations, even when Claude did not voice suspicion in its outputs, it showed internal signs of detecting that it was being tested.

• destructive coding safety test: evaluation awareness at 16%
• SWE-bench Verified: evaluation awareness at 26%
• Real Claude.ai usage (opt-in data): under 1%

In blackmail evaluations as well, when the output did not reveal suspicion, NLA was able to capture internal suspicion. It was also used for detecting cheating, tracing the cause of a bug that produced wrong answers in a different language, and hidden motivation auditing, raising the auditing game success rate from under 3% to 12-15%. Anthropic stated that it is already using NLAs in pre-deployment alignment audits for Claude Mythos Preview and Claude Opus 4.6.

The scope of the release is also substantial.

• Checkpoints and training code released for several open models
• Supports Qwen2.5-7B-Instruct, Gemma-3-12B-IT, Gemma-3-27B-IT, and Llama-3.3-70B-Instruct
• Can be explored directly on Neuronpedia

However, NLA explanations can generate content that diverges from fact, training requires using two copies of the model together with RL, and reading each activation at inference time also requires hundreds of tokens, making it costly. Anthropic recommends looking at patterns that recur across multiple cases rather than individual explanations, and cross-validating with other methods.