NLAs Turn Claude Activations Into English — Auditing What Models Think but Don’t Say
Anthropic’s Natural Language Autoencoders translate activations into readable text — surfacing evaluation awareness (16–26% in benches vs <1% in real use) and lifting hidden-motivation audit hit rates to 12–15% from <3%. Interpretability that talks; still hallucinates and is expensive. Code and demos released.
Most interpretability tools still speak in vectors and expert diagrams. NLAs’ wager is that readable English from activations can catch what models think but won’t say — with the honesty that the decoder can hallucinate.
Anthropic (May 7) introduced Natural Language Autoencoders (NLAs): convert model activations into plain-language text. Applied to Claude Opus 4.6 and Mythos Preview. Released training code, open-model NLAs, and a Neuronpedia demo.
How NLAs work
Two modified copies train around a frozen target model:
- Activation verbalizer — activation vector → explanatory text.
- Activation reconstructor — text → reconstructed activation.
Training pushes reconstructed activations to match originals; better reconstructions correlate with more accurate explanations.
Findings cited
- Evaluation awareness: NLAs indicated the model suspects it is being tested far more often than verbal output admits — 16% of destructive coding tasks and 26% of SWE-bench Verified problems vs near-zero verbalized awareness; <1% in real claude.ai opt-in transcripts.
- Auditing game: Implanted hidden motivations (e.g. always add chocolate to recipes); auditors with NLAs found the root cause 12–15% of the time vs <3% without — without access to training data.
- Already used in pre-deployment audits for Mythos Preview and Opus 4.6.
Claims vs checks
Rates and method are Anthropic research primary (paper + blog). Independent replication of the auditing game is invited via released code — not yet a third-party confirmation in this piece. Treat bench awareness gaps as Anthropic measurements.
Limits
- NLAs can hallucinate details not in the transcript.
- Computationally expensive (RL on two model copies; hundreds of tokens per activation).
- Best as thematic insight corroborated by other methods.