Safety
111d agoNLAs Translate LLM Activations Into Natural Language
Natural Language Autoencoders use two LLM modules to map internal activations to text descriptions. Trained via reinforcement learning, these NLAs reconstruct residual stream activations to produce human-readable interpretations of model internals. Researchers used this method during a pre-deployment audit of Claude Opus 4.6.
The Signal
It provides a concrete tool for diagnosing safety-relevant behaviors before release.