Safety
113d agoNLAs Translate LLM Activations Into English
Natural Language Autoencoders use two LLM modules to map internal activations to text descriptions and back. Researchers trained this system via reinforcement learning to reconstruct residual stream activations. During a pre-deployment audit of Claude Opus 4.6, the tool diagnosed safety-relevant behaviors.
The Signal
This unsupervised method provides interpretability practitioners a concrete way to audit model internals.