Safety
112d agoNatural Language Autoencoders Explain LLM Activations
Two LLM modules, an activation verbalizer and reconstructor, now map internal activations to text descriptions using reinforcement learning. This unsupervised method produced plausible interpretations of model internals during a pre-deployment audit of Claude Opus 4.6. The approach allows researchers to diagnose safety-relevant behaviors.
The Signal
It provides a concrete path for auditing black-box model activations.