Safety
113d agoNLAs Generate Natural Language LLM Explanations
Two LLM modules—an activation verbalizer and reconstructor—now map internal activations to text descriptions via reinforcement learning. This unsupervised method, called Natural Language Autoencoders, translates complex residual stream data into human-readable interpretations. Researchers used the tool to audit Claude Opus 4.6.
The Signal
It provides a concrete path for diagnosing safety-relevant behaviors before deployment.