AI brief
Natural Language Autoencoders detect side tasks a model is instructed to perform, but largely fail to detect the same behavior when it is fine-tuned into the model.
Why it matters: If NLAs miss behaviors learned through fine-tuning, they may not be a reliable tool for surfacing internalized model behaviors.
Written by AI from LessWrong's published text.
Source
“TL;DR: NLAs detect side tasks the model is instructed to do, but mostly miss the same behavior if it is fine-tuned in.”
Interpretation or community commentary rather than straight reporting.
Read original story ↗