AI brief
A LessWrong post reports stress-testing alignment midtraining across model and token budget scales, concluding it cannot address distributional shift and reward underspecification under imperfect conditions.
Written by AI from LessWrong's published text.
Source
“TL;DR We stress-test alignment midtraining (AMT) across model and token budget scales.”
Interpretation or community commentary rather than straight reporting.
Read original story ↗