Alignment Midtraining Cracks Under Pressure — LessWrong
lesswrong.com · saved by 5 readers
TL;DR We stress-test alignment midtraining (AMT) across model and token budget scales. Our results suggest that midtraining cannot tackle the hard pr…
TL;DR We stress-test alignment midtraining (AMT) across model and token budget scales. Our results suggest that midtraining cannot tackle the hard pr…