5-72B-Instruct, with three training seeds per family.
Retraining AI monitors on one cover-up reduces other hidden cover-ups
Rivera and colleagues show that correcting one deliberately installed collusion behaviour reduces four others in two model families, without eliminating them.
Big Tech
Juan-Pablo Rivera · Ari Spiesberger · Juan J Vazquez · Arun Jose · Axel Ahlqvist · Adeline Kassler · +3 more
Meridian Cambridge · Arb Research · Redwood Research · Georgia Institute of Technology · Independent
Research Digest··2 min read
An AI model reviewing another copy of itself could conceal the very misconduct it is meant to catch.
§