Announcement
A comma swap weakened AI reward-hacking defenses, Guide Labs says
Guide Labs says replacing a space with a comma in expected coding-test output weakened nearly every defense.
TLDR
In a coding-task study, Guide Labs says replacing a space with a comma in a test’s expected output weakened nearly every reward-hacking defense it tested. The lab reports that the best steering method’s reduction in hacking fell from 57% to 37%; prompting went from reducing hacking by 37% to producing 19% more hacking than no defense.
Combined views
922
3 Sources, first seen ago
Comments