Game Theory: This makes it MATHEMATICALLY IMPOSSIBLE for AI to Lie
Fractional Block Sensitivity: The Math Trick That Traps Dishonest AI.
The Dual LP Exploit: How to Catch an AI in a Lie EVERY Time.
Forget Human Feedback. AI "Debate" is the Future of AGI.
PlurPO vs. Sycophancy: The Next Massive Leap in AI Alignment!
all rights w/ authors:
MITIGATING SOCIAL SYCOPHANCY VIA PLURALISTIC
PREFERENCE OPTIMIZATION
Stephane Hatgis-Kessell1, Myra Cheng1, Xiaoxuan Hou2,3, Qian Hu2, Rahul Gupta2,
Natasha Jaques3, Emma Brunskill1
1Department of Computer Science, Stanford University, 2Amazon,
3Department of Computer Science, University of Washington
Correspondence: stephh