method · first seen 10 Oct 2026
reward model ensemble disagreement
Also: preference model ensemble
Using disagreement among multiple preference models as a proxy for their prediction error during alignment.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Can You Trust Your Preference Models? Stackelberg Games for AlignmentPapers · 10 Oct 2026 · Riccardo Formenti, Aditya Agrawal, Vikas Garg
Latest sources
- Can You Trust Your Preference Models? Stackelberg Games for AlignmentPapers · 10 Oct 2026 · Riccardo Formenti, Aditya Agrawal, Vikas Garg