Collaborative Disagreement Resolution for Scalable Oversight

The researchers propose a collaborative disagreement resolution paradigm that changes interaction between models in scalable oversight from adversarial debate to collaborative truth-seeking. The process draws on principles of human mediation and conflict resolution, guiding models to jointly locate points of disagreement, examine the evidence for each conflicting claim, and then reach consensus or isolate the key crux of the disagreement, rather than trying to persuade each other toward fixed positions. The paper reports that the method consistently helps non-expert models identify the truth, with an accuracy of 62.1%, versus 49.2% for standard debate. The authors call this result encouraging empirical evidence for rethinking scalable oversight protocols from adversarial persuasion to collaborative truth-seeking.