arXiv paper proposes instance-optimal protocol for AI debate with worst-case guarantees
The new arXiv preprint (arXiv:2610.02557v1) designs a debate protocol for a class of problems with stable decompositions that improves prior work by giving worst-case correctness guarantees, making honesty a dominant-strategy equilibrium for both debaters rather than a Stackelberg equilibrium, and proving black-box lower bounds that show instance-wise optimality. The authors connect stable problem decompositions to fractional block sensitivity from query complexity to obtain these results.