A simpler and better debate protocol for scalable oversight

The researchers propose a new debate protocol for scalable oversight called FBS Debate, claiming its rules are simpler yet provide stronger correctness guarantees than prover-estimator debate. In the protocol, the first debater splits the original proposition into sub-propositions; the second either points out that the split is overly sensitive to small errors in the sub-propositions, or randomly samples a sub-proposition to continue the recursive debate, until a sub-proposition is directly judged by a human. The paper characterizes the verification compute required for recursive debate using fractional block sensitivity, claims the honest strategy is optimal on every problem, and claims that no protocol can be asymptotically more efficient under the assumption that human judgment is treated as a black box. The authors believe these conclusions are also informative for debate experiments that do not use FBS Debate.