Adversarial proceduralism

Two AI systems are assigned opposite sides of a question and debate it in front of a judge. The judge lacks something the two systems have (the source passage, the evidence, the time, or the raw capability) and so cannot settle the question alone. Because each side gains by exposing the other's errors, the judge hears the arguments against each claim before deciding. The papers in this approach measure whether judges actually become more accurate. They test this on passages the judge has not read, on contested COVID and climate claims judged by people with prior beliefs, and in comparison with a single consultant who argues without an opponent. One variant drops fixed sides altogether. The two systems are instead directed to find the exact point on which they disagree and to give the judge that point rather than a finished argument.

A deliberately small and limited checker is trained to question a far more capable arguer that it has no reason to trust. Meanwhile, a second arguer is rewarded for getting wrong answers past the checker. Only answers that the weak checker can follow and confirm survive this training loop. The result is reasoning that a less capable reader (e.g., a small model or a person working under a time limit) can verify. What distinguishes this approach is that the adversary is part of the training loop. The checker also acts alone, as a gatekeeper facing one untrusted prover, rather than as a judge who picks between two contestants.

This approach writes debate protocols in advance so that a dishonest player cannot win merely because it can compute more than an honest one. One design lets the honest side succeed using a manageable number of simulation steps, even when its opponent is allowed exponentially many. Another design addresses the tactic of splitting a question into pieces so obscure that answering them honestly would be intractable. It pairs an arguer with an estimator and keeps the honest player's cost comparable to its opponent's. The approach fixes the rules of the debate, rather than running debates and measuring how well the judge does.

This approach handles disputes over interests and values, where there is no fact to be right about. Each position is represented by its own agent, and the procedure is built to keep the positions from collapsing into one. In one design, agents role-play stakeholders and bargain under a budget with a neutral mediator, and the outcome reports how the worst-off group fared rather than only the total. In another, each value perspective gets its own distinct underlying model, which keeps the answers of a deliberating council spread out instead of converging on one. Two results underlie the design. The first is a study of what happens when a single model acts as generator, debater, and judge on moral dilemmas and contested political claims. The second is a proof that when majority preferences cycle, no reward model can pick a winner at all.

In this approach, people do the arguing and the model acts as a referee rather than a participant. It facilitates, moderates, and summarizes an online discussion, or bridges between the opposing groups within it. This reversal of roles distinguishes the approach. In every other approach in this regime, the model is one of the parties to the dispute, while here it carries out the procedure. The expectation is that a mass argument can be kept understandable and workable at a scale no human moderation team could reach.