Debate is a proposal for scalable oversight: supervising models on tasks where people cannot check an answer directly but can judge arguments about it (Irving et al., 2018).
Concept 86 of 100Scaling & AlignmentStudy phase 12
How to study this pagePredict · Check · Generalize · Continue
PredictSay what you expect.
Pick a concrete input and say what the equation should give before you read the math.
CheckCheck against the math and the papers.
Work through the math, then read the cited papers for the original argument.
GeneralizeName what stays the same.
Identify the part of the idea that survives a change of notation or model, so you can recognize it in another paper or system.
ContinueFollow the connections.
Use the prerequisites and related concepts at the end of the page to fill the next gap.
Why It Matters for Modern Models
Debate is a proposal for scalable oversight: supervising models on tasks where people cannot check an answer directly but can judge arguments about it (Irving et al., 2018).
In an early experiment on MNIST images, debate between two agents raised the accuracy of a judge that sees only 6 pixels from 59.4% to 88.9% (Irving et al., 2018).
With language models, judges who could not read a story answered comprehension questions about it more accurately after watching two models that could read it debate: 88% for human judges and 76% for model judges, against naive baselines of 60% and 48%. Training the debaters to be more persuasive made judges more accurate, not less (Khan et al., 2024).
Questions to Think About
What stops a skilled debater from winning with a convincing but false argument that the judge cannot check, and does that risk grow as debaters get stronger?
The PSPACE result assumes optimal play; what does it say, and not say, about debates between imperfect trained models?
Core Math
Key Equation
πAmaxπBminE[J(τ)]
In debate, two copies of a model, A and B, take turns making statements about a question, and a judge, who may be a person, reads the transcript τ and decides which side argued for the true answer (Irving et al., 2018). Training is a zero-sum game:
πAmaxπBminE[J(τ)]
where J(τ)=1 if the judge sides with A and 0 otherwise.
The hope is that pointing out a flaw in an argument is easier than producing a correct answer, so a judge weaker than the debaters can still tell who is right. Irving et al. support this with an analogy to interactive proofs: with optimal play, debate with a polynomial-time judge can answer any question in the complexity class PSPACE, whereas judging answers directly reaches only NP. This is a statement about optimal debaters, not a guarantee for trained models.