FBS Debate offers a simpler protocol for scalable oversight
The new protocol gives stronger guarantees than the prior best approach for the same class of problems.
AGI & Safety#05
50 of 99
Example: papers supporting one claim
Image: AI Alignment Forum
50 of 99Example: papers supporting one claim
The short version
- FBS Debate is a new debate protocol for scalable oversight, designed for the same class of problems as the current best protocol.
- Its correctness holds in a worst-case sense rather than an average-case sense, and for a stronger equilibrium concept.
- The authors also prove that no protocol for this class can outperform FBS Debate while treating human judgment as a black box.
- In FBS Debate, honest play is the best strategy for each debater, no matter what the other does.
Why it matters
Debate is one proposed way to supervise AI systems on tasks too complex for humans to judge directly, and it is now being applied to frontier-scale models. The paper shows that fractional block sensitivity sets how much computation a recursive debate protocol needs to give an accurate training signal. For developers, the work points to a simpler protocol that is more practical to run.
“FBS Debate applies on a worst-case basis: prover-estimator debate only works on average over a distribution of questions, while FBS Debate guarantees that on each individual question, honesty is the best policy.”