Users are excited about AI debate techniques for boosting oversight and benchmark accuracy because they value scientific understanding of where oversight works and fails and invite others to join the research.
Based on 3 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
If you're excited about this kind of work and agree on the need for a scientific understanding of where oversight works and fails, come work with me at Resolution! We're hiring for our scalable oversight team. https://jobs.ashbyhq.com/resolution/65704205-7ccd-4df6-a7d1-3e3338bd80a8 [10/n]
Debate work usually emphasizes the 2P game structure: use adversarial pressure to surface flaws. We instead see the gains come from better pooling across many noisy critiques. Data shows similar gains from using M critiques by averaging vs adversarially min-reducing. [6/n] https://x.com/jacob_pfau/status/2075552396785152247/photo/1
We build the full game tree per question (50 solutions × 50 critiques × 20 rebuttals, judge-scored at every leaf), then independently dial: proposer vs critic optimization, debate depth, and sample budget. Then we read off exactly which knob buys accuracy. [3/n] https://x.com/jacob_pfau/status/2075552390208516536/photo/1
Headline: debate gives up to +43pp accuracy (38→81% LiveCodeBench, 50→91% ARC-AGI). Where from? Roughly half is consensus (simple majority gives: 50→71% on ARC), half is debate structure (91% vs 71% at matched samples). [5/n] https://x.com/jacob_pfau/status/2075552394318864772/photo/1
That's surprising since recent work found simple majority voting matches all fancy consensus methods. Meanwhile debate-like aggregation outperforms by ~20pp. There may be a continuum between consistency/aggregation methods and debate-like methods! https://arxiv.org/abs/2603.06612 [7/n]
That's surprising since recent work found simple majority voting matches all fancy consensus methods. Meanwhile debate-like aggregation outperforms by ~20pp. There may be a continuum between consistency/aggregation methods and debate-like methods! https://arxiv.org/abs/2603.06612 [7/n]
The difference looks to be coming from the additional rounds of interaction in debate. We find (1) adding critic is very different from direct judge optimization (2) adding a rebuttal after critic adds again. [8/n] https://x.com/jacob_pfau/status/2075552401910591941/photo/1
Prior BoN debate work optimized greedily, one player at a time. We evaluate the full nested min-max tree: the critic best-responds to each proposal, the proposer best-responds to the worst-case critic. Giving us an inference-time sim of zero-sum self-play RL training. [4/n]
There's a lot more to do with min-max BoN as a microscope for diagnosing oversight protocols. In the future we'll look at per-question early stopping, protocol iteration, and the fuzzy, non-verifiable tasks alignment actually needs. [9/n]
Full post: https://www.lesswrong.com/posts/hb8pv3zyAHGJpwz9F/debate-with-self-play-best-of-n-optimization See also Sam's thread! https://x.com/SamMartin589196/status/2075268508716556686?s=20 [11/11]
Users are excited about AI debate techniques for boosting oversight and benchmark accuracy because they value scientific understanding of where oversight works and fails and invite others to join the research.
Based on 3 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
The difference looks to be coming from the additional rounds of interaction in debate. We find (1) adding critic is very different from direct judge optimization (2) adding a rebuttal after critic adds again. [8/n] https://x.com/jacob_pfau/status/2075552401910591941/photo/1