Inference-time Generative Debates on Coding and Reasoning Tasks for Scalable Oversight

Researchers Ethan Elasky and Frank Nakasako have published a significant empirical study testing inference-time generative debates for scalable AI oversight. Their work, using weak-judge/strong-debater setups with GPT-5-mini/nano and Qwen3-8B/4B model pairs on BigCodeBench+ coding tasks and ARC-AGI reasoning problems, delivers mostly negative results. The core finding is that debate formats underperformed consultancy baselines in 11 of 16 tested conditions, challenging assumptions about debate's effectiveness for scalable oversight. Even more striking, removing debate transcripts entirely and showing judges only the proposed answers matched or outperformed both debate and consultancy approaches.

The study reveals a critical mechanism: judges tend to default to endorsing plausible-sounding arguments even when both debaters are wrong, and debate transcripts specifically amplify this problematic tendency. While best-of-4 speech selection showed some promise on ARC-AGI tasks, suggesting potential for RL-trained debaters, the overall results indicate current debate implementations may not provide the expected oversight benefits. This research represents a departure from previous multiple-choice settings by using verifiable ground truth answers, and introduces formats where participants freely choose positions rather than being assigned sides—important considerations for alignment research moving forward.