论文概要
研究领域: ML 作者: Tosca Lechner, Vinayak Pathak, Sivan Sabato 发布时间: 2025-11-13 arXiv: 2511.10852
中文摘要
辩论被提出作为一种训练协议,以期产生更对齐的模型:人类评判两个竞争证明相反标签的模型之间的辩论。我们通过将其形式化为在线学习问题,迈出了理解该协议内隐学习问题的第一步。我们证明:在可通过评判者学习的情况下,最佳学习方式可能与对抗性辩论不同,由此定义了协议的合作版本,并用组合维度刻画其可学习性。我们还表明,评判中的少量错误可能导致学习中的巨大误差。
原文摘要
Debate has been proposed as a training protocol to produce more aligned models. The human judges a debate between two models competing to prove opposite labels. We take the first steps towards understanding the learning problem implicit in the protocol by formulating it as online learning. We show the best way to learn given access to a judge can differ from running adversarial debate, motivating a cooperative version whose learnability we characterize using a combinatorial dimension. We also show that a small number of errors in judgement can lead to a large error in learning.
— 自动采集于 2026-11-15
#论文 #arXiv #ML #小凯
