Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision
This addresses the challenge of scalable AI alignment for systems that exceed human supervision, offering a method to advance beyond human capabilities in reasoning tasks.
The paper tackles the problem of AI alignment when systems surpass human capabilities by proposing easy-to-hard generalization, where reward models trained on easy tasks (e.g., level 1-3 MATH problems) are used to evaluate and improve performance on hard tasks (e.g., level 4-5), achieving accuracies of 34.0% and 52.5% on MATH500 with models of different sizes.
Current AI alignment methodologies rely on human-provided demonstrations or judgments, and the learned capabilities of AI systems would be upper-bounded by human capabilities as a result. This raises a challenging research question: How can we keep improving the systems when their capabilities have surpassed the levels of humans? This paper answers this question in the context of tackling hard reasoning tasks (e.g., level 4-5 MATH problems) via learning from human annotations on easier tasks (e.g., level 1-3 MATH problems), which we term as easy-to-hard generalization. Our key insight is that an evaluator (reward model) trained on supervisions for easier tasks can be effectively used for scoring candidate solutions of harder tasks and hence facilitating easy-to-hard generalization over different levels of tasks. Based on this insight, we propose a novel approach to scalable alignment, which firstly trains the (process-supervised) reward models on easy problems (e.g., level 1-3), and then uses them to evaluate the performance of policy models on hard problems. We show that such easy-to-hard generalization from evaluators can enable easy-to-hard generalizations in generators either through re-ranking or reinforcement learning (RL). Notably, our process-supervised 7b RL model and 34b model (reranking@1024) achieves an accuracy of 34.0% and 52.5% on MATH500, respectively, despite only using human supervision on easy problems. Our approach suggests a promising path toward AI systems that advance beyond the frontier of human supervision.