Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision

TL;DR

This paper introduces a scalable alignment method where reward models trained on easy tasks are used to evaluate and improve performance on harder tasks, enabling AI systems to surpass human-level capabilities in complex reasoning.

Contribution

The paper proposes a novel easy-to-hard generalization approach using reward models trained on simple tasks to evaluate and enhance performance on difficult tasks, advancing AI beyond human supervision.

Findings

Reward models trained on easy tasks effectively evaluate harder tasks.

The approach achieves 34.0% accuracy on MATH500 with minimal supervision.

Enables AI systems to surpass human capabilities in complex reasoning.

Abstract

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.