Scaling Offline RL via Efficient and Expressive Shortcut Models

TL;DR

This paper introduces SORL, a scalable offline reinforcement learning algorithm that uses novel shortcut models to efficiently train and infer policies capable of modeling complex behaviors, demonstrating strong performance and positive scaling.

Contribution

The paper proposes a new class of generative models called shortcut models for offline RL, enabling simple, efficient training and scalable inference with improved performance.

Findings

SORL achieves strong results across various offline RL tasks.

SORL exhibits positive scaling with increased test-time compute.

The method simplifies training to a one-stage process.

Abstract

Diffusion and flow models have emerged as powerful generative approaches capable of modeling diverse and multimodal behavior. However, applying these models to offline reinforcement learning (RL) remains challenging due to the iterative nature of their noise sampling processes, making policy optimization difficult. In this paper, we introduce Scalable Offline Reinforcement Learning (SORL), a new offline RL algorithm that leverages shortcut models - a novel class of generative models - to scale both training and inference. SORL's policy can capture complex data distributions and can be trained simply and efficiently in a one-stage training procedure. At test time, SORL introduces both sequential and parallel inference scaling by using the learned Q-function as a verifier. We demonstrate that SORL achieves strong performance across a range of offline RL tasks and exhibits positive scaling behavior with increased test-time compute. We release the code at nico-espinosadice.github.io/projects/sorl.