Workshop on Reinforcement Learning Beyond Rewards: Towards Scalable General-Purpose Agents

Reinforcement Learning (RL) has traditionally focused on maximizing rewards. However, intelligent agents often rely on reward-free interactions and diverse environmental signals to form abstractions that facilitate rapid adaptation. Recent RL research has begun leveraging reward-free transitions—available through exploratory interactions or expert datasets—to increase decision-making efficiency and task specification. However, unlike in vision or language modeling, RL still lacks scalable methods for learning generalizable representations from unlabeled data. Additionally, difficulties in specifying reward functions have led researchers toward alternative signals, such as human demonstrations, preferences, and implicit feedback. This workshop seeks to advance beyond traditional reward-centric RL by exploring methods like intrinsic motivation, skill discovery, predictive and contrastive representation learning, and leveraging human-centric signals. Building upon recent progress, including foundational models employing scalable alternative signals, the workshop aims to bridge theoretical insights and practical applications, fostering collaborations toward creating more versatile, adaptive decision-making agents.