Dense Reward RL
RL: Reinforcement Learning
Overview
Dense reward reinforcement learning studies how to give agents useful feedback during a task, rather than only scoring the final outcome, to improve learning and credit assignment. Approaches include progress and stage-based rewards, models learned from demonstrations or videos, language- and vision-based signals, and trajectory, graph, or token-level feedback. These methods are used in robotic manipulation, long-horizon language-model agents, coding, and mathematical reasoning, often alongside sparse outcome rewards. A central challenge is making intermediate signals accurate and generalizable without encouraging reward hacking or destabilizing training; in some settings, noisy or poorly delivered dense rewards can perform worse than sparse feedback.
Latest papers
34 papers5 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.