Density-Aware Reward Scaling: A Reinforcement Learning Framework for Reward-Proportional Sampling
A critic-free method for scaling rewards by local outcome density.
Authors: Abhijeet Sinha et al.
Status: Submitted to AAAI 2027
Density-Aware Reward Scaling (DARS) is a critic-free reward-processing method for reward-proportional sampling. It inversely scales rewards according to local outcome density while retaining the standard policy-gradient objective, encouraging policies to cover diverse valuable outcomes instead of over-representing dense regions.