ProRL: Effective Reinforcement Learning for Proactive Recommendation via Rectified Policy Gradient Estimation

Researchers introduced ProRL, a reinforcement-learning method for proactive recommendation that fixes two policy-gradient failure modes: length bias and high-variance gradients.

ProRL uses stepwise reward centering and position-specific advantage estimation, stabilizing training without needing a critic.

The method outperforms baselines on MovieLens-1M, Steam, and Amazon-Book, improving path quality and diversity.

It also transfers better to unseen reward signals and recommendation models, including GRU4Rec, BERT4Rec, and LightSANs.