ABPO Framework Tackles Bandit Feedback in Continual LLM Recommender Updates

To tackle exposure bias and feedback ambiguity in the ongoing updates of generative LLM-based recommenders (LLM-Rec), researchers have introduced the Anchored Bandit Policy Optimization (ABPO). The feedback from post-deployment logs is limited to policy-shaped contextual bandit signals, where results are only noted for items presented by a previous serving policy, resulting in incomplete and skewed information. ABPO merges group-relative policy optimization (GRPO) with a direct approach to these biases by integrating the exposed recommendation as a logged anchor within each GRPO rollout group, adjusting group-relative normalization based on the actions of the prior policy. The study can be found on arXiv with the identifier 2605.18899.