Document Optimization for Black-Box Retrieval via Reinforcement Learning
This research frames document optimization as a reinforcement learning problem where an instruction-tuned model rewrites documents to improve retrieval quality under a target search system. By using Group Relative Policy Optimization (GRPO) and rewards based on counterfactual changes in retrieval metrics like nDCG, the method learns to transform documents into representations better aligned with the retriever's internal space, showing gains across different retrieval types.