AdaTutoRank addresses sparse-credit-assignment problems in setwise document rerankers for RAG, where prior methods reward an entire retrieved set with one shared scalar rubric score, making redundant and decisive documents indistinguishable during training. It introduces Adaptive Tutoring Optimization (ATO), which draws three increasingly specific ‘hint’ forms from a frozen snapshot of the policy (rubrics alone, a rubric-selected sibling set, and a reflective contrast against that sibling set) and matches each training rollout to the hint form suited to its quality level. Re-scoring rollouts under hint-conditioned versus hint-free snapshots distills a token-level advantage that complements the standard group-relative outcome advantage. Across ten RAG, deep-research, and setwise-evaluation benchmarks, AdaTutoRank achieves the best overall performance while issuing fewer retrieval calls than baselines.
