Training a Document Reranker with Search Rubrics for Deep Research Agents

The paper argues that deep research agents need document sets that satisfy query-level requirements such as diversity, concision, and authority, rather than merely individually relevant documents. It introduces hierarchical search rubrics synthesized with a powerful LLM and trains RubricRanker to select a high-quality subset from retrieved candidates. The two-stage training framework combines rubric-guided supervised fine-tuning with rubric-based reinforcement learning. According to the abstract, RubricRanker improves over the strongest baseline by 2.6 points across four deep research benchmarks and generalizes to five RAG benchmarks.