Tencent Proposes RubricRanker: Training Rerankers for Deep Research Agents

_reachsumit · x · 2026-08-05

Existing retrieval systems generally select documents based on relevance matching. However, for deep research agents, individually well-matched documents may not form a high-quality set that satisfies complex query requirements (e.g., diversity, conciseness, and authority).

To address this, Tencent introduced search-oriented rubrics that explicitly define the criteria for high-quality document sets. Based on these hierarchically structured rubrics synthesized by an LLM, the researchers trained a document reranker model named RubricRanker.

The training framework consists of two stages: rubrics-guided supervised fine-tuning (SFT) and rubric-based reinforcement learning (RL). Experiments demonstrate that RubricRanker outperforms the strongest baseline by 2.6 points on four deep research benchmarks and generalizes well to five RAG benchmarks.

Original post →

More from coding & agent

coding & agent channel →