TabRank: A Framework for Training Reasoning Rerankers in Tabular Retrieval
TabRank is a novel framework designed for training reasoning rerankers tailored for tabular retrieval. This initiative responds to the rising significance of neural and LLM-based rerankers within multi-stage retrieval systems, which enhance candidate selections from initial retrievers. Recent progress in Large Reasoning Models (LRMs) utilizing chain-of-thought (CoT) reasoning has led to better ranking outcomes in unstructured passage retrieval. TabRank applies this concept to structured data by assembling a detailed dataset comprising 6,728 reasoning traces for tabular reranking, specifically on the Natural Questions Tables dataset. The framework also investigates two approaches for developing a compact reasoning model. This research has been made available on arXiv under ID 2607.25182.
Key facts
- TabRank is a framework for training reasoning rerankers for tabular retrieval.
- A dataset of 6,728 reasoning traces was created on the Natural Questions Tables dataset.
- The work builds on LRMs with chain-of-thought reasoning for ranking.
- Two variants of training a compact reasoning model are explored.
- The research is published on arXiv (ID 2607.25182).
Entities
Institutions
- arXiv