TabRank boosts table retrieval with chain-of-thought distillation
A new framework improves table retrieval accuracy by up to 52.9%
Adarsh Singh and colleagues present TabRank, a framework that distills chain-of-thought reasoning from large reasoning models into compact rerankers for table retrieval. The approach builds a dataset of 6,728 reasoning traces on the Natural Questions Tables dataset and explores two distillation variants: explicit CoT distillation and conditioning the student model on the teacher's reasoning trace. TabRank significantly improves retrieval accuracy, with Acc@10 gains of 30.5% on HybridQA, 15.2% on SQA, 52.9% on TabFact, and 13.1% on TATQA compared to base models.
The framework excels in out-of-distribution settings and multi-table scenarios, demonstrating strong generalization. The authors release code, data, and models, enabling further research. TabRank addresses the challenge of enriching table retrieval with structured reasoning, offering a practical path to more accurate question answering over tables.
- TabRank uses 6,728 CoT reasoning traces from large reasoning models for table reranker training.
- Acc@10 improvements: 30.5% on HybridQA, 52.9% on TabFact, 13.1% on TATQA.
- Generalizes effectively out-of-distribution and to multi-table reasoning tasks.
Why It Matters
Boosts data retrieval accuracy, enabling more reliable QA on structured tables for enterprise and research.