Open Brain
arXivPaperNeeds Review

FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking

FERN ranks database tables for text-to-SQL without generative model calls at retrieval time. Its results show strong gains on some benchmarks, but weaker transfer on others makes dataset-specific validation essential.

arXiv · Nov 10, 2026
Open Source Open MarkdownOpen JSON
Source Summary

FERN combines three frozen retrievers, describes up to **120 features** for each of 100 candidates, and ranks them with LightGBM fitted on BIRD training questions. On Spider, complete@5 rose from **83.9% to 95.6%**; median latency on BIRD dev was **208 ms** per question on one A100, without hosted model calls.

Practical Implication

Builders retrieving schema context for text-to-SQL can test a learned reranker between retrieval and generation. It may improve which tables reach the model while keeping retrieval independent of hosted LLM calls.

Agent-Ready Context
FERN combines three frozen retrievers, describes up to **120 features** for each of 100 candidates, and ranks them with LightGBM fitted on BIRD training questions. On Spider, complete@5 rose from **83.9% to 95.6%**; median latency on BIRD dev was **208 ms** per question on one A100, without hosted model calls.

Builders retrieving schema context for text-to-SQL can test a learned reranker between retrieval and generation. It may improve which tables reach the model while keeping retrieval independent of hosted LLM calls.

Transfer was uneven: on Fiben, recall at fifteen tables was 54.2% versus RASL’s 69.2%, using different question sets. FERN’s BEAVER complete@15 also trailed two BEAVER-trained retrievers, so results from one catalog should not be assumed to generalize.
Connected Context · Feed7 Judgment

FERN shows that table retrieval can benefit from a learned ranking stage over frozen retrievers, but its uneven transfer limits any general claim that this ranking will work across catalogs. The practical implication is to evaluate retrieval and ranking on the target schema and question distribution.

Context Map
contextcodingdata#retrieval#context-engineering
Uncertainty
Transfer was uneven: on Fiben, recall at fifteen tables was 54.2% versus RASL’s 69.2%, using different question sets. FERN’s BEAVER complete@15 also trailed two BEAVER-trained retrievers, so results from one catalog should not be assumed to generalize.