Verified, then logged.
Sun 22 Nov 2026 · 00:17 KST / 15:17 UTC한국어 | ENSearch /

Breaking01 / AI

FERN improves table retrieval without generative AI

A monitor in a naturally lit database workspace displays numerous data tables alongside five shortlisted tables arranged in ranked order. AI-generated image

FERN researchers reported that their method retrieved every required database table for 95.6% of questions on Spider without training on that benchmark, a researcher-published result.

Their paper, published on 11 Nov, describes FERN as a method for finding database tables needed to answer natural-language questions. It combines retrievers whose model weights remain fixed, then uses LightGBM, a machine-learning model, to reorder the candidates.

According to the paper, the share of questions with every correct table among the top five candidates rose from 83.9% before reranking to 95.6% after it. Both figures are researcher-published results.

The researchers reported training the ranking model once on training questions from BIRD, another benchmark. FERN selects and reorders table candidates without a generative language model.

The paper gives Spider table recall as 96.7 for FERN and 97.0 for RASL, measuring how many required tables were retrieved among the top five candidates. These are researcher-published benchmark figures; the RASL value comes from an earlier publication.

The researchers did not disclose a RASL comparison evaluated on the same Fiben question set in the abstract.