New Polaris system trains LLMs for better table retrieval using preference data

Researchers have developed Polaris, a novel system designed to enhance table retrieval for Natural Language to SQL (NL2SQL) tasks. Polaris trains a large language model to generate table descriptions that are optimized for retrieval effectiveness, rather than just fluency. The system leverages existing table retrieval benchmarks by using query-table relevance judgments to create preference pairs for fine-tuning the LLM via Direct Preference Optimization (DPO). Experiments indicate that Polaris significantly outperforms current state-of-the-art methods like AutoDDG, demonstrating the potential of repurposing retrieval benchmarks for training LLMs to produce retrieval-oriented metadata. AI

IMPACT This research could lead to more effective AI systems for querying structured data, improving how users interact with databases.

RANK_REASON The cluster describes a new research paper detailing a novel system and methodology for improving LLM performance on a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• AutoDDG
• BM25
• Direct Preference Optimization
• Hugging Face
• NL2SQL
• Polaris

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →