Research map / Large language models / Instruction and fine-tuning

ROUTE: Robust Multitask Tuning and Collaboration for Text-to-SQL

ICLR 2025 · Large language models › Instruction and fine-tuning · approach: Calibration and knowledge in LLM tuning · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ROUTE, multitask SFT (Text2SQL, schema linking, noise correction, continuation writing) plus a multitask collaboration prompting pipeline to boost open-source LLMs on Text2SQL.

Problem

Open-source LLMs lag closed-source GPT-4 based Text2SQL pipelines; SFT on Text2SQL alone degrades other SQL-related abilities and prompting pipelines are costly and weak on small models.

Approach

Multitask SFT on four SQL-related tasks (including new continuation writing and noise correction with noisy-correspondence filtering), then a collaborative inference pipeline that chains schema linking (merged with pseudo-SQL-derived schema), SQL generation, executor-checked noise correction, and continuation writing for hard queries.

Key results

• ROUTE + Qwen2.5-14B reaches 87.3 Dev-EX on SPIDER and 60.9 Dev-EX on BIRD (65.2 VES), versus 83.6 and 55.9 for ROUTE + Qwen2.5-7B.
• MCP alone lifts Llama3-8B from 69.3 to 75.0 on SPIDER Dev-EX and from 32.1 to 42.7 on BIRD Dev-EX.
• On SPIDER-variant benchmarks, ROUTE + Llama3-8B averages 75.1 versus 72.5 for SFT alone.

Measured effects reported in the paper

Surprising findings

• Noisy correspondence filtering improved SPIDER accuracy (84.5 to 86.0) but slightly degraded BIRD accuracy (57.4 to 57.3 in the table, with the noise-filtering ablation reporting a drop on BIRD).
• MSFT on a single task such as Text2SQL alone caused other SQL-related tasks to collapse, e.g. Text2SQL-only SFT left schema linking outputs unusable due to format overfitting.

Limitations

• Multi-stage MCP adds reasoning cost at inference.
• Large gap remains to upper-bound performance with ground-truth schema and continuation hints.
• Small 7B/14B models struggle on BIRD's large schemas and long contexts; BIRD test results not reported.

Experimental setup

Datasets: SPIDER BIRD SPIDER-SYN SPIDER-DK SPIDER-Realistic Dr.Spider

Benchmarks: SPIDER BIRD SPIDER-SYN SPIDER-DK SPIDER-Realistic Dr.Spider

Metrics: execution accuracy test-suite accuracy valid efficiency score

Baselines: GPT-4 DIN-SQL DAIL-SQL MAC-SQL MCS-SQL DTS-SQL CodeS SENSE Llama3-8B + SFT Qwen2.5-7B + SFT

Builds on: DIN-SQL MAC-SQL SENSE DTS-SQL CodeS

Closest papers by meaning

• Structure-Guided Large Language Models for Text-to-SQL Generation (ICML 2025 · similarity 0.91)
• CogSQL: A Cognitive Framework for Enhancing Large Language Models in Text-to-SQL Translation (AAAI 2025 · similarity 0.90)
• Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search (ICML 2025 · similarity 0.89)
• SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• CHASE-SQL: Multi-Path Reasoning and Preference Optimized Candidate Selection in Text-to-SQL (ICLR 2025 · similarity 0.88)
• SQLens: An End-to-End Framework for Error Detection and Correction in Text-to-SQL (NeurIPS 2025 · similarity 0.88)