Research map / Multi-agent systems and games / LLM agents

MARS-SQL: A Multi-Agent Reinforcement Learning Framework For Text-To-SQL

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 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 MARS-SQL, a trainable three-agent Text-to-SQL framework (grounding, GRPO-trained multi-turn ReAct generation, and SFT generative validation) reporting SOTA execution accuracy on BIRD-dev and Spider-test using 7B agents.

Problem

Text-to-SQL with complex schemas requires precise logic and schema alignment, but static prompting and one-shot generation lack dynamic self-correction through database interaction.

Approach

Decomposes Text-to-SQL into grounding, generation, and validation agents, trains the schema grounder and the generator with GRPO (the generator is a multi-turn ReAct policy rewarded by execution outcomes), and selects among candidate trajectories with a generative Yes/No validator scored by next-token probability.

Key results

• MARS-SQL reaches 77.84% execution accuracy on BIRD-dev, versus 72.29% for the next-best open-source system (Reasoning SQL).
• MARS-SQL reaches 89.75% on Spider-test and 78.13% on Spider-DK (second best), trained only on BIRD.
• In ablation, the generative verifier gives 77.84% on BIRD-dev versus 72.93% for self-consistency with the same generator setup.

Surprising findings

• A single Qwen2.5-Coder model SFT-trained for all three roles reaches only 55.67% (14B) and 57.82% (32B) on BIRD-dev, suggesting multi-agent role separation matters more than raw scale.
• The full pipeline's 21B total footprint means a single-model baseline with comparable budget (SFT + 16 samples, 64.2%) still underperforms, but the paper notes the approach costs more than a single 7B model.

Limitations

• Three specialized 7B agents give ~21B total inference footprint, and the deployed pipeline costs more than a single 7B model.
• Latency of ~22 s per query and ~13k tokens per query are high for interactive use.
• The reward shaping for grounding is heuristic, and the paper does not clearly separate contributions from training-data differences such as the 16-candidate validator data.

Experimental setup

Datasets: BIRD train Spider 1.0 Spider-DK SParC

Benchmarks: BIRD dev Spider test Spider-DK SParC

Metrics: execution accuracy selection accuracy latency VRAM

Baselines: O3-mini GPT-4o GPT-5 Qwen2.5-Coder-7B Qwen3-Coder-30B CHESS OpenSearch-SQL XiYan-SQL CHASE-SQL CodeS SHARE OmniSQL

Builds on: GRPO ReAct Generative Verifiers Self-Consistency

Closest papers by meaning

• CHASE-SQL: Multi-Path Reasoning and Preference Optimized Candidate Selection in Text-to-SQL (ICLR 2025 · similarity 0.89)
• SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models (ICML 2024 · similarity 0.88)
• Beyond Static Pipelines: Learning Dynamic Workflows for Text-to-SQL (ICML 2026 · similarity 0.88)
• Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.88)
• Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search (ICML 2025 · similarity 0.88)