arXiv:2511.01008v1 Announce Type: new 
Abstract: Translating natural language to SQL remains difficult for complex queries. Such queries often need environmental interaction and self-correction. To address this, we introduce MARS-SQL, a novel multi-agent framework that combines principled task decomposition and interactive reinforcement learning (RL). Our system comprises three specialized agents: a Grounding Agent for schema linking, a Generation Agent for query generation, and a Validation Agent for final selection. The core of our framework is the Generation agent, which is trained via a multi-turn RL policy. Adopting a ReAct-style Think-Act-Observe loop, the agent iteratively generates thoughts, executes SQL actions against a live database, and revises its strategy based on execution feedback, enabling dynamic, stateful reasoning and self-correction. At inference time, we generate multiple interaction trajectories to explore diverse reasoning paths. The Validation agent, then selects the optimal trajectory by modeling verification as a next-token prediction task and choosing the solution with the highest generation probability. This structured workflow pipelines specialized agents. It combines interactive RL for generation with generative modeling for verification. The approach proves highly effective for robust and accurate SQL generation. Experiments show that MARS-SQL achieves state-of-the-art Execution Accuracy of 77.84% on the BIRD dev set and 89.75% on the Spider test set. Our code is available at https://github.com/YangHaolin0526/MARS-SQL.

MARS-SQL هو إطار مبتكر للتعلم المعزز متعدد الوكلاء مصمم لمواجهة تحديات ترجمة اللغة الطبيعية إلى SQL، خاصةً للاستعلامات المعقدة. من خلال استخدام وكلاء متخصصين لربط المخططات، وتوليد الاستعلامات، والتحقق، يحسن هذا النظام دقة وكفاءة إنشاء استعلامات SQL. يعتبر هذا التطور مهمًا لأنه لا يبسط العملية للمستخدمين فحسب، بل يفتح أيضًا آفاقًا جديدة لدمج معالجة اللغة الطبيعية مع إدارة قواعد البيانات، مما يجعل الوصول إلى البيانات أكثر سهولة.

MARS-SQL es un innovador marco de aprendizaje por refuerzo multiagente diseñado para abordar los desafíos de traducir el lenguaje natural a SQL, especialmente para consultas complejas. Al utilizar agentes especializados para el enlace de esquemas, la generación de consultas y la validación, este sistema mejora la precisión y eficiencia en la creación de consultas SQL. Este desarrollo es significativo ya que no solo simplifica el proceso para los usuarios, sino que también abre nuevas posibilidades para integrar el procesamiento del lenguaje natural con la gestión de bases de datos, haciendo que el acceso a los datos sea más intuitivo.

MARS-SQL est un cadre innovant d'apprentissage par renforcement multi-agent conçu pour relever les défis de la traduction du langage naturel en SQL, en particulier pour les requêtes complexes. En utilisant des agents spécialisés pour le lien de schéma, la génération de requêtes et la validation, ce système améliore l'exactitude et l'efficacité de la création de requêtes SQL. Ce développement est important car il simplifie non seulement le processus pour les utilisateurs, mais ouvre également de nouvelles possibilités d'intégration du traitement du langage naturel avec la gestion des bases de données, rendant l'accès aux données plus intuitif.

MARS-SQL is an innovative multi-agent reinforcement learning framework designed to tackle the challenges of translating natural language into SQL, especially for complex queries. By utilizing specialized agents for schema linking, query generation, and validation, this system enhances the accuracy and efficiency of SQL query creation. This development is significant as it not only simplifies the process for users but also opens up new possibilities for integrating natural language processing with database management, making data access more intuitive.

MARS-SQL: A multi-agent reinforcement learning framework for Text-to-SQL

arXiv:2411.00365v2 Announce Type: replace 
Abstract: In the paradigm of decentralized learning, a group of agents collaborate to learn a global model using a distributed dataset without a central server; nevertheless, it is severely challenged by the heterogeneity of the data distribution across the agents. For example, the data may be distributed non-independently and identically, and even be noised or poisoned. To address these data challenges, we propose ROSS, a novel robust decentralized stochastic learning algorithm based on Shapley values, in this paper. Specifically, in each round, each agent aggregates the cross-gradient information from its neighbors, i.e., the derivatives of its local model with respect to the datasets of its neighbors, to update its local model in a momentum like manner, while we innovate in weighting the derivatives according to their contributions measured by Shapley values. We perform solid theoretical analysis to reveal the linear convergence speedup of our ROSS algorithm. We also verify the efficacy of our algorithm through extensive experiments on public datasets. Our results demonstrate that, in face of the above variety of data challenges, our ROSS algorithm has significant advantages over existing state-of-the-art proposals in terms of both convergence and prediction accuracy.

تم اقتراح خوارزمية جديدة للتعلم اللامركزي تُدعى ROSS، والتي تستخدم قيم شابلي لتعزيز قوة التعلم العشوائي بين الوكلاء. تتناول هذه الطريقة التحديات التي تطرحها توزيعات البيانات غير المتجانسة، مما يسمح للوكلاء بتعلم نموذج عالمي بشكل تعاوني دون الحاجة إلى خادم مركزي. يقوم كل وكيل بتحديث نموذجه من خلال تجميع معلومات التدرج المتقاطع من الوكلاء المجاورين، مع وزنها وفقًا لمساهماتهم.

Se ha propuesto un nuevo algoritmo de aprendizaje descentralizado llamado ROSS, que utiliza valores de Shapley para mejorar la robustez del aprendizaje estocástico entre agentes. Este enfoque aborda los desafíos planteados por las distribuciones de datos heterogéneas, permitiendo que los agentes aprendan de manera colaborativa un modelo global sin un servidor central. Cada agente actualiza su modelo agregando información de gradiente cruzado de los agentes vecinos, ponderada por sus contribuciones.

Un nouvel algorithme d'apprentissage décentralisé nommé ROSS a été proposé, utilisant les valeurs de Shapley pour améliorer la robustesse de l'apprentissage stochastique parmi les agents. Cette approche répond aux défis posés par les distributions de données hétérogènes, permettant aux agents d'apprendre de manière collaborative un modèle global sans serveur central. Chaque agent met à jour son modèle en agrégeant des informations de gradient croisé provenant des agents voisins, pondérées par leurs contributions.

A new decentralized learning algorithm named ROSS has been proposed, which utilizes Shapley values to enhance the robustness of stochastic learning among agents. This approach addresses challenges posed by heterogeneous data distributions, allowing agents to collaboratively learn a global model without a central server. Each agent updates its model by aggregating cross-gradient information from neighboring agents, weighted by their contributions.

ROSS: RObust decentralized Stochastic learning based on Shapley values

One More Thing in AI – Your Shortcut to AI Mastery

MARS-SQL: A multi-agent reinforcement learning framework for Text-to-SQL

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

Chattermate

Legion AI

AI2sql

AIvilization

Ready to build your own newsroom?