LoopTool - Shanghai Jiaotong University and Little Red Book open source automated tool to call the data evolution framework

What is LoopTool?

LoopTool is an automated tool-call data evolution framework open-sourced by Shanghai Jiao Tong University and Xiaohongshu team, specifically designed to improve the tool-call capability of large language models. It optimizes data generation and model training through closed-loop iteration, using open-source models (e.g., Qwen3-32B) as data generators and discriminators without relying on closed-source APIs.The core modules include multi-intelligent body dialogue generation, capability detection, label checking, and error-driven data expansion, which dynamically identifies model weaknesses and generates high-quality training data. Experiments show that the LoopTool-based 8B model outperforms the 32B generator model in the tool-call benchmark test, and achieves the best results of the same size on the BFCL-v3 and ACEBench lists.

LoopTool Features

• Competency Diagnosis and Identification of Shortcomings: Through the Greedy Capability Probing (GCP) module, LoopTool is able to diagnose the model's capability on tool-calling tasks, accurately identify the points of capability that the model has and has not mastered, determine the model's shortcomings and error patterns, and provide direction for subsequent data optimization.
• Label Correction and Quality Improvement: With the Judgement-Guided Label Verification (JGLV) module, LoopTool utilizes the open source model as a judge, comparing the accuracy of the model predictions with the original labels, automatically correcting mislabels, improving the overall quality of the dataset, and ensuring the accuracy of the training data.
• Data Extension and Challenging Sample Generation: Through the Error-Driven Data Expansion (EDDE) module, LoopTool generates new challenging samples based on the model's error cases, increasing the number and diversity of difficult samples in the dataset, and prompting the model to continually improve its own tool-calling capabilities in the face of complex problems.
• Closed-loop optimization and dynamic evolutionLoopTool builds a closed loop between data and model, realizing the close integration of data generation, label correction and model training. As the model capability improves, the dataset will also evolve dynamically, forming a positive cycle of "data trains the model, model guides the data", which continuously improves the performance of the model.
• Open Source Ecology and Cost Control: The framework relies entirely on open-source models to complete data generation and evaluation without the need for closed-source APIs, which reduces costs and ensures the flexibility and scalability of data generation and evaluation, making the entire optimization process more efficient and sustainable.

LoopTool's Core Benefits

• Closed-loop optimization and dynamic evolution: By building a closed loop between data and model, it realizes the close integration of data generation, label correction and model training, ensuring that the training data dynamically evolves with the model capability, forming a positive cycle.
• Significant Performance Improvements: On the tool-calling task, LoopTool-trained models significantly outperform other models of the same size, and even outperform larger data generators in some benchmarks, demonstrating superior optimization.
• Enhanced generic capacity: In addition to the enhancement of tool invocation capabilities, LoopTool also enhances the model's performance on other general-purpose tasks such as instruction following and code generation, improving the model's ability to generalize and handle complex tasks.
• Efficient data utilization: Through error-driven data expansion and label correction, LoopTool is able to efficiently utilize data to generate high-quality and challenging samples, improving the model's ability to handle complex problems.
• Wide range of applicability: LoopTool is applicable to models of different sizes and shows good optimization results on both 8B and 32B models, with wide applicability and scalability.

What is LoopTool's official website

• GitHub repository:: https://github.com/Rednote-DeepExperience/LoopTool
• HuggingFace Model Library:: https://huggingface.co/papers/2511.09148
• arXiv Technical Paper:: https://arxiv.org/pdf/2511.09148

Who LoopTool is for

• Large Language Model Developer: Developers who are committed to improving their language modeling tool invocation capabilities can use LoopTool to optimize model performance and enhance the performance of their models in real-world applications.
• artificial intelligence researcher: AI researchers interested in model optimization and data evolution can use LoopTool to explore new model training methods and data enhancement strategies.
• Open Source Community Contributors: Researchers and developers who wish to contribute to and benefit from the open source field, LoopTool relies entirely on the open source model and is suitable for participants in the open source community.
• Corporate Technical Team: Enterprise technology teams that need to improve model efficiency and reduce costs, LoopTool can help them achieve model optimization with limited resources.
• academic research organization: Academic institutions focusing on language modeling and artificial intelligence research can use LoopTool as a research tool to advance academic progress in related fields.