LoopTool - An automated data evolution framework launched by Shanghai Jiao Tong University and Xiaohongshu

What is LoopTool?

LoopTool, developed by Shanghai Jiao Tong University and the Xiaohongshu team, is an automated, model-aware, iterative data evolution framework that improves the performance of Large Language Models (LLMs) on tool-calling tasks. The framework uses closed-loop optimization to tightly integrate data generation, label correction, and model training, forming a dynamic feedback mechanism. LoopTool comprises two main stages: seed data construction and iterative optimization. The latter includes core modules such as greedy capability detection, discriminative guided label verification, and error-driven data expansion, enabling dynamic adjustment of training data and precise optimization of weak points in the model. Experiments show that LoopTool significantly improves model performance on tool-calling tasks, achieving the best results among open-source models on multiple public leaderboards.

LoopTool's main functions

• Automated data generation: Construct high-quality seed datasets to support multi-agent dialogue generation and ensure data diversity and consistency.
• Dynamic data optimizationBased on the model's performance, it automatically identifies and optimizes the model's weaknesses, generating more challenging training samples.
• Tag verification and correctionBy comparing predictions with labels using open-source models, incorrect labels can be corrected, reducing the impact of noisy data on training.
• Model performance improvementSignificantly improves the model's tool invocation capabilities across multiple benchmark tests, while also enhancing the model's generalization reasoning ability.

The technical principle of LoopTool

• Automated tools enhance data building (Seed Generation)API definitions that conform to functional intent and structural specifications are synthesized using semantic trees and constraint trees. A high-quality seed dataset is constructed through a multi-agent dialogue generation process (including Planner Agent, User Agent, Assistant Agent, and Tool Agent).
• Model training and data evolution based on closed-loop iteration:
  • GRPO reinforcement learning training: The ability to call tools to optimize models through binary reward functions.
  • Greedy Ability Detection (GCP)Identify samples that the model has mastered, failed, and boundary samples, and retain high-perplexity samples for the next round of training.
  • Judgment of Guide Tag Verification (JGLV): Use an open-source model to compare the predictions with the original labels and correct incorrect labels.
  • Error-Driven Data Extension (EDDE): Generate new samples with similar structures but diverse contexts based on erroneous samples, thereby enhancing the model's ability to learn from difficult samples.
• Closed-loop iterative updateEach round of training data consists of high-perplexity samples, corrected error samples, newly generated samples, and unused sub-samples, forming a complete closed loop of "training-evaluation-correction-expansion".

LoopTool's project address

• GitHub repository: https://github.com/Rednote-DeepExperience/LoopTool
• HuggingFace model libraryhttps://huggingface.co/papers/2511.09148
• arXiv technical paper: https://arxiv.org/pdf/2511.09148

Application scenarios of LoopTool

• API callsLoopTool is suitable for scenarios such as intelligent customer service and automated tasks, helping models to efficiently complete queries and data interactions.
• Multi-round task planningLoopTool optimizes the model's task planning capabilities in multi-turn dialogues, enabling it to adapt to the step-by-step execution of complex tasks, such as the multi-step task processing of intelligent assistants.
• Knowledge RetrievalTo improve the accuracy and efficiency of information retrieval in question-answering systems and help models better understand user needs.
• Code generation and executionImprove the accuracy of model-generated code and execution tools, applicable to programming assistance and education platforms, and reduce code errors.
• Multimodal tasks: Optimize the model's ability to call multimodal tools and improve the performance of multimodal data processing in scenarios such as intelligent security and image recognition.