Download
ToRA is an open-source framework developed by Microsoft for building tool-integrated reasoning agents powered by large language models. The project focuses on improving the ability of AI systems to solve complex mathematical and analytical problems by combining natural language reasoning with external computational tools. Instead of relying solely on text generation, the system dynamically invokes tools such as symbolic solvers or programming libraries when deeper computation is required. This approach allows the model to reason step by step in natural language and then execute precise calculations or code through tool calls, creating a hybrid reasoning workflow. The framework was designed to address known weaknesses of large language models in mathematical problem solving and formal reasoning tasks. Training data includes tool-use trajectories that teach the model when to reason verbally and when to delegate tasks to specialized tools.
Features
- Tool-integrated reasoning with external computational libraries
- Automatic generation of programs for symbolic or numerical computation
- Hybrid reasoning combining natural language chains with executable code
- Support for multiple model sizes and architectures such as LLaMA-based agents
- Training pipeline based on tool-use trajectories for reasoning supervision
- Evaluation on advanced mathematics datasets such as MATH and GSM8K
Project Samples
