ToRA (Tool-integrated Reasoning Agents) is an open-source research project from Microsoft that trains large language models to solve hard mathematical reasoning problems by calling external tools. The GitHub repository has about 1,100 stars, and the work was published at ICLR 2024.
Pure chain-of-thought reasoning often fails on competition-style maths because models make arithmetic slips and cannot verify their work. ToRA addresses this by teaching models to interleave natural-language reasoning with tool use — writing and executing code, performing exact computation and checking intermediate results — inside a single solution trajectory. The project releases the ToRA model series, the ToRA-Corpus training data built from GPT-4-generated tool-integrated solutions, and the code used for training and evaluation. The resulting models showed large gains on established mathematical benchmarks such as GSM8K, MATH and SVAMP compared with purely textual reasoning at the same scale. The repository serves as a complete, reproducible reference for tool-integrated reasoning research.
ToRA’s approach — interleaving reasoning text with executable code — influenced later work on tool-using reasoning models, and its corpus construction method, distilling tool-integrated trajectories from a stronger teacher model, has been widely reused. Evaluations in the repo cover ten mathematical datasets.
ToRA is free and open source under the MIT license. Models, data and code are available on GitHub and model hubs at no cost; you supply your own compute for training or inference.
LLM researchers and engineers working on mathematical reasoning, tool use and agent training who want a proven open recipe — data, models and code — for tool-integrated problem solving.