START - A self-learning reasoning model launched by Alibaba in collaboration with the University of Science and Technology of China

What is START?

START (Self-Taught Reasoner with Tools) is a novel tool-enhanced reasoning model developed by Alibaba Group and the University of Science and Technology of China. It combines external tools (such as Python code executors) to improve the reasoning capabilities of large language models (LLMs). START uses "Hint-infer" technology to insert hints during the reasoning process, prompting the model to use external tools, and employs a "Hint-RFT" framework for self-learning and fine-tuning. Building upon Long CoT (Long Chain Reasoning), START introduces tool calls, significantly improving accuracy and efficiency in complex mathematical problems, scientific question answering, and programming challenges. START has performed exceptionally well in multiple benchmark tests, surpassing existing models and becoming the first open-source model combining long chain reasoning with tool integration.

The main function of START

• Complex computation and verification: Call the Python code executor to perform complex mathematical calculations, logical verifications, and simulations.
• Self-tuning and optimizationSTART uses tools to execute code and verify output, automatically detects errors and debugs them, improving the accuracy of answers.
• Multi-strategy explorationHints guide the model to try multiple reasoning paths and methods, enhancing the model's flexibility and adaptability when facing complex problems.
• Improve reasoning efficiencyBased on tool calls and self-verification, it reduces the hallucination phenomenon of models in complex tasks and improves inference efficiency and reliability.

START's technical principles

• Long chain reasoningInheriting the advantages of long-chain reasoning, it decomposes the problem into multiple intermediate reasoning steps, simulates the deep thinking process of humans, and improves the model's reasoning ability in complex tasks.
• Tool IntegrationSTART overcomes the shortcomings of traditional long-chain inference by calling external tools (such as Python code executors). The model generates code during inference, and the tool is used to execute and verify the results.
• Hint-inferInsert manually designed hints during the inference process to encourage the model to call external tools. Guide the model to call tools at specific nodes without requiring additional demonstration data.
• Hint-RFTBy combining Hint-infer and Rejection Sample Fine-tuning (RFT), the inference trajectories generated by the model are scored, filtered, and modified to further optimize the model's tool usage capabilities.
• Self-learning frameworkBased on active learning methods, valuable data is selected from the inference trajectories generated by the model for fine-tuning, allowing the model to learn how to use tools more effectively.
• Expansion during testingInserting hints at the end of the reasoning process increases the model's thinking time and the number of tool calls, thereby improving the accuracy and success rate of the reasoning.

START project address

• arXiv technical paper:https://arxiv.org/pdf/2503.04625

Application scenarios of START

• Solving mathematical problemsSolve complex math problems, such as math competitions and advanced math problems, and improve accuracy by verifying with code.
• Scientific research assistanceIt helps solve complex computational and scientific problems in fields such as physics, chemistry, and biology.
• Programming and DebuggingGenerates code and automatically debugs it, solving programming problems and improving development efficiency.
• Interdisciplinary problem solvingIt integrates knowledge from multiple disciplines to solve complex tasks such as engineering design and data analysis.
• Education and LearningAs an intelligent tutoring tool, it assists students in learning mathematics and science by providing detailed problem-solving processes and feedback.