What is AtomThink
AtomThink is a multimodal mathematical reasoning framework jointly launched by researchers from Sun Yat-sen University, the Hong Kong University of Science and Technology, Shanghai Jiao Tong University, the University of Hong Kong, and Huawei Noah's Ark Lab. The framework guides multimodal large language models (MLLMs) in complex reasoning by constructing long chain-of-thought (CoT), and includes an automatic CoT annotation engine, atomic step fine-tuning, and multiple search strategies. AtomThink significantly enhances the reasoning ability of MLLMs in solving mathematical problems by improving the quality of atomic steps, providing a new direction for developing general slow-thinking models.
Main features of AtomThink
- CoT annotation engine: Automatically generates high-quality chain-of-thought (CoT) annotations, addressing the problem of insufficient quality in visual mathematical data.
- Atomic step fine-tuning strategy: Jointly optimizes the multimodal large language model (MLLM) and the policy reward model (PRM) to achieve step-by-step reasoning.
- Search strategies: Provides four different search strategies, used in combination with PRM to accomplish complex reasoning tasks.
- Dataset construction: Proposes AtomMATH, a large-scale multimodal dataset containing long CoTs, used for training and evaluating models.
- Atomic capability evaluation: Designs an atomic capability evaluation method based on outcome supervision to evaluate the ability of MLLMs in generating each type of atomic step.
Technical principles of AtomThink
- Slow-thinking framework: The core idea of the AtomThink framework is "slow thinking," guiding MLLMs in complex reasoning by gradually constructing long CoT, rather than relying on fast direct prediction.
- Dynamic prompting strategy: Based on a dynamic prompting strategy, it drives MLLMs to iteratively construct state reasoning paths, where each path node represents a reasoning step, including the previous stage, the current state, and possible actions.
- Short CoT enhancement: Based on LLMs, it semantically segments existing short CoT annotations into multiple discrete steps, focusing on solving a single atomic problem in the reasoning process.
- Multimodal dataset: Samples mathematical data from multiple data sources, generates multi-step reasoning paths based on dynamic prompting and short CoT enhancement, and constructs the AtomMATH dataset.
- Atomic step fine-tuning: Fine-tunes on the AtomMATH dataset, enabling the MLLM to learn reasoning patterns based on atomic steps.
Project address of AtomThink
- GitHub repository：https://github.com/Quinn777/AtomThink
- arXiv technical paper：https://arxiv.org/pdf/2411.11930
Application scenarios of AtomThink
- Educational assistance: As an intelligent tutoring system, it provides students with step-by-step solutions and explanations for mathematical problems.
- Automated testing and evaluation: Automatically generates and grades mathematical test questions in online examination systems.
- Academic research: Assists researchers in exploring and solving complex mathematical problems.
- Software development: Helps developers automatically generate and debug code related to mathematical computation.
- Intelligent customer service and technical support: Provides support in customer service scenarios that require mathematical computation or reasoning.