Summary of Atomthink: a Slow Thinking Framework For Multimodal Mathematical Reasoning, by Kun Xiang et al.
AtomThink: A Slow Thinking Framework for Multimodal Mathematical Reasoning
by Kun Xiang, Zhili Liu, Zihao Jiang, Yunshuang Nie, Runhui Huang, Haoxiang Fan, Hanhui Li, Weiran Huang, Yihan Zeng, Jianhua Han, Lanqing Hong, Hang Xu, Xiaodan Liang
First submitted to arxiv on: 18 Nov 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper introduces AtomThink, a novel framework that enables multimodal large language models (MLLMs) to perform complex mathematical reasoning through the incorporation of “slow thinking.” This approach involves constructing long chains of thought (CoT) consisting of atomic actions in a step-by-step manner. The framework consists of three key modules: CoT annotation engine, atomic step fine-tuning strategy, and four different search strategies that can be applied with a policy reward model (PRM). The paper also proposes AtomMATH, a large-scale multimodal dataset of long CoTs, and an atomic capability evaluation metric for mathematical tasks. Experimental results show that AtomThink significantly improves the performance of baseline MLLMs on MathVista and MathVerse, achieving approximately 50% relative accuracy gains. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps AI models think more like humans by creating a step-by-step process to solve math problems. It’s called “slow thinking” because it takes time to reason through each step. The new framework, AtomThink, has three main parts: one that creates good instructions for the model, another that fine-tunes the model to follow these steps, and four ways to search for answers using a special reward system. The team also created a big dataset of math problems and a way to measure how well the models do on these tasks. This can help AI systems get better at solving complex math problems. |
Keywords
» Artificial intelligence » Fine tuning
