Can Atomic Step Decomposition Enhance the Self-structured Reasoning of Multimodal Large Models?
Kun Xiang, Zhili Liu, Zihao Jiang, Yunshuang Nie, Kaixin Cai, Yiyang Yin, Runhui Huang, Haoxiang Fan, Hanhui Li, Weiran Huang, Yihan Zeng, Yu-Jie Yuan, Jianhua Han, Lanqing Hong, Hang Xu, Xiaodan Liang · 2025
Read Can Atomic Step Decomposition Enhance the Self-structured Reasoning of Multimodal Large Models? by Kun Xiang, Zhili Liu, Zihao Jiang, Yunshuang Nie, Kaixin Cai, Yiyang Yin, Runhui Huang, Haoxiang Fan, Hanhui Li, Weiran Huang, Yihan Zeng, Yu-Jie Yuan, Jianhua Han, Lanqing Hong, Hang Xu, Xiaodan Liang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-06252.txt.
Description
Kun Xiang and co-authors teach multimodal models to think in small, named steps in this 2025 paper. Models that see images and do math often ramble or overthink. The team's idea is that reasoning skill comes in levels, and a question should draw only the level it needs.
Their framework breaks each solution into minimal semantic steps and lets the model assemble its own chain from them. A data engine first manufactures clean step-by-step paths for training. Fine-tuning on those paths installs the habit, and a policy-guided loop runs it over several turns at test time. A new metric scores how well each single step gets used. On two visual math benchmarks the method lifted baseline accuracy by over ten points on average, used training data five times more efficiently than template rivals, and answered much faster.
Catalog record
| CREATOR | Kun Xiang, Zhili Liu, Zihao Jiang, Yunshuang Nie, Kaixin Cai, Yiyang Yin, Runhui Huang, Haoxiang Fan, Hanhui Li, Weiran Huang, Yihan Zeng, Yu-Jie Yuan, Jianhua Han, Lanqing Hong, Hang Xu, Xiaodan Liang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-06252