Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs
Zhiyu Pan , Yizheng Wu , Jiashen Hua , Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, Jieping Ye, [0.8em] Huazhong University of Science, Technology, Alibaba Cloud · 2026
Read Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs by Zhiyu Pan , Yizheng Wu , Jiashen Hua , Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, Jieping Ye, [0.8em] Huazhong University of Science, Technology, Alibaba Cloud (2026) 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-2603-02556.txt.
Description
Zhiyu Pan, Yizheng Wu, Jiashen Hua, and colleagues improve visual reasoning through contrast in this 2026 paper. Language models sharpen reasoning by refining their own reasoning paths. Vision-language models cannot copy that trick directly, since visual hallucinations poison self-made paths. No reliable check exists for a wrong visual claim inside a rationale.
Their insight is that paired, similar images force sharper visual attention. Given two near-identical images with like questions, models pinpoint relevant cues more precisely. Their framework builds such contrastive pairs and generates rationales under that discipline. The resulting data set holds tens of thousands of visual reasoning examples. Fine-tuning various models on it beat prior self-improvement methods and top curated data sets. A model's own contrastive skill bootstraps its reasoning.
Catalog record
| CREATOR | Zhiyu Pan , Yizheng Wu , Jiashen Hua , Junyi Feng, Shaotian Yan, Bing Deng, Zhiguo Cao, Jieping Ye, [0.8em] Huazhong University of Science, Technology, Alibaba Cloud |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41048 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.02556