CURV framework boosts chart AI reasoning by 20.5% with curriculum learning

Chart question answering (CQA) remains hard for multimodal large language models (MLLMs), which often fail to ground their reasoning in the actual visual data. A new approach called CURV tackles this by reformulating CQA as multi-step visual grounded reasoning. Instead of relying on external prompts or cues, CURV makes the model intrinsically coordinate logical reasoning with spatial attention concentration, dynamically focusing on relevant chart regions at each reasoning step. The team also built CCQA, a three-level curriculum dataset with synthetic generation covering diverse chart types and reasoning patterns—from basic single-operation questions to complex multi-chart compositional tasks. This curriculum is key: models learn progressively, building up robust visual reasoning skills.

In experiments, CURV delivered up to 20.50% improvements over existing baselines, and gains held on real-world benchmarks (up to 12.30%) and out-of-domain multimodal reasoning tasks (up to 10.20%). The work, authored by Xuehang Guo, Pingyue Zhang, Ruiyi Zhang, Zhenhailong Wang, Hanrui Lyu, Heng Ji, Tong Sun, Qingyun Wang, and Manling Li, is available on arXiv with code released publicly. For professionals working with document analysis, BI tools, or data-intensive applications, CURV suggests a path toward MLLMs that not only see charts but genuinely understand them—aligning visual evidence with logical inference rather than guessing from patterns.