Charts are a core medium for conveying quantitative and qualitative information, yet small multimodal large language models still struggle to reliably ground visual primitives and connect them to downstream reasoning. This thesis proposes VisDoT, a fr...
Charts are a core medium for conveying quantitative and qualitative information, yet small multimodal large language models still struggle to reliably ground visual primitives and connect them to downstream reasoning. This thesis proposes VisDoT, a framework that transfers visual reasoning from large multimodal models to compact models by combining Decomposition-of-Thought prompting with perceptualtask supervision. The proposed method decomposes a chart question into perception-oriented and logic-oriented sub-questions and trains small models to recover visual evidence before performing higher-level reasoning. To strengthen visual grounding, the framework formalizes four perceptual tasks, namely Position, Length, Pattern, and Extract, and constructs the VisDoTQA dataset with stepwise reasoning annotations. Experimental results show that VisDoT improves InternVL-4B on ChartQA from 75.08% to 84.36%, and to 85.60% under the short-answer setting. On the VisDoTQA benchmark, the model improves from 43.30% to 76.52%, outperforming GPT-4o and Gemini-Flash 2.0 on this benchmark. Additional comparisons with public instruction and distillation pipelines show that VisDoT remains highly data efficient, achieving competitive performance with only 7.4K question-answer pairs in the controlled comparison setting. Zero-shot gains on POPE and MMMU further suggest that perception-first decomposition can extend beyond chart-specific tasks. These findings indicate that human-like interpretation grounding and decomposition-based supervision provide an effective and interpretable path for robust chart understanding in small multimodal large language models.