EACL 2026 (Findings) paper accepted! 🎉

Excited to share that our paper, “VisDoT: Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought,” has been accepted to EACL 2026 (Findings)! 🎉

While recent large vision-language models (LVLMs) show strong language-centric reasoning capabilities, they often struggle with perceptual grounding, failing to reliably align visual primitives (e.g., position, length, color, and patterns) with semantic reasoning—especially in visualized data such as charts and graphs.

💡 Key Contributions

VisDoT addresses this limitation by combining two key ideas:

1️⃣ Human-like interpretation grounding based on graphical perception theory. We formalize four core perceptual tasks—Position, Length, Pattern, and Extract—to explicitly guide models toward accurate visual decoding before reasoning.

2️⃣ Decomposition-of-Thought (DoT) prompting, which reformulates visual question answering as a compositional process. DoT decomposes complex questions into sequential perception-oriented and logic-oriented sub-questions, enabling structured, interpretable, and grounded reasoning within a single model and a single inference pass.

🔍 Major Results

Extensive experiments demonstrate that VisDoT substantially improves performance on challenging chart reasoning benchmarks such as ChartQA, ChartQAPro, and the newly introduced VisDoTQA:

• Significant Improvement: Achieve up to +33.2% absolute improvement on VisDoTQA.
• Strong Performance: Match or surpass strong closed-source models (e.g., GPT-4o) on ChartQAPro.
• Zero-Shot Generalization: Exhibit consistent zero-shot gains on open-domain VQA benchmarks, indicating strong generalization beyond charts.

Overall, VisDoT shows that explicit perceptual grounding combined with structured thought decomposition is a powerful and general strategy for reliable and interpretable visual reasoning.