Chart-FR1: Focus-Driven Reasoning for Dense Charts
The introduction of Chart-FR1 aims to enhance the effectiveness of multimodal large language models (MLLMs) when dealing with high information density (HID) charts. These charts, characterized by numerous subplots, legends, and intricate annotations, present three primary difficulties: insufficient fine-grained perception that leads to missing vital visual signals, excessive visual redundancy that hampers reasoning, and a deficiency in adaptive deep reasoning. To tackle these issues, Chart-FR1 employs Focus-CoT, a visual focusing chain-of-thought that connects reasoning processes to essential visual elements such as local image areas and OCR signals. This model improves perception, efficiency in focus, and adaptive deep reasoning for HID charts. The findings are detailed in arXiv paper 2605.01882v1.
Key facts
- Chart-FR1 is a focus-driven fine-grained chart reasoning model.
- It targets high information density (HID) charts with multiple subplots, legends, and dense annotations.
- Three challenges addressed: limited fine-grained perception, redundant visual information, lack of adaptive deep reasoning.
- Focus-CoT is a visual focusing chain-of-thought that links reasoning steps to key visual cues.
- Key visual cues include local image regions and OCR signals.
- The model improves perception, focusing efficiency, and adaptive deep reasoning.
- The research is published on arXiv with ID 2605.01882v1.
Entities
Institutions
- arXiv