Research map / Vision-language and multimodal / Multimodal LLMs

Chart-FR1: Visual Focus-Driven Fine-Grained Reasoning on Dense Charts

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Chart-FR1, a Qwen2.5-VL-7B model trained with focus-tagged chain-of-thought (Focus-CoT) SFT and a Focus-GRPO RL variant with an information-efficiency reward and adaptive KL, plus the HID-Chart benchmark for dense charts.

Problem

MLLMs struggle with high information density charts (multiple subplots, legends, dense annotations) due to weak fine-grained perception, redundant visual information, and lack of adaptive deep reasoning.

Approach

Interleave reasoning with explicit <focus> actions (OCR text and bounding-box regions) as Focus-CoT, then optimize with GRPO variants that penalize redundant focus cues, relax numerical accuracy matching, and scale the KL penalty by the number of visual cues.

Key results

• Chart-FR1-7B averages 67.1 across five chart benchmarks vs 61.0 for Qwen2.5-VL-7B and 64.0 for GPT-4o.
• On HID-Chart, Chart-FR1-7B reaches 53.0 average accuracy, outperforming Qwen2.5-VL-72B (51.5) and GPT-4o (51.2).
• Focus-GRPO improves average accuracy to 67.1 vs 64.1 for standard GRPO; removing adaptive KL drops it to 65.8.

Surprising findings

• Replacing the GPT-5 teacher with open-source Qwen2.5-VL-72B or Qwen3-VL-32B yields only modest drops in average accuracy (65.5 and 66.7 vs 67.1).

Limitations

• Relies on GPT-5 for Focus-CoT reconstruction and for benchmark information-density scoring and question generation.
• HID-Chart judged by GPT-5 mini, with LLM-based answer matching that may introduce grading noise.
• Redundancy penalty uses string similarity and IoU heuristics that may not capture semantic redundancy.

Experimental setup

Datasets: NovaChart EvoChart ChartQA

Benchmarks: ChartQA CharXiv EvoChart ChartBench PlotQA HID-Chart

Metrics: accuracy

Baselines: GPT-4o Gemini-1.5-Pro Qwen2.5-VL-7B Qwen2.5-VL-72B InternVL2.5-78B ChartSketcher-72B EvoChart ChartReasoner ChartPoint-7B ChartGemma R1-VL-7B R1-Onevision-7B

Builds on: GRPO Chain-of-Thought Vision-R1 Qwen2.5-VL

Closest papers by meaning

• ChartSketcher: Reasoning with Multimodal Feedback and Reflection for Chart Understanding (NeurIPS 2025 · similarity 0.91)
• ChartMuseum: Testing Visual Reasoning Capabilities of Large Vision-Language Models (NeurIPS 2025 · similarity 0.91)
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding (CVPR 2026 · similarity 0.91)
• SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More (CVPR 2026 · similarity 0.89)
• Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models (ICML 2026 · similarity 0.89)
• Visual Self-Refine: A Pixel-Guided Paradigm for Accurate Chart Parsing (ICLR 2026 · similarity 0.89)