Research map / Vision-language and multimodal / Multimodal LLMs

EvoChart: A Benchmark and a Self-Training Approach Towards Real-World Chart Understanding

AAAI 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 2 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 EvoChart, a multi-stage self-training pipeline that synthesizes chart data and trains a Phi-3-Vision-based chart model, plus EvoChart-QA, a 625-chart real-world benchmark for basic chart comprehension.

Problem

VLMs score highly on ChartQA but perform poorly on real-world chart understanding, partly due to a scarcity of high-quality chart training data and an overly optimistic benchmark with few chart sources.

Approach

A cyclical self-training loop in which the current-stage model evaluates synthetic charts with directness-based QA probes, drops or modifies charts based on score change, and then trains on self-selected QA pairs to produce the next-stage model and data.

Key results

• EvoChart-4B reaches 54.2% overall accuracy on EvoChart-QA, above GPT-4o at 49.8%.
• GPT-4o scores 85.7% on ChartQA but only 49.8% on EvoChart-QA; open-source chart expert models fall below 30% on EvoChart-QA.
• On ChartQA, EvoChart-4B scores 81.5%, comparable to Intern-VL-2.0-8B (81.5%) and Phi3-Vision (81.4%).

Measured effects reported in the paper

Surprising findings

• Chart expert models fine-tuned on ChartQA-style data drop below 30% on EvoChart-QA despite strong ChartQA scores.
• Ablation shows the model trained without chart refinement at stage 3 scores lower (49.0%) than with refinement (54.2%) despite more training data.

Limitations

• Benchmark questions are limited to basic chart understanding (direct and complex retrieval), not numerical reasoning.
• Pipeline relies on GPT-4 for seed generation and on ECharts rendering, so synthetic charts may not fully match real charts.
• Complex retrieval performance of EvoChart trails larger models such as InternVL-2.0-40B and GPT-4o.

Experimental setup

Datasets: ChartQA EvoChart corpus (synthetic, ECharts-rendered)

Benchmarks: EvoChart-QA ChartQA

Metrics: accuracy (strict zero-tolerance / flex 5% tolerance)

Baselines: GPT-4o GPT-4-turbo Gemini-1.5-Pro Gemini-1.5-Flash Qwen-VL-Max Qwen-VL-Plus Intern-VL-2.0-40B Intern-VL-2.0-8B Phi3-Vision-4B ChartGemma-2B TinyChart-3B ChartAst-S-13B

Builds on: ChartLlama ReST ENVISIONS vCoT ECharts

Closest papers by meaning

• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding (CVPR 2026 · similarity 0.91)
• CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs (NeurIPS 2024 · similarity 0.90)
• ChartSketcher: Reasoning with Multimodal Feedback and Reflection for Chart Understanding (NeurIPS 2025 · similarity 0.90)
• ChartMuseum: Testing Visual Reasoning Capabilities of Large Vision-Language Models (NeurIPS 2025 · similarity 0.90)
• Synthesize Step-by-Step: Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA (CVPR 2024 · similarity 0.89)
• Chart-FR1: Visual Focus-Driven Fine-Grained Reasoning on Dense Charts (CVPR 2026 · similarity 0.88)