ChartEvo: Self-Evolving Agent Harnesses for Long-Tail Chart Reasoning

Key points

• Multimodal large language models (MLLMs) have substantially advanced chart understanding and reasoning, yet their performance remains highly uneven across chart types.1
• We introduce ChartEvo, a recursive self-improvement framework that enables chart agents to adapt to underperforming chart types through experience-driven harness-level self-improvement.1
• ChartEvo jointly evolves reusable chart-specific Skills and executable Tools, while learning reusable local tool-call chains to compose evolved capabilities and control the inference overhead of tool-augmented reasoning.1
• On long-tail charts from ChartBench and ChartX, ChartEvo improves accuracy over its base MLLM, Qwen3.7-Plus, by 45.6% and 32.5%, respectively, demonstrating the effectiveness of recursive self-improvement for long-tail chart adaptation.1

Sources (1)

• [1]ChartEvo: Self-Evolving Agent Harnesses for Long-Tail Chart Reasoning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 02:34 PM

  Tool-augmented agents can improve complex chart reasoning by introducing specialized visual capabilities, but manually engineering strategies and tools for the heterogeneous and evolving long tail is difficult to scale.

  Multimodal large language models (MLLMs) have substantially advanced chart understanding and reasoning, yet their performance remains highly uneven across chart types.

Extractive summary: sentences quoted from the sources.