Reproduce “Improving VLMs on Scientific Chart Analysis via Autonomous Evolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Abhinav Sudhakar Dubey, Vidushi Goel, Kiran Rana, Shikhar Mathur, Vaggos Chatziafratis
The authors evolve inference programs around fixed vision-language models for chart question answering, changing prompts, answer-length budgets, and output cleanup without retraining. Across 12 combinations of three models and four benchmarks, mean descriptive exact-match accuracy rises from 16.6% to 41.3%. The largest gain is Qwen3-VL-2B on ChartQA, from 0.5% to 55.7%. These results concern descriptive answers, not improved visual perception.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.