Reproduce “Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGER” 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
Gregory Kang Ruey Lau, Minh Huynh Nguyen, Bryan Kian Hsiang Low
TIGER trains models on rendered-image problems they answer correctly from text but fail from images, using image-conditioned reinforcement learning. On MATH-500, Qwen3-VL-8B scored 75.53 on text, 76.40 on extracted text, and 70.87 on images despite 96.73 OCR accuracy. For Qwen2.5-VL-7B, accuracy rose 3.6 points on MathVerse and 3.5 on EMMA; training uses synthetic text images, limiting real-world visual coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.