Reproduce “Look Where You Say You're Looking: Self-Grounded Attention for Visual Reasoning” 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
Uri Berger, Gal Chechik, Gal Dalal
Self-Saliency trains a vision-language model to reward attention to image regions grounded from its own observation steps, rather than regions chosen in advance. It uses phrase grounding during post-training, with no external grounding model needed at inference. On 25 visual-reasoning benchmarks, Qwen3-VL-8B-Instruct reaches a mean score of 64.3 versus 63.5 for No-Sal; attention remains a proxy, not proof of understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.