Reproduce “Look Where It Counts: A Free, Label-Free Visual Evidence Signal for Fine-Grained Vision-Language 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

Santi Ram Tiwari, Nihal Naik, Devbrat Pandey, Nishant Sinha

The paper presents a label-free method for selecting image crops that make a vision-language model’s answer distribution most decisive. Using a 16-crop grid, it locates regions at rates of 0.835, 0.800, and 0.922 across models and improves Qwen3-VL-8B accuracy from 0.704 to 0.852 on fine-detail questions. Evidence-sensitive scores improve AUC, but training does not beat the base model on HR-Bench.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.