EvidenceConnect trains VLMs to connect visual evidence
Lai Wei and 12 other authors submitted “Zooming without Zooming v2” to arXiv on October 27, 2026, introducing EvidenceConnect, a synthetic-data framework built around zoomed views of related regions in single images. The authors train base vision-language models on questions requiring visual-evidence comparison, aggregation, or chaining using RLVR, producing Zooming without Zooming v2 (ZwZ2). They report gains in fine-grained and spatial perception, embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning.
The paper offers a concrete way to generate training data for connecting visual evidence, with reported improvements across spatial perception, long-context reasoning, and agent tasks.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.37037 (cs)
[Submitted on 27 Oct 2026]
Title:Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception
Authors:Lai Wei, Yatai Li, Liangbo He, Jun Lan, Renjie Gu, Linhan Cao, Junkai Lin, Huijia Zhu, Weiqiang Wang, Linghe Kong, Yue Wang, Zhuosheng Zhang, Weiran Huang
View a PDF of the paper titled Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception, by Lai Wei and 12 other authors