Summary
Understanding complex scenes and long conversations that mix pictures and words is hard because training data is expensive to make. The authors noticed that a key skill shared by these tasks is 'connecting evidence' — finding important visual details, linking them to the right things, and putting the information together to answer questions. They created a method called EvidenceConnect that generates training questions by zooming in on parts of an image to make the model practice comparing and combining clues. Teaching models this way improved their ability to see and reason about visual details, even for tasks the training didn’t directly focus on.
What this means in practice
- •For embodied ai developers: Train AI agents to better understand spatial layouts and long visual contexts for navigation and interaction in environments.
- •For multimodal system engineers: Improve models that combine images and text to answer detailed questions involving multiple parts of a scene or long conversations.
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
Abstract
Multimodal perception forms the foundation of multimodal intelligence. In particular, spatial perception and multimodal long-context understanding are two pivotal tasks that underpin embodied AI and multimodal agents. However, synthesizing targeted, high-quality training data for these tasks remains difficult and costly. Our key insight is that these tasks share an underlying capability, which we call connecting evidence: identifying relevant visual facts, binding them to the correct entities, recovering their relations, and composing them into an answer. Crucially, targeted training data are far easier to synthesize for this capability than for these tasks themselves. To strengthen this capability, we propose EvidenceConnect, an easy-to-implement data synthesis framework that uses zoomed views of related regions within a single image to generate questions requiring comparison, aggregation, or chaining of visual evidence. To ensure quality, a reference answer is retained only when independent VLMs agree on it, and region-ablation tests discard questions whose answer survives the removal of selected regions. We then train base VLMs on the resulting data with RLVR, obtaining Zooming without Zooming v2 (ZwZ2). Although the synthesized data are not tailored to any downstream benchmark, ZwZ2 improves fine-grained and spatial perception, with further gains on embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning. These results establish EvidenceConnect as an effective framework for supervising evidence connection and suggest that strengthening this shared capability leads to more generalizable multimodal perception.