EvidenceConnect data training produces ZwZ2 multimodal perception gains
Original title: Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception
Zoomed-region questions plus RLVR improved spatial, long-context, agent, and general visual reasoning without benchmark-specific data.
The paper argues that spatial perception and long-context multimodal understanding share a capability it calls connecting evidence. EvidenceConnect synthesizes comparison, aggregation, and chaining questions from zoomed regions of one image, keeps answers only when independent VLMs agree, and drops questions that still answer after region ablation. Base VLMs trained with RLVR become Zooming without Zooming v2 (ZwZ2). The authors report gains on fine-grained and spatial perception, embodied spatial reasoning, long-context understanding, agent tasks, and general visual reasoning.
Unconfirmed: Reported gains are the authors' results; independent replication is not described.
- Training data are synthesized from related regions inside a single image.
- A reference answer is kept only when independent VLMs agree.
- Region-ablation tests discard questions whose answers survive removing selected regions.
- The synthesized set is not tailored to any downstream benchmark.
From the original article
1 min read at huggingface.co
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.
By Lai Wei, Yatai Li, Liangbo He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.