Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception

Zooming without Zooming v2. From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception. By Lai Wei and twelve others, from Shanghai Jiao Tong University, Ant Group, and three other institutions. Published on the twenty-seventh of October.

Multimodal perception forms the foundation of multimodal intelligence. Two tasks within it are singled out as pivotal: spatial perception, and multimodal long-context understanding. Together they underpin embodied AI and multimodal agents. Both share a practical difficulty, though. Synthesizing targeted, high-quality training data for them remains hard, and it remains costly. A model cannot be taught a skill well without examples of it, and here the examples are the expensive part.

The key insight is that the two tasks share an underlying capability, which the authors call connecting evidence. It has four parts: identifying the relevant visual facts, binding them to the correct entities, recovering their relations, and composing them into an answer. Crucially, targeted training data is far easier to synthesize for that capability than for the tasks themselves. So the plan is indirect. Rather than generating data for each hard task, strengthen the shared capability that sits beneath them.

The framework for doing that is EvidenceConnect, described as easy to implement. It works within a single image. Zoomed views of related regions are used to generate questions that require comparing, aggregating, or chaining pieces of visual evidence. Answering one means finding several facts and putting them together, which is the capability being targeted.

Two filters guard quality. A reference answer is retained only when independent VLMs agree on it. And region-ablation tests discard any question whose answer survives the removal of the selected regions, so a question that could be answered without that evidence does not make it through. Base VLMs are then trained on the resulting data with RLVR, and the model that comes out is Zooming without Zooming v2, or ZwZ2.

The synthesized data are not tailored to any downstream benchmark. Even so, ZwZ2 improves fine-grained perception and spatial perception. The gains go further than that, reaching embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning. Several of those are exactly the hard tasks the work set out to help, reached without any data built for them.

The route here is sideways. Rather than chase each hard task with its own costly data, this work trains the capability underneath them, using questions cheap enough to synthesize from single images. The authors take the results as establishing EvidenceConnect as an effective way to supervise evidence connection, and as a sign that strengthening that one shared capability leads to more generalizable multimodal perception.