Train models to link scattered evidence, not just zoom in: ZwZ2-27B averages 60.51% across 25 benchmarks

The training target for multimodal models shifts from seeing one region clearly to joining several dispersed pieces of evidence in a single image into one answer. Spatial perception asks a model to recover how entities are arranged and how several spatial relations compose, while long-context understanding asks it to locate facts across images and dialogue turns and bind them to entities; the two were previously treated as separate data-construction problems. Earlier synthetic data either supervised fine-grained perception around a single region or concentrated on a few multi-image and multi-turn formats; spatial relation labels need precise localization and relational grounding, which is expensive by hand and often hallucinated by automatic generators. On the same 4B base model, training on 20.4K synthesized examples raises the average accuracy over 11 benchmarks from the base model's 47.16 to 53.86; by contrast, the single-region ZwZ dataset with 73K examples reaches only 49.86, and four other synthetic datasets, TVG, F2RVLM, CRIT and MMHops, all train to averages below the base model.

Each training question is built on a coherent group of regions from one high-resolution image, so the model must locate those regions itself, bind facts to the right entities and compose an answer. EVIDENCECONNECT is the name of the fully automated synthesis pipeline: a strong vision-language model proposes candidate regions and groups them by task family, and native-resolution crops of the selected regions are arranged by relative position on a canvas to form a privileged zoomed evidence view used only during data construction. Questions are generated from that view and require comparing, aggregating or chaining evidence; a reference answer is kept only when vision-language models from two different families mostly agree, and each region is then masked in turn, with a region marked redundant if the answer survives its removal and a question discarded when more than one region is redundant.

During training the model never sees the zoomed view, only the original image with the selected regions outlined in color, so it still has to perceive and connect the evidence itself. The synthesized data are converted into four-option multiple-choice questions to support rule-based rewards, and a lightweight vision-language model filters out examples it always answers correctly. The final set holds 20.4K examples, each involving 3 to 7 regions, with 65.74% using three regions and 24.34% using four; by difficulty, easy, medium, hard and very hard examples account for 33.75%, 22.04%, 19.07% and 25.14%. All three scales share the same data and the same GRPO configuration for two epochs.

The demonstrated setting is 20.4K multiple-choice examples synthesized from SA-1B single images, used to train Qwen3.5-4B/9B and Qwen3.6-27B with GRPO for two epochs and evaluated on 25 benchmarks; the authors propose extending synthesis to evidence spread across video frames and studying whether the same supervision helps larger models or earlier stages such as pre-training.

The authors report that gains are not uniform: in the 7K comparison, drop-one verification lifts the 11-benchmark average from 51.03 to 52.65 but ERQA falls by 4.00 points, and privileged-view synthesis is 1.27 points higher on average yet lower on Chartography and ERQA. Long-context gains narrow at the longest contexts, with ZwZ2-4B and ZwZ2-9B merely matching their base models at 128K on the MMLongBench visual core, which the authors attribute to every training example being a single image with a short context.