inclusionAI's ZwZ2 Improves Base VLMs on 25 Perception Tests

inclusionAI's new Zooming without Zooming v2 paper describes a family of 4B, 9B and 27B vision-language models that it reports improve over their base versions on all 25 benchmarks across six domains.

The recipe is a data pipeline, not a model trick. EvidenceConnect, as the authors name it, 'uses zoomed views of related regions within a single image to generate questions requiring comparison, aggregation, or chaining of visual evidence.' A candidate question survives only when 'independent VLMs agree on it, and region-ablation tests discard questions whose answer survives the removal of selected regions.' The filtered corpus is then used to train base VLMs with RLVR.

The capability the authors are after is what the paper calls connecting evidence: 'identifying relevant visual facts, binding them to the correct entities, recovering their relations, and composing them into an answer.' Although the training set is single-image, the paper reports gains on embodied spatial reasoning, multimodal long-context understanding and multimodal agent tasks.

The abstract reports no per-benchmark deltas, and the comparison baseline is each model's own base rather than any outside VLM. Code, weights for all three sizes and the ZwZ2-RL-VQA dataset are posted to the project's Hugging Face collection, one of 354 open-source AI stories we have logged in the past 90 days.