ZoomGain turns zoom-only details into exam questions: 9K samples of reinforcement learning let a 4B model beat its ZwZ-trained counterpart on all 14 benchmarks
Synopsis
ZoomGain is a data-construction pipeline that compares a full-image description of a region with a zoomed-in description of the same region and rewrites details visible only after zooming into questions about the whole image, verifying answers on the zoomed view with several models and discarding items answerable from language priors alone. On the resulting 1,000-question bilingual ZoomGainBench, the best of 28 vision-language models, GPT-5.6-Sol, scores only 61.08; after GRPO reinforcement learning on just 9K ZoomGain samples, Qwen3.5-9B averages 66.10 across 13 held-out benchmarks, above Qwen3.5-397B-A17B at 65.55.
Interpretation
The way fine-grained visual supervision is built has changed: questions are no longer written item by item by people or generated directly by a model, but reverse-engineered from details that are unclear in the full image and become definite only when zoomed in. Questions in earlier benchmarks of this kind were mostly annotated by hand, or model-generated and then heavily edited by people, which made scalable curation costly and rarely filtered out items answerable from language common sense alone. On ZoomGainBench, 1,000 questions over 941 images, the top-scoring of 28 vision-language models, GPT-5.6-Sol, reaches 61.08, and only five models exceed 50.
The pipeline first grounds an anchor object that can be uniquely referred to in the full image, then describes the same region separately in the global view and the zoomed view, and uses the difference between the two descriptions as the basis for questions. Incremental facts are kept only when the zoomed-view caption states them explicitly while the global-view caption omits them, marks them unclear, or states them differently, so answering requires that detail itself. An answer is retained only if several models, sampled repeatedly on the zoomed view, each meet a per-model accuracy threshold; a question is discarded if all blind models answer it correctly without an image, and multiple-choice distractors come from three sources: other instances in the image, blind guesses, and full-image errors.
The same data also serves directly as training supervision, and the gains are not confined to fine-grained perception. Against ZwZ under identical source images, identical reinforcement-learning settings and 9,000 samples per side, ZoomGain-trained models outperform their ZwZ-trained counterparts on all 14 benchmarks. Trained on ZoomGain data, Qwen3.5-9B gains 11.01 points on O3-Bench and averages 66.10 over 13 held-out benchmarks; the Qwen3.5-4B version averages 63.96, above Qwen3.5-122B-A10B at 63.04.
Perspective
The demonstrated scope is a 1,000-question, 941-image bilingual benchmark with attribute-recognition, counting and OCR questions, built on high-resolution images collected by ZwZ; the training conclusions are limited to Qwen3.5-4B and Qwen3.5-9B under GRPO reinforcement learning with 9,000 samples and image-matched, identically configured ZwZ comparisons.
The error analysis finds that 71.7% of the 28 models' incorrect choices fall on full-image-error distractors, above the 54.5% expected under random choice, which the authors read as a weakness shared across current vision-language models rather than one specific to individual models; the authors also note that the strongest large models still lead on general perception and difficult perception-and-reasoning tasks, which remain dependent on model scale and reasoning capability.