In plain words
ZoomGain is a way to make and check questions about small details in pictures. It compares a full image with a zoomed crop, asks about details that become clear in the crop, checks answers using the crop, and adds plausible wrong choices.
In tests on 13 separate image-question tasks, a 9-billion-parameter model trained on ZoomGain examples scored 3.58 points higher on average than a same-size model trained on a matched set of questions. That result covers the tested models and tasks.
The research
A question about a small image feature is useful only if answering it depends on seeing that feature, rather than guessing from its wording or the surrounding scene.
ZoomGain’s results make the design of training questions a consequential choice: examples built around details revealed by zooming can outperform questions about the same images on the tested models and tasks.
ZoomGain turns details visible only in a crop into verified questions about the full image, then uses those questions to train visual models.
Researchers trained Qwen3.5 models with 4B and 9B parameters on image-matched sets of 9,000 questions, using the same training settings. They compared average scores across 13 held-out tests, leaving out ZoomGainBench because its questions share the training data’s generation pipeline.
The 9B model trained on ZoomGain questions averaged 66.10, 3.58 points above its ZwZ-trained counterpart. The 4B model also scored higher on average. These comparisons support the view that ZoomGain questions provide useful training signals beyond the specific questions used for training.
The training comparison covered two Qwen3.5 model sizes and 9,000 examples per dataset, with ZwZ images retaining bounding-box overlays while ZoomGain used original images. The component checks covered Qwen3.5-4B only.
What the evidence shows
- Questions need visual evidence
ZoomGain is an automated pipeline for building questions about fine image details. The challenge is that a detail hard to see in a full image can lead to a wrong answer, an unclear target, or a question that can be guessed without looking.
ZoomGain uses the difference between full and zoomed views to identify details for questions, then checks answers and constructs plausible wrong options.
- The questions challenge models
ZoomGainBench asks about details, counts, and text in full images. Its percentage scores measure model answers across these question types; the highest-scoring of the 28 tested models reached 61.08%, showing that the selected questions challenged even the leader in this evaluation.
- Wrong choices reveal a pattern
Researchers compared the models’ incorrect multiple-choice selections by distractor source. The share choosing options based on full-image errors exceeded the random-choice baseline, suggesting that models often select a plausible scene detail instead of resolving the detail the question asks about. This share counts incorrect choices, not all answers.
- Matched training favors ZoomGain
To test whether the question design mattered beyond the images, researchers trained 4B and 9B Qwen3.5 models on image-matched sets with the same training settings, then averaged scores across 13 tests excluded from training-data generation.
The 9B model trained on ZoomGain questions averaged 3.58 points above its ZwZ-trained counterpart. This supports a benefit in the tested comparison, though the largest models still led on some task groups.
- Two design choices matter
Researchers tested two design choices by training Qwen3.5-4B variants on 9,000 examples with the same settings and comparing averages across 10 tests excluded from training-data generation. Removing either zoom-revealed details or dedicated wrong answer choices lowered the average relative to the full version.
This supports both choices in this setup, while leaving their effects on other models and conditions open.
Deep
- Original paper authors
- Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang
- Submitted to arXiv
- License
- CC BY 4.0