Fine-grained visual perception requires Vision-Language Models (VLMs) to locate and interpret subtle evidence within complex scenes. Constructing supervision for this ability is challenging: questions must have correct, unambiguous answers that depend on actual fine-grained visual details rather than coarse cues or language priors. Verifying these properties makes manual annotation labor-intensive and automatic generation difficult for models that themselves struggle with fine-grained perception. Yet the same models can often recognize such details once the region is zoomed in. We exploit this global-to-zoom perception gap and introduce ZoomGain, a data construction pipeline that constructs questions about the full image from visual information gained only by zooming in. It verifies each answer on the zoomed view with multiple models, ensures that each question remains unambiguous in the full image and cannot be answered from language priors, and specifically generates high-quality distractors for multiple-choice questions. With this pipeline and human-only screening (no manual QA editing), we build ZoomGainBench, a 1,000-question bilingual benchmark on which the best of 28 VLMs achieves only 61.08. Failure analysis reveals that models favor a particular type of distractor when answering multiple-choice questions incorrectly. Beyond evaluation, training on just 9K ZoomGain samples substantially improves the visual perception capabilities of our 4B/9B models, enabling them to outperform several hundred-billion-parameter models across various perception benchmarks. The trained 9B model achieves SoTA performance on several benchmarks, such as 90.25 on HR-Bench 4K. These results establish global-to-zoom perception gaps as a practical basis for scaling fine-grained visual supervision for both evaluation and training. Our code is available at https://github.com/inclusionAI/ZoomGain.
Figures & tables
Figure 1: ZoomGain: turning global-to-zoom perception gaps into fine-grained visual supervision. (a) The runner’s shoe color is unclear in the global view but discernible in the zoomed view; ZoomGain asks about such details in the full image, using a global referring expression to identify the target. The paired views Vgi and Vzi , and the referring ei are defined in Section 3.2 . (b) ZoomGainBench, obtained by human screening without QA editing, challenges all 28 evaluated VLMs (c) RLVR on 9K training QA pairs improves Qwen3.5-4B and Qwen3.5-9B over their base models and ZwZ-trained counterparts (ICML’26 Wei et al. (2026) ).
Figure 2: Overview of ZoomGain. (a) Discover anchors and obtain consensus regions through independent grounding. (b) Mine global-to-zoom incremental evidence from paired views to construct questions and reference answers. (c) Verify answers and visual dependence, then assemble and recheck multiple-choice questions with multi-source distractors.
Benchmark
#Questions
QA authoring
Language
Evidence
Blind filter
Local verif.
Dual-view
Dedicated distractor generation
ZoomBench Wei et al. (2026)
845
Model+Human
EN
Auto
✗
✗
✓
✗
V* Bench Wu and Xie (2023)
191
Human
EN
Human
✗
✗
✗
✗
HR-Bench Wang et al. (2025b)
200
Model+Human
EN
✗
✗
✗
✗
✗
VisualProbe Lai et al. (2025)
500
–
EN
✗
✗
✗
✗
N/A
TreeBench Wang et al. (2025a)
405
Model+Human
EN
Human
✗
✗
✗
✗
FINERS-4k Zhang et al. (2026b)
2,427
Human
EN
–
✗
✗
✗
✗
Table 1: Comparison with existing fine-grained perception benchmarks. QA authoring : who writes the questions and answers; Model+Human denotes model-generated questions refined by human annotators, and Model denotes questions authored entirely by models, with humans only deciding which candidates to include. Blind filter : questions answerable without the image are removed during construction. Local verif. : answers are automatically verified on the zoomed evidence region. Dual-view : both the full image and the evidence region are provided. Dedicated distractor generation : a separate stage for constructing distractor options (Section 3.3 ). In this column, ✗ indicates that no dedicated stage is reported, and N/A denotes open-ended evaluation. “–” denotes not provided or not specified.
English
Chinese
Overall
Model
Attr
Count
OCR
Overall
Attr
Count
OCR
Overall
Closed-source Frontier VLMs
GPT-5.5 OpenAI ()
52.75
57.45
60.40
56.51
52.12
63.83
60.96
57.06
56.79
Gemini-3.6-Flash Google DeepMind (2026)
51.27
42.55
74.71
60.62
48.52
45.74
71.42
58.20
59.41
Claude-Opus-5 Anthropic (2026)
48.09
63.83
71.86
59.89
48.73
62.77
71.29
59.84
59.86
GPT-5.6-Sol OpenAI ()
56.57
62.77
67.93
62.08
54.03
63.83
65.84
60.07
61.08
Table 2: Evaluation results on ZoomGainBench. Scores are reported as percentages for English and Chinese questions. Each language-specific Overall is weighted by the number of questions in each category. Bold and underlined scores indicate the best and second-best results across all models in each column, respectively.
Figure 3: Where models fail on ZoomGain Bench. (a) How often each model group chooses each type of distractor compared with the random baseline, considering only incorrect answers. Values above 1 indicate over-selection. Global-view error + blind guess denotes distractors produced both as incorrect answers when models view the full image and as guesses based on language priors without an image. (b) How often the 26 models not used in construction (all evaluated models except Qwen3.5-397B-A17B and Qwen3-VL-235B-A22B-Instruct) choose global-view errors, grouped by each error’s frequency in the 30 full-image answers collected during construction. (c) Accuracy by quartile of the anchor’s area in the full image.
Held-in
Fine-grained Perception
General Perception
Difficult Percept & Reason
Model
ZoomGain Bench
Zoom Bench
V* Bench
HR 4K
HR 8K
MME-RW Lite
Tree Bench
Visual Probe
CV- Bench
MMStar
MMVP
Baby Vision
Perception Bench
O3- Bench
Avg. Only Held-out
Several Large Parameter VLMs
GPT-5.1
23.22
47.22
70.16
67.00
65.25
45.86
47.65
29.90
84.22
71.60
32.67
13.92
28.97
31.01
48.88
Qwen3-VL-235B-A22B
40.51
49.11
87.96
84.50
81.62
58.16
53.58
46.41
86.72
76.33
64.00
18.30
35.50
60.58
61.75
Qwen3.5-122B-A10B
41.11
51.83
84.82
86.50
83.75
57.89
54.07
44.85
88.06
74.60
70.00
19.07
38.80
65.22
63.04
Qwen3.5-397B-A17B
45.16
54.91
87.43
86.50
85.12
57.06
53.33
49.51
89.57
76.00
72.67
20.36
47.50
72.17
65.55
Table 3: Results of RL training on ZoomGain and ZwZ data. ΔZoomGain−ZwZ and ΔZoomGain−Base denote score gains over the ZwZ-trained and base models, respectively; Avg. is computed over the 13 held-out benchmarks, excluding ZoomGain Bench .
Model
Incremental Evidence
Dedicated Distractors
Held-in
Fine-grained Perception
Difficult Perception & Reason
Avg. Only Held-out
ZoomGain Bench
Zoom Bench
V* Bench
HR 4K
HR 8K
MME-RW Lite
Tree Bench
Visual Probe
Baby Vision
Perception Bench
O3- Bench
Qwen3.5-4B
–
–
30.03
48.88
81.68
81.25
77.63
57.89
49.38
43.30
14.69
33.10
54.20
54.20
+ZoomGain 9K
✗
✓
49.71
56.45
88.48
85.50
81.75
60.24
48.40
52.82
18.56
34.70
58.84
58.57
+ZoomGain 9K
✓
✗
46.85
57.04
90.05
85.75
80.25
62.69
48.89
54.76
17.53
34.97
59.42
59.14
+ZoomGain 9K
✓
✓
55.81
59.88
90.05
87.25
83.88
61.59
49.14
57.28
21.13
36.23
63.19
60.96
Table 4: Ablation of incremental evidence and dedicated distractors. Avg. is computed over the 10 held-out benchmarks.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Conditional anchor markers in the zoomed view. (a) Two people are present, including the partially visible person on the left; the red rectangle selects the person wearing a red-and-black plaid shawl. (b) Only one person is present, so no marker is added. Both views are rendered from the original images and recorded crop/anchor coordinates using the pipeline’s marker and resizing functions. The rectangle identifies the anchor rather than the answer-bearing detail; it is used only in the zoomed view and is not part of the photographed scene.
Model
Roles
Qwen3.5-397B-A17B Team (2026)
Initial global scene captioning and paired-view captioning ( Mcap ); grounding; crop anchor enumeration; joint crop/full-image audit and rewriting; blind answering; global-view error collection; visual distractor generation; final MCQ verification.
Candidate distractor clustering and option-pool curation.
Appendix
Table 5: Model assignments in the ZoomGain construction pipeline.
Figure 5: Attribute recognition examples. (a) Mirror housing color; (b) tabletop material; (c) logo shape. Correct answers and their corresponding options are highlighted in green.
Figure 6: Counting examples. (a) Eight burger images on a menu; (b) three people seated inside a shelter; (c) two chest pockets on a shirt. Each question requires an open-ended numerical answer.
Figure 7: Text recognition examples. (a) A supplementary road sign; (b) lettering on a truck’s rear door; (c) the largest Chinese text on a bottle label. Reference transcriptions are shown in green.
Figure 8: Representative failures on ZoomGainBench. Each case shows the full image with the queried detail outlined in red, an enlarged view of that region, and the responses of the 56 evaluations (28 models × 2 languages). (a) Nearly all evaluations name Bauer, a hockey equipment brand, instead of the NHL logo, while the blind guesses (Reebok, Nike) are never chosen. (b) The white tote bag is taken for a plastic bag; the visual-evidence distractors, drawn from the black bag on her shoulder and from another pedestrian’s purse, are never chosen. (c) Most evaluations (46/56) call the white object on the truck’s dashboard orange, likely the color of the vest worn by the worker in front of the truck. (d) The salient reflected silhouettes draw choices away from the smaller crane at the base of the mast.
Figure 9: Attribute QA construction example. The global-view caption mentions logos on a visor without specifying their shapes. The zoomed-view caption identifies a small star-shaped logo. This fact supports a shape question, which is rewritten using the global referring expression while preserving the answer.
Figure 10: Counting QA construction example. The global-view caption describes food images on digital menu boards without giving a count. The zoomed-view caption identifies eight burger images arranged in two rows of four. The selected fact yields a counting question with the same answer in both views.
Figure 11: OCR QA construction example. The global-view caption describes a bottle with a blue label without transcribing its text. The zoomed-view caption reads the Chinese characters on the label. The selected text fact supports a local OCR question, which is rewritten to identify the bottle in the full image while retaining the original transcription as the answer.
Figure 12: Examples rejected by crop-based verification. The captions incorrectly describe (A) the race-car digit as “3” instead of “8” and (B) the door handle as silver instead of white. Both verification models give the visually supported answer in all 10 samples, yielding zero matches to the generated reference answer and failing the 6/10 threshold. The original crop inputs have no red anchor markers; orange boxes and enlarged details are added only for display.
Figure 13: Examples rejected by blind verification. The reference answers, (A) drumsticks and (B) steam, are supported by the images, but are also predictable from the questions alone. Without images, the three models obtain 30/30 and 29/30 reference matches in total, respectively, with every model meeting the rejection threshold of at least 3/10 matches. The images and blue crop boxes are shown only for the reader; no images are supplied to the blind models.