tiiuae/visres_bench

VisRes Bench

VisRes Bench is a benchmark for evaluating the visual reasoning capabilities of Vision-Language Models (VLMs) in naturalistic settings without contextual language supervision. It is introduced in the paper VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs.

Paper Summary

Vision-Language Models excel at captioning and VQA, but it is unclear how much they rely on visual reasoning versus linguistic priors. VisRes addresses this by using image-only, four-choice tasks on real-world images (~19,000 samples) so that performance reflects visual reasoning rather than textual shortcuts.

The benchmark is organized in three levels of increasing complexity:

• Level 1 — Perceptual grounding: Local patch completion (masked tile + 4 candidate patches) under perturbations (blur, brightness, rotation, edges, location) and global occlusion (50% or 80% of the image masked). Tests robustness and amodal completion.
• Level 2 — Single-attribute rule: Raven-style 3×3 grids with one missing cell; one attribute (color, count, or orientation) follows a row-wise rule. Includes uniform, 3-different, 2-similar-1-different, count progression, arithmetic, and min-max subtasks (~5,956 samples).
• Level 3 — Multi-attribute composition: Same 3×3 format but multiple attributes (color, count, orientation, object identity) with row-wise, grid-wise, or spiral rules (~2,522 samples).

Main findings: State-of-the-art VLMs perform near random (25%) on many subtasks under subtle perceptual changes. Performance is stronger on color than count, and weakest on orientation. When the same logical structure is given as text, models do much better, indicating a visual-to-symbolic bottleneck rather than a pure reasoning limit. Higher resolution and guided/thinking prompts help but do not close the gap to human baselines.

---

Main Results (Guided Prompting, Thinking Mode When Available)

Accuracy (%) across levels and subtasks. Random chance = 25%.

Setting	GPT-5	GPT-4o	Gemini-2.5	Qwen3-VL-4B	Qwen3-VL-30B	Mimo-VL-7B
Level-1
Edges	27.17	23.91	25.00	16.67	25.00	22.30
Location	23.71	20.62	26.00	23.16	22.40	25.77
Rotation	35.42	26.04	34.38	37.50	36.05	29.17
Brightness	25.26	27.37	27.37	31.52	29.47	27.37
Blur	31.18	25.26	26.32	24.73	24.28	26.32
Global@50%	42.86	20.88	57.14	37.50	47.25	48.35
Global@80%	32.61	22.83	36.96	25.88	35.87	30.43
Level-1 Average	31.10	23.86	33.28	28.17	31.20	29.22
Level-2
Uniform Color	96.00	21.00	97.00	66.20	88.00	78.95
Uniform Count	61.00	25.00	90.91	40.82	59.00	52.75
Uniform Orientation	22.22	25.25	26.53	26.00	23.00	19.19
Count Progression	50.00	13.00	77.00	37.20	48.00	36.96
Count Arithmetic	52.00	22.00	75.76	43.20	49.00	33.33
Level-2 Average	49.79	24.12	62.29	37.18	46.75	39.15
Level-3
Independent Color-Object-Orientation	34.00	25.25	38.00	27.39	32.60	19.00
Independent Count-Object-Color	34.00	24.00	44.00	29.45	36.34	29.00
Coupled Color-Orientation	24.24	24.00	16.33	26.13	29.43	20.00
Coupled Color-Count	30.00	22.00	21.21	27.46	33.33	28.00
Spiral Color-Count-Object	56.00	30.00	54.17	28.63	36.00	33.00
Level-3 Average	34.39	23.86	33.73	26.31	31.36	25.17

---

Finetuning on Level-1 (Qwen2.5-VL-3B)

Setting	Original	Finetuned	Human Baseline
Location	24.3	42.8	94.1
Blur	23.9	37.5	84.3
Brightness	23.7	39.8	85.6
Rotation	25.5	50.8	92.0
Edges	25.1	33.2	82.6
Global (50%)	24.9	52.2	96.1
Global (80%)	23.9	38.6	98.0
Average	24.5	43.7	90.4

---

Single-Attribute Recognition (Perceptual Grounding)

Accuracy (%) when models are asked to report a single attribute (color, orientation, or count) for one grid cell.

Attribute	GPT-4o	GPT-5
Color	84.6	97.6
Orientation	39.8	49.6
Count	72.4	94.2

---

Impact of Thinking Mode

Accuracy (%) with thinking mode enabled (✓) vs disabled (✗). Open-source models improve substantially with thinking.

Level	GPT-5 (high)	GPT-5 (low)	Mimo-VL ✓	Mimo-VL ✗	Qwen3-4B ✓	Qwen3-4B ✗	Qwen3-30B ✓	Qwen3-30B ✗
Level-1	32.61	31.43	29.22	23.91	28.17	23.16	31.20	23.60
Level-2	49.79	47.01	39.15	26.68	37.18	24.08	46.75	28.25
Level-3	34.39	32.89	25.17	25.23	26.31	23.50	31.36	24.00

---

Impact of Image Resolution (GPT-5)

Accuracy (%) at different input resolutions. All levels improve with higher resolution.

Resolution	Level-1	Level-2	Level-3
512×512	45.17	42.83	31.63
1024×1024	54.01	49.61	35.48
2048×2048	56.51	48.99	40.07

---

Citation

@article{visres2025,
  title={VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs},
  author={Malagurski T{\"o}rtei, Brigitta and Dahou, Yasser and Huynh, Ngoc Dung and others},
  journal={arXiv preprint arXiv:2512.21194},
  year={2025}
}