Benchmark Almanac

SpaceDG-Bench

SpaceDG-SFT-Qwen3-VL-8B-Instruct 66.1% tops its leaderboard Gemini 3.1 Pro 56.7% leads the models we track
Not in model cards

Models answer spatial questions under clean images and nine forms of visual degradation. Choice and binary answers use accuracy, while numerical answers use relative-error thresholds. The test has 1,102 verified questions from scenes excluded from the training set.

This record covers SpaceDG-Bench, not the million-question SpaceDG training collection. The benchmark renders clean and degraded versions of the same questions, so its roughly ten thousand instances are not independent questions.

Released
May 2026
Kind
Spatial reasoning and counting
Level
Everyday
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images → Text
Interaction
Single turn
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Vision
Strongest model tested
GPT-5.4 #28 on the capability index
Last model added
April 2026
Scores refreshed
November 18th, 2026

In the authors’ words

Multimodal Large Language Models (MLLMs) have made rapid progress in spatial intelligence, yet existing spatial reasoning benchmarks largely assume pristine visual inputs and overlook the degradations that commonly occur in real-world deployment, such as motion blur, low light, adverse weather, lens distortion, and compression artifacts. This raises a fundamental question: how robust is the spatial intelligence of current MLLMs when visual observations are imperfect? To answer this question, we introduce SpaceDG, the first large-scale dataset for degradation-aware spatial understanding. It is constructed with a physically grounded degradation synthesis engine that embeds degradation formation process into 3D Gaussian Splatting (3DGS) rendering, enabling realistic simulation of nine degradation types. The resulting dataset contains approximately 1M QA pairs from nearly 1,000 indoor scenes. We further introduce SpaceDG-Bench, an human-verified benchmark with 1,102 questions spanning 11 reasoning categories and 9 visual degradation types, yielding over 10K VQA instances. Evaluating 25 open- and closed-source MLLMs reveals that visual degradations consistently and substantially impair spatial reasoning, exposing a critical robustness gap. Finally, we show that finetuning on SpaceDG markedly improves degradation robustness and can even surpass human performance under degraded conditions without any performance drop on clean images, highlighting the promise of degradation-aware training for robust spatial intelligence.

Abstract of SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation (arXiv:2605.22536, CC0)
27
Entries in its paper, 20 models we track

SpaceDG-SFT-Qwen3-VL-8B-Instruct tops the leaderboard with 66.1% Gemini 3.1 Pro leads with 56.7% among models released by April 2026

Best score per entry from the benchmark's paper, read on November 18th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Gemini 3.1 Pro, at 56.7%. This view shows only models we track, each at its best.

The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 SpaceDG-SFT-Qwen3-VL-8B-Instruct 9918 degraded instances 66.1%
  2. 2 SpaceDG-SFT-InternVL-3.5-8B 9918 degraded instances 65.2%
  3. 3 Gemini 3.1 Pro 9918 degraded instances 56.7%
  4. 4 SenseNova-SI-InternVL3-8B 9918 degraded instances 53.3%
  5. 5 Gemini 3.1 Flash-Lite 9918 degraded instances 48.8%
  6. 6 InternVL3.5-38B 9918 degraded instances 47.7%
  7. 7 GPT-5.4 9918 degraded instances 46.2%
  8. 8 Qwen 3.6 Plus 9918 degraded instances 46%
  9. 9 RynnBrain-8B 9918 degraded instances 45.2%
  10. 10 Qwen3-VL-32B-Instruct 9918 degraded instances 45.2%
  11. 11 ACE-Brain-0-8B 9918 degraded instances 45.1%
  12. 12 Claude Sonnet 4.6 9918 degraded instances 44.2%
  13. 13 Qwen3.5-27B 9918 degraded instances 43.8%
  14. 14 Qwen 3.6 35B-A3B 9918 degraded instances 42.4%
  15. 15 Qwen3-VL-8B-Instruct 9918 degraded instances 42.1%
  16. 16 Qwen3.5-9B 9918 degraded instances 41.7%
  17. 17 InternVL3.5-8B 9918 degraded instances 41.4%
  18. 18 VST-7B 9918 degraded instances 41.1%
  19. 19 Qwen3.5-4B 9918 degraded instances 40.4%
  20. 20 InternVL3-8B 9918 degraded instances 40.1%
  21. 21 Qwen3-VL-4B-Instruct 9918 degraded instances 39.7%
  22. 22 Grok 4.1 Fast 9918 degraded instances 35.6%
  23. 23 Llama 4 Maverick 9918 degraded instances 34.3%
  24. 24 Kimi-VL-A3B-Instruct 9918 degraded instances 34%
  25. 25 LLaVA-OneVision 7B (SI) 9918 degraded instances 33.8%
  26. 26 Gemma 4 26B A4B 9918 degraded instances 31%
  27. 27 Cambrian-S-7B 9918 degraded instances 26.9%

Source: The SpaceDG-Bench paper

Source: The SpaceDG-Bench paper

#EntryScoreSettingDateSource
1 SpaceDG-SFT-Qwen3-VL-8B-Instruct 66.1% 9918 degraded instances Official leaderboard
2 SpaceDG-SFT-InternVL-3.5-8B 65.2% 9918 degraded instances Official leaderboard
3 Gemini 3.1 Pro 56.7% 9918 degraded instances Feb 2026 Official leaderboard
4 SenseNova-SI-InternVL3-8B 53.3% 9918 degraded instances Official leaderboard
5 Gemini 3.1 Flash-Lite 48.8% 9918 degraded instances Mar 2026 Official leaderboard
6 InternVL3.5-38B 47.7% 9918 degraded instances Aug 2025 Official leaderboard
7 GPT-5.4 46.2% 9918 degraded instances Mar 2026 Official leaderboard
8 Qwen 3.6 Plus 46% 9918 degraded instances Mar 2026 Official leaderboard
9 RynnBrain-8B 45.2% 9918 degraded instances Official leaderboard
10 Qwen3-VL-32B-Instruct 45.2% 9918 degraded instances Oct 2025 Official leaderboard
11 ACE-Brain-0-8B 45.1% 9918 degraded instances Official leaderboard
12 Claude Sonnet 4.6 44.2% 9918 degraded instances Feb 2026 Official leaderboard
13 Qwen3.5-27B 43.8% 9918 degraded instances Feb 2026 Official leaderboard
14 Qwen 3.6 35B-A3B 42.4% 9918 degraded instances Apr 2026 Official leaderboard
15 Qwen3-VL-8B-Instruct 42.1% 9918 degraded instances Oct 2025 Official leaderboard
16 Qwen3.5-9B 41.7% 9918 degraded instances Feb 2026 Official leaderboard
17 InternVL3.5-8B 41.4% 9918 degraded instances Aug 2025 Official leaderboard
18 VST-7B 41.1% 9918 degraded instances Official leaderboard
19 Qwen3.5-4B 40.4% 9918 degraded instances Feb 2026 Official leaderboard
20 InternVL3-8B 40.1% 9918 degraded instances Apr 2025 Official leaderboard
21 Qwen3-VL-4B-Instruct 39.7% 9918 degraded instances Oct 2025 Official leaderboard
22 Grok 4.1 Fast 35.6% 9918 degraded instances Nov 2025 Official leaderboard
23 Llama 4 Maverick 34.3% 9918 degraded instances Apr 2025 Official leaderboard
24 Kimi-VL-A3B-Instruct 34% 9918 degraded instances Apr 2025 Official leaderboard
25 LLaVA-OneVision 7B (SI) 33.8% 9918 degraded instances Aug 2024 Official leaderboard
26 Gemma 4 26B A4B 31% 9918 degraded instances Apr 2026 Official leaderboard
27 Cambrian-S-7B 26.9% 9918 degraded instances Official leaderboard
#ModelScoreSettingDateSource
1 Gemini 3.1 Pro 56.7% 9918 degraded instances Feb 2026 Official leaderboard
2 Gemini 3.1 Flash-Lite 48.8% 9918 degraded instances Mar 2026 Official leaderboard
3 InternVL3.5-38B 47.7% 9918 degraded instances Aug 2025 Official leaderboard
4 GPT-5.4 46.2% 9918 degraded instances Mar 2026 Official leaderboard
5 Qwen 3.6 Plus 46% 9918 degraded instances Mar 2026 Official leaderboard
6 Qwen3-VL-32B-Instruct 45.2% 9918 degraded instances Oct 2025 Official leaderboard
7 Claude Sonnet 4.6 44.2% 9918 degraded instances Feb 2026 Official leaderboard
8 Qwen3.5-27B 43.8% 9918 degraded instances Feb 2026 Official leaderboard
9 Qwen 3.6 35B-A3B 42.4% 9918 degraded instances Apr 2026 Official leaderboard
10 Qwen3-VL-8B-Instruct 42.1% 9918 degraded instances Oct 2025 Official leaderboard
11 Qwen3.5-9B 41.7% 9918 degraded instances Feb 2026 Official leaderboard
12 InternVL3.5-8B 41.4% 9918 degraded instances Aug 2025 Official leaderboard
13 Qwen3.5-4B 40.4% 9918 degraded instances Feb 2026 Official leaderboard
14 InternVL3-8B 40.1% 9918 degraded instances Apr 2025 Official leaderboard
15 Qwen3-VL-4B-Instruct 39.7% 9918 degraded instances Oct 2025 Official leaderboard
16 Grok 4.1 Fast 35.6% 9918 degraded instances Nov 2025 Official leaderboard
17 Llama 4 Maverick 34.3% 9918 degraded instances Apr 2025 Official leaderboard
18 Kimi-VL-A3B-Instruct 34% 9918 degraded instances Apr 2025 Official leaderboard
19 LLaVA-OneVision 7B (SI) 33.8% 9918 degraded instances Aug 2024 Official leaderboard
20 Gemma 4 26B A4B 31% 9918 degraded instances Apr 2026 Official leaderboard

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Google DeepMind
Benchmark paper Not tested since May 2026

Source: The SpaceDG-Bench paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemini 3.1 Pro InternVL3.5-38B InternVL3-8B Llama 4 Maverick LLaVA-OneVision 7B (SI) 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemini 3.1 Pro InternVL3.5-38B

Source: Official SpaceDG-Bench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Vision

All 636 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. Qwen 3.6 Plus 86%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy. Seed2.1 Pro 90.7%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for Qwen 3.6 35B-A3B. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (27 entries).
  • Pass Models scored 20 models we track, plus 7 other entries on its leaderboard.
  • Warning Newest score The newest came 7 months ago, in April 2026, for Qwen 3.6 35B-A3B.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.4 (#28).
  • Pass Headroom The best is 56.7%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 22 rows for models we track and 7 other entries at the last read.
  • Pass Last read November 18th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check