List your company
ResearchFormerly WizWand

Spatial Reasoning on CV-Bench-3D (Multi-choice)

96.3AccuracyN3D-VLM-3B
25.673644.009362.34580.6807Dec 7, 2025Jan 3, 2026Jan 30, 2026Feb 26, 2026Mar 25, 2026Apr 21, 2026May 19, 2026
Updated 5mo ago

Evaluation results

MethodLinks
N3D-VLM-3BCategory=Ours2025.1296.3—
N3D-VLM-7BCategory=Ours2025.1293.3—
Qwen3-VL-8BCategory=Open-source2025.1291.3—
Gemini-2.5-FlashCategory=Closed-source2025.1286—
NashStrategy=Step-wise verification2026.0582.34—
LLaVA CriticStrategy=Step-wise verification2026.0581.58—
SpatialReasoner-7BCategory=Open-source2025.1280.3—
Qwen2.5-VL-7BFT-Data=-, Fine-tuned on identical training data=false2026.0380.1—
VG-LLMFT-Data=385k, Fine-tuned on identical training data=false2026.0379.1—
GeoSenseFT-Data=940K, Fine-tuned on identical training data=false2026.0378.7—
ViLASR-7BFT-Data=73K, Fine-tuned on identical training data=false2026.0377.2—
BaseStrategy=Step-wise verification2026.0576.39—
Qwen2.5-VL-7BCategory=Open-source2025.1275.8—
SpaceR-sft-7BFT-Data=151K, Fine-tuned on identical training data=false2026.0375.7—
Cambrain-S-3BFT-Data=10M, Fine-tuned on identical training data=false2026.0375.2—
SpatialLadder-3BCategory=Open-source2025.1274.9—
InternVL3-2BFT-Data=-, Fine-tuned on identical training data=false2026.0374.3—
SpatialLadder-3BFT-Data=26K, Fine-tuned on identical training data=false2026.0373.7—
GPT-4oCategory=Closed-source2025.1272.4—
Qwen2.5-VL-3BFT-Data=-, Fine-tuned on identical training data=false2026.0371.8—
VG-LLM*FT-Data=940k, Fine-tuned on identical training data=true2026.0371.7—
Qwen2.5-VL-3B*FT-Data=940k, Fine-tuned on identical training data=true2026.0367.9—
SpatialRGPT-VILA-1.5-8BCategory=Open-source2025.1263.3—
SiTebackbone=Qwen2-flickr, variant=rand2025.1251.67—
Qwen2variant=flickr2025.1251.57—
SiTebackbone=Qwen2-flickr, variant=ratio2025.1251.5—
SiTebackbone=Qwen2, variant=ratio2025.1250.4—
SiTebackbone=Qwen2, variant=rand2025.1250.36—
Qwen2variant=Baseline2025.1247.63—
SiTebackbone=LLaVA-flickr, variant=ratio2025.1239.8—
SiTebackbone=LLaVA-flickr, variant=rand2025.1238.66—
SiTebackbone=LLaVA, variant=ratio2025.1236.53—
SiTebackbone=LLaVA, variant=rand2025.1235.5—
VisionSR1Strategy=Step-wise verification2026.0534.28—
LLaVAvariant=Baseline2025.1233.43—
SherlockStrategy=Step-wise verification2026.0528.67—
LLaVAvariant=flickr2025.1228.39—
CoT promptingBackbone=Qwen2.5-VL-7B-Instruct, Model Scale=7B, Reasoning Strategy=CoT prompting2026.04—76.09
CoT promptingBackbone=Qwen2.5-VL-3B-Instruct, Model Scale=3B, Reasoning Strategy=CoT prompting2026.04—68
FGRPOBackbone=Qwen2.5-VL-7B-Instruct, Model Scale=7B, Reasoning Strategy=Faithful GRPO2026.04—87.04
FGRPOBackbone=Qwen2.5-VL-3B-Instruct, Model Scale=3B, Reasoning Strategy=Faithful GRPO2026.04—82.5
GPT-4o (CoT)Reasoning Strategy=CoT2026.04—86.42
GPT-5-nano (CoT)Reasoning Strategy=CoT2026.04—86.75
GRPO-TBackbone=Qwen2.5-VL-7B-Instruct, Model Scale=7B, Reasoning Strategy=RL (Task Reward Only)2026.04—85.92
GRPO-TBackbone=Qwen2.5-VL-3B-Instruct, Model Scale=3B, Reasoning Strategy=RL (Task Reward Only)2026.04—80.44
Non-reasoningBackbone=Qwen2.5-VL-7B-Instruct, Model Scale=7B, Reasoning Strategy=Non-reasoning2026.04—83.78
Non-reasoningBackbone=Qwen2.5-VL-3B-Instruct, Model Scale=3B, Reasoning Strategy=Non-reasoning2026.04—74
R1-OneVisionReasoning Strategy=MRM baseline2026.04—58
TreeVGRReasoning Strategy=MRM baseline2026.04—73.92
ViGoRL-SpatialReasoning Strategy=MRM baseline2026.04—86.14
Vision-R1Reasoning Strategy=MRM baseline2026.04—75.83
VL-RethinkerReasoning Strategy=MRM baseline2026.04—80.75