List your company

Spatial Reasoning on VSI-Bench (test)

79.2Avg ScoreHuman level
30.8443.39555.9568.505Feb 24, 2026Mar 8, 2026Mar 20, 2026Apr 1, 2026Apr 13, 2026Apr 25, 2026May 8, 2026

Evaluation results

MethodLinks
Human level2026.0579.294.34760.445.994.795.895.8100—
3D-RFT-4BTraining Stage=Reinforcement Fine-Tuned2026.0362.871.253.570.363.260.877.937.667.8—
VLM-3R-7BTraining Stage=Supervised Fine-Tuned2026.0360.970.249.469.267.165.480.545.440.1—
VST-SFT-3BTraining Stage=Supervised Fine-Tuned2026.0357.969.345.471.862.4594638.770.2—
VST-RL-3BTraining Stage=Reinforcement Fine-Tuned2026.0357.766.64572.860.959.947.640.768.3—
Cambrian-S-3BTraining Stage=Supervised Fine-Tuned2026.0357.370.740.66846.364.861.927.378.8—
Gemini-2.5 ProTraining Stage=Zero-shot2026.0351.543.834.964.342.861.147.845.971.3—
Spatial-MLLM-4BTraining Stage=Reinforcement Fine-Tuned2026.0348.465.334.863.145.141.346.233.546.3—
Spatial-MLLM-4BAccess Type=Open-source, Input Frames=16-322026.0548.465.334.863.145.141.346.233.546.31
VG LLM-4BTraining Stage=Supervised Fine-Tuned2026.0347.36637.855.259.244.645.633.536.4—
Proxy3DAccess Type=Open-source, Backbone=Qwen2.5-VL-7B, Input Frames=162026.054763.941.967.242.850.346.531.4322
SpatialLadder-3BTraining Stage=Reinforcement Fine-Tuned2026.0345.763.534.361.743.945.444.835.636.4—
ViLaSR-7BTraining Stage=Reinforcement Fine-Tuned2026.0345.463.534.460.630.948.945.230.449.2—
Gemini-1.5 ProAccess Type=Proprietary via API, Input Frames=16-322026.0545.456.230.964.143.651.346.33634.63
PyVision-VideoBackbone=Qwen2.5-VL-7B, Training=RL, Tooling=dynamic tooling2026.024453.825.850.838.244.846.326.358.6—
SpaceR-7BTraining Stage=Reinforcement Fine-Tuned2026.0343.561.928.660.935.238.24631.445.6—
InternVL3-8BTraining Stage=Zero-shot2026.0342.168.13948.433.648.336.427.335.4—
VITALAgent Type=agentic video model, Tools=predefined video clipping tools, Backbone=Qwen2.5-VL-7B, Training=RL2026.0241.8—————————
LLaVA-Video-72BAccess Type=Open-source, Input Frames=16-322026.0540.948.922.857.435.342.436.73548.64
vsGRPO-7BTraining Stage=Reinforcement Fine-Tuned2026.0340.759.929.650.848.335.435.63431.5—
LLaVA-OV-72BAccess Type=Open-source, Input Frames=16-322026.0540.243.523.957.637.542.539.932.544.65
Video-R1Reasoning Type=pure textual reasoning, Backbone=Qwen2.5-VL-7B, Training=RL2026.0237.1—————————
Qwen2.5-VL-72BAccess Type=Open-source, Input Frames=162026.053725.129.354.538.838.2373428.96
Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.0236.741.921.450.436.838.540.929.934.1—
InternVL2-40BAccess Type=Open-source, Input Frames=16-322026.053634.926.946.531.842.132.23439.67
LLaVA-Video-7BTraining Stage=Zero-shot2026.0335.648.51447.824.243.542.43430.6—
vsGRPO-2BTraining Stage=Reinforcement Fine-Tuned2026.0335.453.62952.743.428.130.926.818.9—
GPT-4oTraining Stage=Zero-shot2026.033446.25.343.838.23741.331.528.5—
GPT-4oAccess Type=Proprietary via API, Input Frames=16-322026.053446.25.343.838.23741.331.528.58
Qwen2.5-VL-7BAccess Type=Open-source, Input Frames=162026.053340.914.843.410.738.638.53329.89
Qwen2.5-VL-7BTraining Stage=Zero-shot2026.0332.734.519.447.640.832.824.532.529.4—