Spatial-SSRL4B Release - Spatial understanding: leaderboard
Spatial relation/geometry understanding aggregate under sourceprompting; finalSSRL usesCoT. OriginalNov24 Spatial-SSRL Qwen3VL4B release. CoT checkmarks are promptingprotocol in thisfigure, not proof of a separatelytrained QwenThinkingcheckpoint. FinalSSRL4B is exact releasedQwen3VL4B-trained model. Spatial understanding aggregate andgeneralVQA aggregate kept distinct; plusvaluesareimprovements, notadditionalobservations. Metric: Average score (%). Higher is better. Exact printed figure values; https://huggingface.co/internlm/Spatial-SSRL-Qwen3VL-4B/resolve/main/assets/exp_result_new3.png
Metric: Average score (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Spatial-SSRL-Qwen3VL-4B | 61.43 | |
| 2 | Qwen3-VL-4B (Spatial-SSRL baseline with CoT) | 60.23 | |
| 3 | Qwen3-VL-4B (Spatial-SSRL baseline without CoT) | 60.14 |
Interactive version: theaggregate.ai/benchmark?slug=spatial-ssrl4b-release-spatial-understanding · How It Works · Data refreshed daily, snapshot 2026-10-17.