DK-VLM VSI-Bench - Object Size: leaderboard
DK-VLM v1; LoRA/geometric attention fine-tunes remain distinct from Qwen2.5VL7BInstruct and Qwen3VL8BInstruct backbones. Frozen VGGT-Omega 1B frontend; official shared six-benchmark GeoThinker evaluation settings. Comparison baseline scores are source-reported, starred GeoThinker values reproduced by authors; no automatic primary-board fusion. Metric: Mean relative accuracy (%). Higher is better. Source: https://arxiv.org/html/2610.22356v1
Metric: Mean relative accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DK-VLM-8B | 77.82 | |
| 2 | Q-GeoMem-7B | 76.5 | |
| 3 | Qwen3-VL-8B (DK-VLM paper baseline) | 76.3 | |
| 4 | GeoThinker-8B | 76.23 | |
| 5 | Cambrian-S-7B | 74.9 | |
| 6 | GeoThinker-7B | 74.34 | |
| 7 | DK-VLM-7B | 74.01 | |
| 8 | VLM-3R-7B | 69.2 | |
| 9 | SpatialStack-5B | 69.1 | |
| 10 | VG-LLM-8B | 69 |
Interactive version: theaggregate.ai/benchmark?slug=dk-vlm-vsi-bench-object-size · How It Works · Data refreshed daily, snapshot 2026-10-17.