Wearable AI assistants must track objects across egocentric video to help users in physical spaces. Current vision-language models (VLMs) struggle with spatial questions that require combining observations from different moments. An overall score does not reveal whether their errors come from perceiving the scene or reasoning about it. We introduce SPLIT, a training-free harness that reconstructs one metric 3-D point cloud from a video and gives a planner VLM tools to query it. Frozen reconstruction and grounding models locate objects and measure their geometry, and the planner VLM combines those measurements using Python. We train no model parameters on the evaluated benchmarks. SPLIT improves accuracy on VSIBench, VSTIBench, and ReVSI. On VSTIBench, it raises accuracy from 57.7% to 78.3%, eliminating 49% of the baseline VLM's errors; across all three benchmarks, the largest gains occur when questions relate objects seen in different frames. SPLIT achieves the best published training-free score on VSIBench and exceeds published scores on VSTIBench and ReVSI. To measure how much error comes from perception, we rerun SPLIT with ground-truth measurements in place of the tools' estimates. On the VSTIBench analysis subset, the baseline VLM scores 59.0%, SPLIT scores 79.7%, and SPLIT + GT reaches 95.7%. This gain shows that inaccurate perception remains a major source of error.
| Method | VSIBench | VSTIBench | ReVSI |
|---|---|---|---|
| Proprietary | |||
| GPT-4o | 34.0 | 38.2 | – |
| Gemini-2.5 Pro | 51.5 | 42.4 | – |
| GPT-5.2 | 49.2 | – | 50.9 |
| Gemini 3 Flash | 55.9 | – | 57.6 |
| Gemini 3 Pro | 60.5 | – | 60.9 |
| Open-source | |||
| Qwen3-VL-8B | 57.4 | – | 49.1 |
| InternVL3.5-38B | 60.8 | – | 54.1 |
| LLaVA-Video-72B | 39.7 | 44.0 | 37.8 |
| Supervised spatial specialists | |||
| VLM-3R-7B | 60.9 | 58.8 | 50.1 |
| Cambrian-S-7B | 67.5 | – | 49.1 |
| GeoThinker-8B | 72.6 | 67.4 | – |
| Cambrian-P | 73.7 | 68.9 | – |
| SSR-3D | 73.9 | 44.8 | – |
| Baseline VLM (Gemini 3.1 Pro) | 60.4 | 57.7 | 63.8 |
| SPLIT w/ Gemini 3.1 Pro (Ours) | 66.9 (+6.5) | 78.3 (+20.6) | 71.5 (+7.7) |
| SPLIT (Ours) + GT | 91.8† | 95.7† | 91.0† |
Bold and underlining mark the best and second-best full-benchmark scores. † Ground-truth cells use the corresponding analysis subsets; VSIBench and ReVSI counting are oracle ceilings.
† SPLIT (Ours) + GT is measured on each benchmark's analysis subset; VSIBench and ReVSI counting are oracle ceilings.
VSIBench
| Numerical | Multiple choice | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | Train | Avg. | Obj. Count | Abs. Dist. | Obj. Size | Room Size |
Rel. Dist. | Rel. Dir. | Route Plan | Appr. Order |
| Human | – | 79.2 | 94.3 | 47.0 | 60.4 | 45.9 | 94.7 | 95.8 | 95.8 | 100.0 |
| General-purpose VLMs | ||||||||||
| GPT-4o | TF | 34.0 | 46.2 | 5.3 | 43.8 | 38.2 | 37.0 | 41.3 | 31.5 | 28.5 |
| Gemini-1.5 Pro | TF | 45.4 | 56.2 | 30.9 | 64.1 | 43.6 | 51.3 | 46.3 | 36.0 | 34.6 |
| Gemini-2.5 Pro | TF | 51.5 | 43.8 | 34.9 | 64.3 | 42.8 | 61.1 | 47.8 | 45.9 | 71.3 |
| Supervised spatial specialists | ||||||||||
| Think-with-Spatial-Code | RL | 54.9 | 58.3 | 39.0 | 73.0 | 52.4 | 57.8 | 55.9 | 38.7 | 63.9 |
| + 2D box | RL | 60.1 | 95.2 | 60.7 | 50.8 | 33.1 | 87.1 | 62.0 | 32.5 | 59.0 |
| VLM-3R-7B | SFT | 60.9 | 70.2 | 49.4 | 69.2 | 67.1 | 65.4 | 80.5 | 45.4 | 40.1 |
| Cambrian-S-7B | SFT | 67.5 | 73.2 | 50.5 | 74.9 | 72.2 | 71.1 | 76.2 | 41.8 | 80.1 |
| SpaceMind++ (InternVL3-8B) | SFT | 73.4 | 74.5 | 62.4 | 77.9 | 76.9 | 73.5 | 89.7 | 48.2 | 84.1 |
| Cambrian-P (Qwen2.5-7B) | SFT | 73.7 | 74.9 | 60.1 | 76.0 | 76.9 | 74.8 | 89.5 | 52.6 | 85.0 |
| SSR-3D (openPangu-VL-7B) | SFT | 73.9 | 70.5 | 67.1 | 76.0 | 79.5 | 71.3 | 93.4 | 48.5 | 85.0 |
| Training-free methods | ||||||||||
| ViSRA (Qwen2.5-VL-7B)† | TF | 50.3 | 46.4 | 27.8 | 60.1 | 47.7 | 51.3 | 71.7 | 43.8 | 53.7 |
| TRACE (Gemini 3 Pro) | TF | 57.8 | 47.6 | 38.8 | 73.9 | 45.6 | 63.9 | 61.7 | 58.0 | 73.0 |
| Full benchmark | ||||||||||
| Baseline VLM | TF | 60.4 | 51.2 | 43.6 | 73.3 | 53.1 | 69.2 | 62.4 | 58.8 | 71.7 |
| SPLIT (Ours) | TF | 66.9 | 50.0 | 58.3 | 66.1 | 56.3 | 69.9 | 90.0 | 73.7 | 70.7 |
| Δ | +6.5 | -1.2 | +14.7 | -7.2 | +3.2 | +0.7 | +27.6 | +15.0 | -1.0 | |
| Debiased subset | ||||||||||
| Baseline VLM | TF | 57.5 | 46.4 | 41.2 | 62.9 | 54.6 | 69.0 | 62.3 | 51.8 | 71.7 |
| SPLIT (Ours) | TF | 65.8 | 50.9 | 56.4 | 60.2 | 57.5 | 71.6 | 86.7 | 71.1 | 72.3 |
| Δ | +8.3 | +4.5 | +15.2 | -2.7 | +2.8 | +2.6 | +24.4 | +19.3 | +0.6 | |
| Ground-truth diagnostic‡ | ||||||||||
| SPLIT (Ours) + GT | TF | 91.8 | 95.4 | 92.4 | 97.0 | 92.4 | 94.0 | 99.3 | 64.0 | 100.0 |
Bold and underlining mark the best and second-best full-benchmark model scores. “TF” is training-free, “RL” reinforcement learning, and “SFT” supervised fine-tuning. † Subset result reported by ViSRA. ‡ Ground-truth diagnostic; the analysis subset is described in the paper, counting is an oracle ceiling, and appearance order uses a fixed 50-question cohort shared by every arm.
VSTIBench
| Method | Train | Avg. | Cam-Obj Abs (MRA) | Cam Disp (MRA) | Cam Mov Dir |
Obj-Obj Rel Pos | Cam-Obj Rel Dist |
|---|---|---|---|---|---|---|---|
| Published comparisons* | |||||||
| VLM-3R-7B | SFT | 58.8 | 39.4 | 39.6 | 60.6 | 86.5 | 68.6 |
| GeoThinker-8B | SFT | 67.4 | 38.4 | 45.8 | 84.2 | 93.6 | 75.2 |
| Cambrian-P | SFT | 68.9 | 42.5 | 46.6 | 87.7 | 94.3 | 73.2 |
| Baseline VLM | TF | 57.7 | 49.0 | 18.5 | 45.3 | 93.5 | 82.0 |
| SPLIT (Ours) | TF | 78.3 | 58.2 | 60.2 | 96.7 | 90.4 | 85.7 |
| Δ | +20.6 | +9.3 | +41.7 | +51.4 | -3.2 | +3.7 | |
| Ground-truth diagnostic† | |||||||
| Baseline VLM | TF | 59.0 | 55.0 | 20.6 | 44.0 | 98.7 | 76.7 |
| SPLIT (Ours) | TF | 79.7 | 63.8 | 56.2 | 96.0 | 97.3 | 85.3 |
| SPLIT (Ours) + GT | TF | 95.7 | 87.8 | 97.6 | 98.0 | 98.0 | 97.3 |
Bold and underlining mark the best and second-best full-benchmark scores. * Published comparison rows use original-release labels. We identified a camera-center bug in camera displacement, camera–object distance (absolute and relative) labels and notified the VSTIBench authors; they have corrected the release, public erratum. Our rows use the corrected release labels. † Ground-truth diagnostic; the analysis subset is described in the paper.
ReVSI
| Numerical | Multiple choice | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | Train | Frames | Avg. | Obj. Count | Abs. Dist. | Obj. Size | Room Size |
Rel. Dist. | Rel. Dir. | Route Plan |
| GPT-5.2 | TF | 64 | 50.9 | 56.2 | 41.5 | 73.9 | 63.0 | 48.4 | 34.9 | 38.2 |
| InternVL3.5-38B | TF | 64 | 54.1 | 43.8 | 60.6 | 70.2 | 58.4 | 57.4 | 45.9 | 42.7 |
| VLM-3R-7B | SFT | 32 | 50.1 | 41.6 | 61.6 | 64.8 | 52.5 | 46.5 | 49.5 | 34.1 |
| Baseline VLM | TF | 32 | 63.8 | 64.6 | 59.6 | 76.8 | 61.9 | 67.0 | 58.0 | 58.5 |
| SPLIT (Ours) | TF | 32 | 71.5 | 58.8 | 66.8 | 64.2 | 50.3 | 84.2 | 93.2 | 82.9 |
| Δ | +7.7 | -5.8 | +7.2 | -12.5 | -11.6 | +17.2 | +35.3 | +24.4 | ||
| Ground-truth diagnostic† | ||||||||||
| SPLIT (Ours) + GT | TF | 32 | 91.0 | 86.7 | 98.2 | 85.0 | 79.7 | 97.9 | 98.3 | 91.1 |
Bold and underlining mark the best and second-best full-benchmark scores. † Ground-truth diagnostic; the analysis subset is described in the paper, and counting is an oracle ceiling.
† SPLIT (Ours) + GT is measured on each benchmark's analysis subset; counting is an oracle ceiling on VSIBench and ReVSI. On VSIBench, appearance order uses a 50-question answerable cohort. On ReVSI, the baseline and SPLIT bars are full-benchmark scores and the ground-truth bars use the 981-question subset that excludes 18 route questions without complete annotations.
With ground-truth measurements in place of the tools’ estimates, SPLIT rises from 78.3 to 95.7 on VSTIBench: perception, not reasoning, is the bottleneck.
SPLIT improves VSIBench, VSTIBench, and ReVSI by 6.5, 20.6, and 7.7 points over the same planner VLM without tools.
@misc{yeung2026split,
title = {Perception, Not Reasoning, Limits Video Spatial Understanding},
author = {Yeung, Jacob and Goyal, Mohit and Dubost, Florian and Vondran Jr., Gary L.
and Tombari, Federico and Ramanan, Deva and Tarr, Michael J.},
year = {2026}
}