VG-LLM Release v3 - VSIRelDirection: leaderboard
VGLLM originalrelease v3; QA checkpoint uses Qwen2.5VL7BInstruct plus VGGT geometry encoder, greedy32uniformframes VSI-Bench. Different3Dscenecheckpoint and later S2/S1+S2 trainedablations are distinct; no generic8Balias for them. NumericalVSI uses mean-relative-accuracy rather than exact accuracy. Metric: Accuracy (%). Higher is better. Source: https://arxiv.org/html/2505.24625v3
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 46.3 | #443 |
| 2 | GPT-4o | 41.3 | #339 |
| 3 | Gemini 1.5 Flash | 41 | #615 |
| 4 | InternVL2-8B | 30.7 | #813 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=vg-llm-release-v3-vsireldirection · How It Works · Data refreshed daily, snapshot 2026-10-17.