TempCore Study - Query-only MaxProb - NExTQA: leaderboard
TempCore paper v3,Query-only MaxProb, NExTQA native accuracy percent;32frames,zero-shot multiplechoice. Query-only MaxProb ranks10fps candidates withSigLIP2-400M and per-second diversity, without answer-label access. Uniform Intersection uses the complete-results window-oracle intersection but its own uniform selector, not oracle-selected windows. Full source benchmark cohorts, not the released2560-example TempCore subset. QA-mode and window-oracle cells are held; no aggregate across unlike datasets.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3.5-8B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 79.5 | |
| 2 | LLaVA-OV-7B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 78 | |
| 3 | Qwen3-VL-8B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 70.9 | |
| 4 | SmolVLM2-2.2B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 60.8 | |
| 5 | LLaVA-OV-0.5B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 57.2 | |
| 6 | SmolVLM2-500M (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 46.9 | |
| 7 | VidLLaMA3-2B (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 46.9 | |
| 8 | SmolVLM2-256M (TempCore Study Run: selector=Query-only MaxProb;frames=32;checkpoint_variant=unreported;precision=unreported) | 27.6 |
Interactive version: theaggregate.ai/benchmark?slug=tempcore-study-query-only-maxprob-nextqa · How It Works · Data refreshed daily, snapshot 2026-11-12.