Video-MemComp - MLVU KV 3K: leaderboard
Video-MemComp v1: streaming evaluation of official development sets. Table 1 mixes reproduced starred rows and results from prior papers; frame sampling and average KV budget differ by system and remain in row identity. Backbone checkpoint revision and decoding are not fully specified; source-qualified systems are retained. This is video QA under cache compression, not an unmodified backbone score. Appendix E.1: Qwen2-VL-7B, MLVU dev, matched global KV capacity equivalent to 3K per layer. Full-KV row is a single reference spanning columns and is not replicated into budgets. MLVU KV 3K. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Video-MemComp Qwen2-VL-7B Video-MemComp | 65.48 | |
| 2 | Video-MemComp Qwen2-VL-7B InfiniPot-V | 64.7 | |
| 3 | Video-MemComp Qwen2-VL-7B SnapKV | 58.99 | |
| 4 | Video-MemComp Qwen2-VL-7B InfiniPot | 58.36 | |
| 5 | Video-MemComp Qwen2-VL-7B Uniform | 55.59 | |
| 6 | Video-MemComp Qwen2-VL-7B FastV | 55.01 |
Interactive version: theaggregate.ai/benchmark?slug=video-memcomp-mlvu-kv-3k · How It Works · Data refreshed daily, snapshot 2026-10-27.