Video-MemComp - Offline Aggregate: 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. Offline Aggregate. Metric: Accuracy (%).

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 19 models tracked.

Top models

#ModelScoreOverall rank
1Video-MemComp GPT4-o [frames 1fps; KV –]71.9
2Video-MemComp Qwen2-VL-7B + Video-MemComp [frames 768; KV 6K]65.5
3Video-MemComp Qwen2-VL-7B + StreamMem [frames 0.5/0.2fps; KV 6K]65.1
4Video-MemComp Qwen2-VL-7B [frames 768; KV 100K]64.9
5Video-MemComp Qwen2-VL-7B + InfiniPot-V [frames 768; KV 6K]64.7
6Video-MemComp LongVU 7B [frames 1fps; KV 8K]64.5
7Video-MemComp Qwen2.5-VL-3B + Video-MemComp [frames 768; KV 6K]64.1
8Video-MemComp LLaVA-OneVision-7B + StreamMem [frames 0.5/0.2 fps; KV 6K]63.1
9Video-MemComp LLaVA-OneVision-7B + Video-MemComp [frames 0.5/0.2 fps; KV 6K]62.5
10Video-MemComp InternVL3.5-2B + Video-MemComp [frames 64/100; KV 6K]62.4

Interactive version: theaggregate.ai/benchmark?slug=video-memcomp-offline-aggregate · How It Works · Data refreshed daily, snapshot 2026-10-27.