VideoMemBench - 128 Frames - Dependency Mean Accuracy: leaderboard
Video-grounded multi-turn dialogue memory under sequential self-rollout, with model answers retained in history. Turn accuracy, episodic dependency accuracy and answer-chain accuracy are separate, overlapping views; the overall score is their equal mean. 326 dialogues; sequential self-rollout with full previous model answers; uniform timestamped video frames; no audio/subtitles; complete conversation history; exact deployment checkpoints and thinking modes unreported; native MC extraction and GPT-5.4 mini open-answer grading; percentages. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 (Benchmark Sweep Run: source=VideoMemBench; frames=128; self-rollout; exact-checkpoint-and-thinking-mode=unreported) | 38.1 | |
| 2 | GPT-5.6 Sol (Benchmark Sweep Run: source=VideoMemBench; frames=128; self-rollout; exact-checkpoint-and-thinking-mode=unreported) | 37.5 | |
| 3 | Claude Sonnet 5 (Benchmark Sweep Run: source=VideoMemBench; frames=128; self-rollout; exact-checkpoint-and-thinking-mode=unreported) | 36.3 | |
| 4 | Qwen 3.5 27B (Benchmark Sweep Run: source=VideoMemBench; frames=128; self-rollout; exact-checkpoint-and-thinking-mode=unreported) | 30.3 | |
| 5 | Qwen 3 VL 235B A22B (Benchmark Sweep Run: source=VideoMemBench; frames=128; self-rollout; exact-checkpoint-and-thinking-mode=unreported) | 30.1 |
Interactive version: theaggregate.ai/benchmark?slug=videomembench-128-frames-dependency-mean-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-06.