Video-IFBench: leaderboard

Metric: Task-gated instruction satisfaction rate (%; TISR over all four instruction types; checklist-based instruction following on 1.5K video samples with 39 semantic and format constraint types; a response counts only if it addresses the active task (the correct branch for conditional instructions) and satisfies every checklist item, judged by Qwen3.5-397B-A17B-Instruct or verified programmatically; Gemini and Doubao get video at 1 fps, GPT-5.4 50 frames, vision-only models frames plus timestamped subtitles). Source: arxiv.org. Saturation forecast: Around June 2027. 38 models tracked.

Top models

#ModelScore
1Gemini 3 Pro54.5
2Gemini 3 Flash49.5
3Qwen 3.5 397B A17B (Thinking)46.1
4Qwen 3.5 122B A10B (Thinking)41.7
5Doubao-Seed-2.0-Pro-26021540.8
6Qwen 3.5 27B (Thinking)37.5
7Qwen 3.5 35B A3B (Thinking)37.2
8Gemma 4 31B (IT)35.4
9Qwen 3 VL 235B A22B (Thinking)33.5
10Qwen 3.5 9B (Thinking)32.2
11Qwen 3.5 397B A17B (Non-reasoning)30.4
12GPT-5.430
13Gemma 4 26B A4B (IT)29.7
14Qwen 3 VL 235B A22B Instruct25.8
15Qwen 3.5 122B A10B (Non-reasoning)25.3

Interactive version: theaggregate.ai/benchmark?slug=video-ifbench · How It Works · Data refreshed daily, snapshot 2026-09-26.