Video-IFBench - Selection: leaderboard

Metric: Task-gated instruction satisfaction rate (%; TISR on selection instructions, where the model must pick the one true video-grounded condition branch; 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 December 2026. 38 models tracked.

Top models

#ModelScore
1Gemini 3 Pro59.2
2Gemini 3 Flash54.8
3Qwen 3.5 397B A17B (Thinking)44.9
4Qwen 3.5 122B A10B (Thinking)39.4
5Doubao-Seed-2.0-Pro-26021538.2
6Qwen 3.5 27B (Thinking)34
7Qwen 3.5 9B (Thinking)32
8Qwen 3.5 35B A3B (Thinking)31.5
9Gemma 4 31B (IT)25.4
10Qwen 3 VL 235B A22B (Thinking)24.9
11Gemma 4 26B A4B (IT)19.4
12Qwen 3.5 397B A17B (Non-reasoning)18.9
13Qwen 3.5 27B (Non-reasoning)18
14Qwen 3 VL 30B A3B (Thinking)17.1
15GPT-5.416.9

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