IPIBench - Task Modification: leaderboard

IPIBench (2026) streams videos to multimodal LLMs at one frame per second and checks whether they respond proactively at the right moment, manage monitoring tasks the user cancels or changes, and handle reactive questions interleaved with proactive tasks. This board is accuracy on task modification (update the monitoring target after the user changes it); percent, higher is better.

Metric: Accuracy (%) on task modification (update the monitoring target after the user changes it), on IPIBench (1,831 videos, 3,738 interaction instances from egocentric and exocentric datasets) under a unified streaming protocol at 1 FPS (offline models see a sliding window of the latest 16 frames); a proactive response counts only if the first trigger falls within one second of the ground-truth moment, multi-turn tasks only if every required trigger and answer is correct, reactive answers by bidirectional substring matching against up to 20 reference answers; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 15 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4o31.88#751
2Qwen 3.5 Plus31.54#242
3Qwen 3.5 9B (Harness: AURA2 IPIBench offline30frames; Thinking: On)25.17
4GPT-5.424.5#135
5GLM-4.6V24.16#710
6Gemini 3 Pro21.84#145
7Gemini 2.5 Pro20.81#276
8Qwen 3 VL 8B17.45#917
9MOSS-VL-Realtime9.4
10Qwen 3.5 9B (Harness: AURA2 IPIBench offline30frames; Thinking: Off)7.38
11InternVL3-8B6.04#1782
12JoyAI-VL-Interaction2.68
13VideoLLM-online-8B1.68
14LLaVA-OneVision-7B1.01#2981
15Flash-VStream-7B0

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ipibench-task-modification · How It Works · Data refreshed daily, snapshot 2026-11-21.