VIABench is a comprehensive egocentric video benchmark for evaluating multimodal large language models in real-world visual impairment assistance scenarios. Collected from videos recorded or shared by blind individuals, VIABench contains 761 videos, 46.9 hours of footage, and 14,526 manually curated annotations across three core tasks: Proactive Reminder, Visual Question Answering, and Vision-Guided Interaction.
The benchmark is designed to test whether video MLLMs can provide practical assistance beyond passive video understanding, including real-time reminders, environment-aware question answering, and interactive guidance for user goals.
The task instructions and VQA evaluation prompt used by VIABench are available in prompts. The module also provides a standalone prompt builder:
from prompts import VQA_JUDGE_PROMPT, build_task_prompt
reminder_prompt = build_task_prompt("OA")
vqa_prompt = build_task_prompt("VQA", question="What is in front of me?")- 2026/08/13: 🔥🔥🔥 Release the task prompts and VQA evaluation prompt used by VIABench.
- 2026/07/17: 🔥🔥🔥 Release the paper of VIABench, a comprehensive video benchmark for evaluating MLLMs in visually impaired assistance scenarios.
- 2026/07/17: 🔥🔥🔥 Release the VIABench dataset.