VIABench - Vision-Guided Interaction: leaderboard
VIABench (Nanjing University, 2026) evaluates multimodal LLMs as assistants for visually impaired people on first-person videos recorded or shared by blind users. In this multi-turn task the model guides the user through an intended interaction, with the reference history of earlier turns supplied; GPT-5 scores similarity to the reference guidance from 0 to 100, higher is better.
Metric: Vision-guided interaction score (0-100): multi-turn guidance toward an intended interaction with the environment, with the ground-truth history of earlier turns supplied, judged by GPT-5 as the semantic similarity (0 to 1, reported times 100) between the model answer and the ground-truth annotation, on first-person videos recorded or shared by visually impaired people, with each model at its default frame sampling; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 51.7 | #167 |
| 2 | Gemini 2.5 Pro | 43.5 | #266 |
| 3 | GPT-4o (2024-08-06) | 43.5 | #740 |
| 4 | Qwen 2.5 VL 7B | 27.1 | #1787 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=viabench-vision-guided-interaction · How It Works · Data refreshed daily, snapshot 2026-11-13.