ActivityNet-QA
Multimodal · 2019-06-06
ActivityNet-QA holds 58,000 human-written question-answer pairs about 5,800 ActivityNet web videos; this row is its 8,000-question test split. Answers are open-ended and scored as accuracy, in recent model reports by an LLM judge that compares each response with the reference answer; higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-VL-2B | 71.2 |
| GPT-5.6 Luna | 66.8 |
| Gemini 2.5 Pro | 66.7 |
| Gemini 3.5 Flash-Lite | 65.6 |
| Gemini 2.5 Flash | 65.1 |
| Qwen3.5-4B | 64.1 |
| GPT-4o (2024-05-13) | 61.9 |
| GPT-4.1 | 60.4 |
| GPT-4 Turbo | 59.5 |
| Qwen3.5-2B | 59.4 |
| Gemini 1.5 Pro (May 2024) | 57.5 |
| Gemini 1.5 Pro 002 | 57.3 |
| Gemma 4 12B IT | 57.2 |
| Gemini 1.5 Pro (February 2024 report) | 56.7 |
| Gemini 2.0 Flash | 56.4 |
| Gemini 1.5 Flash 002 | 56.2 |
| Gemini 1.5 Flash (May 2024 report) | 55.3 |
| Gemini 2.0 Flash-Lite | 55.3 |
| Gemma 4 E2B IT | 47.7 |