Research map / Video / Video-language understanding
V2P-Bench: Evaluating Video-Language Understanding with Visual Prompts for Better Human-Model Interaction
Yiming Zhao, Yu Zeng, Yukun Qi, YaoYang Liu, Xikun Bao, Lin Chen, Zehui Chen, Qing Miao, Chenxi Liu, Jie Zhao, Feng Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| convert visual prompts to text prompts (ablation) | Gemini-2.5-Pro · V2P-Bench | accuracy (%): 69.8 -> 54.7 | Hurt |
| convert visual prompts to text prompts (ablation) | GPT-4o · V2P-Bench | accuracy (%): 65.4 -> 53.0 | Hurt |
| use visual prompts instead of text prompts for users (ablation) | user experience study · user study (20 volunteers) | completion time (s): 25.2 -> 18.1 | Improved |
| use visual prompts instead of text prompts for users (ablation) | user experience study · user study (20 volunteers) | user satisfaction: 5.3 -> 7.5 | Improved |
| fine-tune LLaVA-NeXT on INST-IT visual prompt data | LLaVA-NeXT-7B · V2P-Bench | average accuracy (%): 46.0 -> 46.3 | No clear effect |
| scale model from 7B to 72B (ablation) | Qwen2.5-VL · V2P-Bench | average accuracy (%): 48.1 -> 59.8 | Improved |
| use doodle-style visual prompts instead of standard shapes (ablation) | Qwen2.5-VL-7B and MiMo-VL-7B · V2P-Bench 217-instance subset | accuracy (%): decrease of 0.7% and 0.8% | Hurt |
| use arrow instead of rectangle visual prompts (ablation) | Qwen2.5-VL-7B · V2P-Bench 217-instance subset | accuracy (%): 47.3 -> 43.6 | Hurt |
| increase video length from short/medium to long (ablation) | o1 · V2P-Bench | accuracy (%): 83.9 -> 60.4 | Hurt |
| reduce sampled frames from 128 to 4 (ablation) | Qwen2.5-VL-7B · V2P-Bench | Hack Ratio (%): 8.0 -> 18.7 | Hurt |
| increase sampled frames from 1 to 64 (ablation) | Qwen2.5-VL-7B · V2P-Bench short videos, Temporal Understanding | accuracy (%): 28.6 -> 40.1 | Improved |
| increase sampled frames from 1 to 64 (ablation) | Qwen2.5-VL-7B · V2P-Bench long videos, Basic Perception | accuracy (%): 47.6 -> 43.1 | Hurt |
Closest papers by meaning
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.92)
- VP-Bench: A Comprehensive Benchmark for Visual Prompting in Multimodal Large Language Models (AAAI 2026 · similarity 0.92)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.91)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.91)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.90)
- ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
