DraftMind: Efficient Video Understanding in Vision-Language Models via Chain of Draft
Year: 2026
Type: conference-paper
Abstract: Large vision-language models (LVLMs) such as Qwen-VL series have demonstrated strong capabilities in visual understanding and reasoning, particularly for static images. However, extending them to video reasoning remains computationally expensive, often incurring substantial latency and token usage under conventional Chain-of-Thought (CoT) prompting. In this paper, we present a unified framework for efficient video reasoning that combines Chain of Draft (CoD) prompting with adaptive sampling, event compression, and supervised fine-tuning. CoD encourages the model to produce concise and informative intermediate reasoning steps rather than verbose rationales, making it particularly suitable for temporally complex video understanding. We adapt this paradigm to video reasoning and show that our approach achieves comparable or superior accuracy to CoT while reducing token consumption by up to 78% and inference latency by up to 65%. We evaluate our method on multiple video reasoning benchmarks, including MVBench, EgoSchema, PerceptionTest, and VideoMME, and demonstrate consistent gains across diverse tasks. Overall, our results show that concise draft-style reasoning offers an effective path toward deploying video reasoning systems in resource-constrained and real-time settings.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225414705