ResearchPod Summary
Long-form video question answering (QA) is often constrained by fixed input token budgets, forcing models to choose between processing fewer frames at high resolution or many frames at low resolution. Existing methods often rely on uniform sampling, which frequently misses critical evidence in long videos, or similarity-based selection, which struggles when the question lacks explicit visual keywords. This paper asks: can we develop a cost-efficient, question-adaptive frame selection strategy that improves reasoning accuracy without requiring fine-tuning of the underlying Multimodal Large Language Model (MLLM)?
The authors propose ReQuest, a three-part pipeline that optimizes frame selection for long-form video:
ReQuest consistently improves performance across major long-video benchmarks (Video-MME, MLVU, and LongVideoBench) without modifying or fine-tuning the base MLLM. The method demonstrates strong generalization, as a selector trained on one MLLM (e.g., LLaVA-Video) can be effectively transferred to others (e.g., Qwen3-VL). Crucially, the authors show that question-adaptive frame allocation is more effective than simply increasing the number of input frames, as it preserves higher visual resolution by focusing tokens on the most relevant moments rather than redundant content.
This work addresses the "needle-in-a-haystack" problem in long-form video understanding. By treating frame selection as an uncertainty-aware, adaptive process, ReQuest enables models to achieve state-of-the-art accuracy while maintaining competitive computational costs. It provides a practical, plug-and-play solution for researchers looking to enhance existing MLLMs for long-context tasks without the high overhead of full model fine-tuning.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.