Less from More: Reinforcing Sparse Video Reasoning from Dense References
- 类型：arxiv
- 标识：2610.10893
- 链接：https://arxiv.org/abs/2610.10893
- 主分类：multimodal
- 形态：method
- TLDR：Video-language models commonly assume that more temporal observations lead to more reliable reasoning. We question this assumption and argue that the key challenge is not merely processing more video frames efficiently, but learning to reason reliably under limited temporal evidence. We propose SAVER, a dense-to-sparse post-training framework that uses dense video views as training-time references for sparse-frame inference. During reinforcement post-training, paired dense and sparse views are optimized with grounding rewards and a reliability-gated reference reward, encouraging sparse view pr
- 副分类：engineering
- 待LLM分类：否
- 标题中文：以少胜多：基于密集参考强化稀疏视频推理
- TLDR中文：视频-语言模型通常假设更多时序观测会带来更可靠的推理。我们质疑该假设，认为核心挑战并非仅是高效处理更多视频帧，而是在有限时序证据下学会可靠推理。我们提出 SAVER，一个由密集到稀疏的后训练框架，将密集视频视图作为稀疏帧推理的训练时参考。在强化后训练阶段，配对的密集与稀疏视图通过 grounding 奖励与可靠性门控参考奖励进行优化，鼓励稀疏视图……
- 来源文件：
- /inbox/tom/_candidates/2026-10-12-agent-rag-longcontext-candidates.json