用视觉缓存解决视频推理中信息衰减问题，让模型更精准记忆画面。
Latent Visual Cache for Video Reasoning
- 在解码器中加入可循环的隐空间视觉缓存，持续保存关键视觉信息。
- 在6个视频基准上表现优于强基线，长视频和依赖视觉的任务提升明显。
- 回答更短但更准，适合需要精炼推理的视频理解场景。
视频推理要求大模型始终基于密集视觉证据，但现有系统多采用‘读一次、生成多次’模式，导致推理过程中视觉锚定逐渐减弱，称为视觉锚定衰减。为解决此问题，我们提出隐空间视频缓存（Latent-VC），一种插入解码器的循环隐空间缓存，用于在整个推理过程中保持紧凑的视觉记忆。该缓存通过监督对比对齐与视觉锚定的GRPO训练，并结合隐空间锚定奖励，同时通过原生解码器隐藏状态确保训练与推理的一致性。基于Qwen3.5-9B构建，Latent-VC在六个视频基准上均优于强基线，尤其在依赖视觉锚定和长视频任务中表现突出。此外，其响应显著更短且准确率更高，表明隐空间缓存通过保留视觉证据而非依赖冗长文本链来提升推理性能。
原文摘要 · Abstract (English)
Video reasoning requires Large Multimodal Models (LMMs) to remain grounded in dense evidence, yet existing systems largely adopt "read-once, generate-many" paradigm, in which visual grounding weakens during generation. This phenomenon has been widely observed and is known as Visual Anchoring Decay. To fill this gap, we introduce Latent Video Cache (Latent-VC), a recurrent latent visual cache inserted into the decoder to preserve compact visual memories throughout reasoning. The cache is trained with supervised contrastive cache alignment and vision-grounded GRPO with a latent grounding reward, while maintaining strict train-inference alignment through native decoder hidden states. Built on Qwen3.5-9B, Latent-VC consistently outperforms strong CoT and SFT+GRPO baselines across six video benchmarks, with especially clear gains on grounding-intensive and long-video tasks. In addition, it also achieves higher accuracy with substantially shorter responses, suggesting that latent visual caching improves video reasoning by preserving visual evidence rather than relying on longer textual chains.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。