Latent Visual Cache for Video Reasoning
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 o...
| Authors | Yongheng Zhang, Zhipeng Xu, Hao Wu et al. |
|---|---|
| Primary category | cs.CV |
| Published | 1 July 2026 |
| Last updated at source | 1 July 2026 |
Connected in the knowledge graph
- BRZ911/Latent-VC-DataDataset · cited by
Models and datasets referencing this paper
- BRZ911/Latent-VC-DataDataset
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.