Latent-VC-Data
Latent Video Cache for Video Reasoning
Introduction
Latent-VC (Latent Visual Cache) introduces a recurrent latent visual cache inside the decoder of a large multimodal model to mitigate Visual Anchoring Decay in long-form video reasoning. Instead of relying on a pure read-once, generate-many pipeline, Latent-VC constructs a compact latent visual memory before answer generation, enabling the model to preserve grounding to visual evidence throughout reasoning.
This dataset contains the video training data used to train the Latent-VC-9B model, built on Qwen3.5-9B-Base and trained with two stages: Supervised Fine-Tuning (SFT) with contrastive cache alignment, followed by GRPO with vision-grounded rewards and latent grounding supervision.
Dataset Details
- Task: Grounded long-form video reasoning
- Usage: Training the Latent-VC model (SFT and GRPO stages)
- Companion model: BRZ911/Latent-VC-9B
- Paper: Latent Visual Cache for Video Reasoning
Usage
Please refer to the official GitHub repository for instructions on how to download and use this dataset for training and evaluation.
License
Please refer to the GitHub repository for license details.
Citation
If you find this work useful, please cite:
@misc{zhang2026latentvisualcachevideo,
title={Latent Visual Cache for Video Reasoning},
author={Yongheng Zhang and Zhipeng Xu and Hao Wu and Yinghui Li and Di Yin and Xing Sun and Philip S. Yu},
year={2026},
eprint={2607.02607},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.02607},
}