Mid training language models using raw video improves vision without losing text skills
Mid-Training Language Models on Raw Video
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
Summary
Most large language models learn from images paired with text or from labeled videos. This paper explores whether videos with no captions or text can help improve a language model’s understanding of visuals. The researchers encoded video frames as visual tokens and trained a language model to predict the next token. The model trained this way performed better on video and image tasks while keeping its ability to understand text. They found that predicting visual tokens works better than using generated captions, making the training simpler and fully self-supervised.
What this means in practice
- •For computer vision engineers: Train multimodal models on unlabeled video data to improve image and video understanding without extra text annotations.
- •For ai product teams: Develop applications that improve visual recognition performance without compromising natural language capabilities by mid-training on raw video.$Commercial implications: Enables building better visual AI features in products by using abundant unlabeled video data for training.
Authors
Jaedong Hwang, Xiaoqian Shen, Ernie Chang, Changsheng Zhao, Chong Zhou, Saksham Suri, Qi Qian, Zechun Liu, Lemeng Wu, Qinsi Wang, Raghuraman Krishnamoorthi, Wei Wen
Abstract
Multimodal large language models learn mostly from paired image-text data or annotated video, and raw web video is rarely used to further train an existing language model. We study whether raw video, with no captions and no text loss, can serve as mid-training data for a pretrained language model. Frames are encoded into continuous visual tokens, and the language model learns to predict the next visual token. We mid-train Qwen3-1.7B on raw clips from YT-Temporal-1B and then apply the same image-text instruction tuning to it and to the model without mid-training, so that the two differ only in mid-training. The mid-trained model scores 2.9 points higher on average across four video benchmarks and 5.1 points higher across ten image benchmarks, spanning perception, document, and chart tasks. Text performance is preserved even though mid-training includes no text, with an average of 48.9 across 14 text benchmarks compared with 48.0 for the model without mid-training. Analyses across training show that the image and video gains emerge within 30% of training and plateau thereafter, varying by less than 0.5 points. Predicting captions fails to outperform next-visual-token prediction, demonstrating that video mid-training can remain purely self-supervised without the computational overhead or labeling noise of automated captioning.