Mid-Training Language Models on Raw Video

2026-10-07

AI Summary

This study addresses the challenge of leveraging unlabeled video data for multimodal large language models by exploring the feasibility of using raw videos as intermediate training data. Methodologically, video frames are encoded into continuous visual tokens, enabling a Qwen3-1.7B base model to undergo intermediate training via self-supervised next-token prediction, followed by unified instruction tuning. This work provides the first demonstration that purely self-supervised intermediate training on raw videos outperforms approaches relying on automatic captioning, enhancing perceptual capabilities while preserving linguistic performance without requiring text-based losses. Experimental results indicate that the proposed method yields average improvements of 2.9 and 5.1 points on video and image benchmarks, respectively, while maintaining textual performance at 48.9 points, thereby validating its effectiveness and generalizability.

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.

Problem

• multimodal large language models
• raw video
• mid-training
• self-supervised learning
• next visual token prediction

Innovation

• raw video mid-training
• next visual token prediction
• self-supervised learning
• multimodal large language models
• continuous visual tokens