Beyond Language Modeling: An Exploration of Multimodal Pretraining

• URL: http://arxiv.org/abs/2603.03276v1
• Date: Tue, 03 Mar 2026 18:58:00 GMT
• Title: Beyond Language Modeling: An Exploration of Multimodal Pretraining
• Authors: Shengbang Tong, David Fan, John Nguyen, Ellis Brown, Gaoyue Zhou, Shengyi Qian, Boyang Zheng, Théophane Vallaeys, Junlin Han, Rob Fergus, Naila Murray, Marjan Ghazvininejad, Mike Lewis, Nicolas Ballas, Amir Bar, Michael Rabbat, Jakob Verbeek, Luke Zettlemoyer, Koustuv Sinha, Yann LeCun, Saining Xie,
• Abstract summary: We provide empirical clarity through controlled, from-scratch pretraining experiments.<n>We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision.<n>We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language.
• Score: 125.34714978184638
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision, to train on diverse data including text, video, image-text pairs, and even action-conditioned video. Our experiments yield four key insights: (i) Representation Autoencoder (RAE) provides an optimal unified visual representation by excelling at both visual understanding and generation; (ii) visual and language data are complementary and yield synergy for downstream capabilities; (iii) unified multimodal pretraining leads naturally to world modeling, with capabilities emerging from general training; and (iv) Mixture-of-Experts (MoE) enables efficient and effective multimodal scaling while naturally inducing modality specialization. Through IsoFLOP analysis, we compute scaling laws for both modalities and uncover a scaling asymmetry: vision is significantly more data-hungry than language. We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language while accommodating the data-intensive nature of vision, paving the way for truly unified multimodal models.

Related papers

• Towards Understanding Multimodal Fine-Tuning: Spatial Features [25.349396112139214]
  Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model.<n>We present the first mechanistic analysis of VLM adaptation using stage-wise model diffing.
  arXiv  Detail & Related papers  (2026-02-06T18:48:18Z)
• Simple o3: Towards Interleaved Vision-Language Reasoning [38.46230601239066]
  We propose Simple o3, an end-to-end framework that integrates dynamic tool interactions into interleaved vision-language reasoning.<n>Our approach features a scalable data synthesis pipeline that generates high-quality interleaved vision-language reasoning chains.<n> Experimental results demonstrate Simple o3's superior performance on diverse benchmarks, outperforming existing approaches.
  arXiv  Detail & Related papers  (2025-08-16T17:15:39Z)
• Aya Vision: Advancing the Frontier of Multilingual Multimodality [15.981889066681424]
  We develop a synthetic annotation framework that curates high-quality, diverse multilingual multimodal instruction data.<n>We also propose a cross-modal model merging technique that mitigates catastrophic forgetting.<n>Our work advances multilingual progress on the multi-modal frontier, and provides insights into techniques that effectively bend the need for compute.
  arXiv  Detail & Related papers  (2025-05-13T17:03:48Z)
• EMMA: Efficient Visual Alignment in Multi-Modal LLMs [56.03417732498859]
  EMMA is a lightweight cross-modality module designed to efficiently fuse visual and textual encodings.<n>EMMA boosts performance across multiple tasks by up to 9.3% while significantly improving robustness against hallucinations.
  arXiv  Detail & Related papers  (2024-10-02T23:00:31Z)
• MIO: A Foundation Model on Multimodal Tokens [75.45159462076937]
  We introduce MIO, a novel foundation model built on multimodal tokens.<n>MIO is capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner.
  arXiv  Detail & Related papers  (2024-09-26T09:57:16Z)
• IAA: Inner-Adaptor Architecture Empowers Frozen Large Language Model with Multimodal Capabilities [4.269326314400742]
  We introduce the Inner-Adaptor Architecture for multimodal large language models (MLLMs)<n>The architecture incorporates multiple multimodal adaptors at varying depths within the large language model to facilitate direct interaction with the inherently text-oriented transformer layers.<n>Unlike previous approaches of freezing language models that require large-scale aligned data, our proposed architecture is able to achieve superior performance on small-scale datasets.
  arXiv  Detail & Related papers  (2024-08-23T08:10:13Z)
• StableLLaVA: Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data [129.92449761766025]
  We propose a novel data collection methodology that synchronously synthesizes images and dialogues for visual instruction tuning. This approach harnesses the power of generative models, marrying the abilities of ChatGPT and text-to-image generative models. Our research includes comprehensive experiments conducted on various datasets.
  arXiv  Detail & Related papers  (2023-08-20T12:43:52Z)
• UniDiff: Advancing Vision-Language Models with Generative and Discriminative Learning [86.91893533388628]
  This paper presents UniDiff, a unified multi-modal model that integrates image-text contrastive learning (ITC), text-conditioned image synthesis learning (IS), and reciprocal semantic consistency modeling (RSC) UniDiff demonstrates versatility in both multi-modal understanding and generative tasks.
  arXiv  Detail & Related papers  (2023-06-01T15:39:38Z)
• PaLM-E: An Embodied Multimodal Language Model [101.29116156731762]
  We propose embodied language models to incorporate real-world continuous sensor modalities into language models. We train these encodings end-to-end, in conjunction with a pre-trained large language model, for multiple embodied tasks. Our largest model, PaLM-E-562B with 562B parameters, is a visual-language generalist with state-of-the-art performance on OK-VQA.
  arXiv  Detail & Related papers  (2023-03-06T18:58:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.