Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs

Despite the impressive results achieved by multimodal large language models (MLLMs), their training typically relies on jointly curated multimodal data, requiring substantial human effort to construct multi-way aligned datasets and thereby limiting scalability across domains. In this work, we explore training MLLMs by only leveraging multiple paired modalities as a surrogate for the full joint multimodal distribution. Specifically, we first provide a theoretical analysis of the conditions under which the representations are identifiable with only observing pairwise modalities. Building on this analysis, we propose a representation learning framework for aligning latent representations across modalities using only pairwise data. The framework consists of two stages: latent representation alignment and cross-modal recomposition. Specifically, in the first stage, we learn the shared latent space across modalities by both self-modal reconstruction and pair-wise contrastive learning. We also incorporate an inductive bias in the contrastive learning process by partially aligning and minimal latent specification. In stage two, we integrate the encoder of newly introduced modalities with the decoders of the pre-trained modalities to facilitate cross-modal transfer and generation. We evaluate our method by newly adding 3D point clouds and tactile modalities into pre-trained MLLMs with three modality pairs and show that, by learning an aligned latent representation space, our model achieves strong cross-modal performance.

Multimodal large language models (MLLMs) achieve strong performance by jointly processing inputs from multiple modalities, such as vision, audio, and language. However, building such models or extending them to new modalities often requires large paired datasets and substantial computational resources. Since many pretrained MLLMs (e.g., vision-language or audio-language) are publicly available, we ask whether we can merge them into a single MLLM that can handle multiple modalities? Merging MLLMs with different input modalities remains challenging, partly because of differences in the learned representations and interference between their parameter spaces. To address these challenges, we propose Singular Subspace Alignment and Merging (SSAM), a training-free model merging framework that unifies independently trained specialist MLLMs into a single model capable of handling any combination of input modalities. SSAM maintains modality-specific parameter updates separately and identifies a shared low-rank subspace for language-related parameter updates, aligns them within this subspace, and merges them to preserve complementary knowledge while minimizing parameter interference. Without using any multimodal training data, SSAM achieves state-of-the-art performance across four datasets, surpassing prior training-free merging methods and even jointly trained multimodal models. These results demonstrate that aligning models in parameter space provides a scalable and resource-efficient alternative to conventional joint multimodal training.

Multimodal Large Language Models (MLLMs) rely on multimodal pre-training over diverse data sources, where different datasets often induce complementary cross-modal alignment capabilities. Model merging provides a cost-effective mechanism for integrating multiple expert MLLMs with complementary strengths into a unified model. However, existing model merging research mainly focuses on post-finetuning scenarios, leaving the pre-training stage largely unexplored. We argue that the core of MLLM pre-training lies in establishing effective cross-modal alignment, which bridges visual and textual representations into a unified semantic space. Motivated by this insight, we introduce the post-alignment merging task, which aims to integrate cross-modal alignment capabilities learned from heterogeneous multimodal pre-training. This setting introduces two key challenges: cross-domain parameter interference, where parameter updates learned from different data distributions conflict during merging, and layer-wise alignment contribution disparity, where different layers and projectors contribute unevenly to cross-modal alignment. To address them, we propose PivotMerge, a post-alignment merging framework for cross-modal projectors. PivotMerge incorporates two key components: Shared-space Decomposition and Filtering, which disentangles shared alignment patterns from domain-specific variations and suppresses conflicting directions, and Alignment-guided Layer-wise Merging, which assigns layer-specific merging weights based on differing alignment contributions. We construct systematic CC12M-based post-alignment merging scenarios for evaluation. Extensive experiments on multiple multimodal benchmarks show that PivotMerge consistently outperforms existing baselines, demonstrating its effectiveness and generalization ability.