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 \textbf{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.
Multimodal pre-training demonstrates strong generalization performance, but this paradigm is often impractical in domains where paired data are scarce. A promising alternative is post-hoc multimodal alignment, which aligns separately pre-trained unimodal encoders using a limited number of paired examples. However, existing methods focus primarily on aligning global representations, missing patch-token relations. This may hinder transfer to tasks that require fine-grained cross-modal matching beyond coarse sample-level semantics. To address this issue, we propose a post-hoc alignment method that learns token-level cross-modal structure using relative representations. Specifically, we represent images and texts through their token-level similarities to a set of learnable anchors in each modality space, which are trained to induce consistent cross-modal similarity patterns for matched pairs. Despite learning only the anchors without heavy projection layers, our approach consistently outperforms existing methods in zero-shot classification, cross-modal retrieval, and zero-shot segmentation by a substantial margin. This highlights the importance of modeling fine-grained cross-modal structure for effective post-hoc multimodal alignment with limited paired data.
Efficiently merging several models fine-tuned for different tasks, but stemming from the same pretrained base model, is of great practical interest. Despite extensive prior work, most evaluations of model merging in computer vision are restricted to image classification using CLIP, where different classification datasets define different tasks. In this work, our goal is to make model merging more practical and show its relevance on challenging scenarios beyond this specific setting. In most vision scenarios, different tasks rely on trainable and usually heterogeneous decoders. Differently from previous studies with frozen decoders, where merged models can be evaluated right away, the non-trivial cost of decoder training renders hyperparameter selection based on downstream performance impractical. To address this, we introduce the task alignment proxy, and show how it can be used to speed up hyperparameter selection by orders of magnitude while retaining performance. Equipped with the task alignment proxy, we extend the applicability of model merging to multi-task vision models beyond CLIP-based classification. Project page: https://europe.naverlabs.com/task-alignment
Pau de Jorge, César Roberto de Souza, Björn Michele +5
The choice between cross-attention and concatenation for multimodal fusion remains governed by practitioner intuition rather than principled understanding. In this paper, we demonstrate that feature alignment quality, not data scale alone, is the primary determinant of which fusion strategy excels. Through controlled experiments on Flickr8k using two feature extraction backbones (ResNet18 and CLIP ViT-B/32), we show that concatenation outperforms cross-attention by 4.1-5.1 percentage points across all tested scales (2048-16384 samples) when features are pre-aligned by a vision-language pretraining objective. We provide a theoretical explanation grounded in sample complexity analysis: concatenation requires O(d_v + d_t) samples to learn its fusion projection, while cross-attention requires O(d_v * d_t) samples to learn bilinear attention weights, over 256 times as many for 512-dimensional CLIP features. When features are already aligned, the approximation error gap between the two methods vanishes, and concatenation's sample efficiency dominates at all practical dataset sizes. An alignment degradation study confirms a monotonic trend: as feature alignment degrades, concatenation's advantage grows from 1.3% to 2.8%. These findings provide a principled decision framework for fusion method selection in multimodal systems, with direct implications for the design of Multimodal Large Language Models.