PivotMerge: Bridging Heterogeneous Multimodal Pre-training via Post-Alignment Model Merging

• URL: http://arxiv.org/abs/2604.22823v1
• Date: Sat, 18 Apr 2026 09:38:03 GMT
• Title: PivotMerge: Bridging Heterogeneous Multimodal Pre-training via Post-Alignment Model Merging
• Abstract summary: We argue that the core of Multimodal Large Language Models pre-training lies in establishing effective cross-modal alignment.<n>Motivated by this insight, we introduce the post-alignment merging task, which aims to integrate cross-modal alignment capabilities learned from multimodal pre-training.<n>We propose PivotMerge, a post-alignment merging framework for cross-modal projectors.
• Score: 80.17966517536102
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• Multimodal LLMs under Pairwise Modalities [16.75545711899814]
  We propose a representation learning framework for aligning latent representations across modalities using only pairwise data.<n>Specifically, in the first stage, we learn the shared latent space across modalities by both self-modal reconstruction and pair-wise contrastive learning.<n>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.
  arXiv  Detail & Related papers  (2026-05-20T11:44:01Z)
• Chain of Modality: From Static Fusion to Dynamic Orchestration in Omni-MLLMs [84.3271821505699]
  Chain of Modality (CoM) is an agentic framework that transitions multimodal fusion from passive concatenation to dynamic orchestration.<n> operating in either a training-free or a data-efficient SFT setting, CoM achieves robust and consistent generalization across diverse benchmarks.
  arXiv  Detail & Related papers  (2026-04-16T01:21:14Z)
• Bridging Domains through Subspace-Aware Model Merging [27.981999508897847]
  Model merging integrates multiple task-specific models into a single consolidated one.<n>Recent research has made progress in improving merging performance for in-distribution or multi-task scenarios.<n>We investigate how merging models fine-tuned on distinct domains affects generalization to unseen domains.
  arXiv  Detail & Related papers  (2026-03-06T00:04:24Z)
• Model Merging in the Essential Subspace [78.5390284258307]
  Model merging aims to integrate multiple task-specific fine-tuned models into a single multi-task model without additional training.<n>Despite extensive research, task interference remains a major obstacle that often undermines the performance of merged models.<n>We propose ESM (Essential Subspace Merging), a robust framework for effective model merging.
  arXiv  Detail & Related papers  (2026-02-23T00:33:38Z)
• Beyond Unimodal Shortcuts: MLLMs as Cross-Modal Reasoners for Grounded Named Entity Recognition [51.68340973140949]
  Multimodal Named Entity Recognition (GMNER) aims to extract text-based entities, assign them semantic categories, and ground them to corresponding visual regions.<n> MLLMs exhibit $textbfmodality bias$, including visual bias and textual bias, which stems from their tendency to take unimodal shortcuts.<n>We propose Modality-aware Consistency Reasoning ($bfMCR$), which enforces structured cross-modal reasoning.
  arXiv  Detail & Related papers  (2026-02-04T12:12:49Z)
• Bridging Modalities via Progressive Re-alignment for Multimodal Test-Time Adaptation [39.02105398462778]
  Test-time adaptation (TTA) enables online model adaptation using only unlabeled test data.<n>In multimodal scenarios, varying degrees of distribution shift across different modalities give rise to a complex coupling effect.<n>We propose a novel TTA framework, termed as Bridging Modalities via Progressive Re-alignment (BriMPR)
  arXiv  Detail & Related papers  (2025-11-28T03:33:42Z)
• UniAlignment: Semantic Alignment for Unified Image Generation, Understanding, Manipulation and Perception [54.53657134205492]
  UniAlignment is a unified multimodal generation framework within a single diffusion transformer.<n>It incorporates both intrinsic-modal semantic alignment and cross-modal semantic alignment, thereby enhancing the model's cross-modal consistency and instruction-following robustness.<n>We present SemGen-Bench, a new benchmark specifically designed to evaluate multimodal semantic consistency under complex textual instructions.
  arXiv  Detail & Related papers  (2025-09-28T09:11:30Z)
• Distilling Transitional Pattern to Large Language Models for Multimodal Session-based Recommendation [67.84581846180458]
  Session-based recommendation (SBR) predicts the next item based on anonymous sessions.<n>Recent Multimodal SBR methods utilize simplistic pre-trained models for modality learning but have limitations in semantic richness.<n>We propose a multimodal LLM-enhanced framework TPAD, which extends a distillation paradigm to decouple and align transitional patterns for promoting MSBR.
  arXiv  Detail & Related papers  (2025-04-13T07:49:08Z)
• AsymLoRA: Harmonizing Data Conflicts and Commonalities in MLLMs [5.018961516699825]
  AsymLoRA is a parameter-efficient tuning framework that unifies knowledge modularization and cross-modal coordination.<n>AsymLoRA consistently surpasses both vanilla LoRA, which captures only commonalities, and LoRA-MoE, which focuses solely on conflicts.
  arXiv  Detail & Related papers  (2025-02-27T12:21:02Z)
• Distributional Vision-Language Alignment by Cauchy-Schwarz Divergence [83.15764564701706]
  We propose a novel framework that performs vision-language alignment by integrating Cauchy-Schwarz divergence with mutual information.<n>We find that the CS divergence seamlessly addresses the InfoNCE's alignment-uniformity conflict and serves complementary roles with InfoNCE.<n> Experiments on text-to-image generation and cross-modality retrieval tasks demonstrate the effectiveness of our method on vision-language alignment.
  arXiv  Detail & Related papers  (2025-02-24T10:29:15Z)
• Understanding Alignment in Multimodal LLMs: A Comprehensive Study [46.33812471516309]
  We analyze each aspect of preference alignment in Multimodal Large Language Models (MLLMs) We show that combining offline and online methods can improve the performance of the model in certain scenarios. We introduce a novel way of creating multimodal preference data called Bias-Driven Hallucination Sampling (BDHS)
  arXiv  Detail & Related papers  (2024-07-02T17:55:03Z)
• Abstractive Sentence Summarization with Guidance of Selective Multimodal Reference [3.505062507621494]
  We propose a Multimodal Hierarchical Selective Transformer (mhsf) model that considers reciprocal relationships among modalities. We evaluate the generalism of proposed mhsf model with the pre-trained+fine-tuning and fresh training strategies.
  arXiv  Detail & Related papers  (2021-08-11T09:59:34Z)

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.