Vision-Language Models (VLMs) have demonstrated strong performance in multimodal understanding and generation. However, fine-tuning of VLMs typically relies on centralized data, which raises privacy concerns in certain domains (e.g. healthcare). Federated Learning (FL) provides a natural solution by enabling model training without sharing raw data. However, applying FL to VLM instruction tuning is highly challenging. VLMs have substantial parameter scales, and in real-world scenarios, clients exhibit significant heterogeneity in tasks, modalities, and model architectures. Existing methods mainly focus on simplified settings and are unable to handle such multi-dimensional heterogeneous scenarios. In this work, we study federated instruction tuning under joint heterogeneity in tasks, modalities, and model architectures. We propose UniFed-VLM, a unified federated instruction tuning framework for VLMs that addresses multiple types of heterogeneity. It consists of two key components: 1) Federated Compensated Subspace Aggregation (FedCSA), which performs subspace-aligned aggregation of parameter-efficient adapters with dynamic weighting and compensation to mitigate heterogeneity-induced conflicts; 2) Two-stage Collaborative Distillation (TCoD), which enables effective knowledge transfer across heterogeneous models via a Mutual Distillation Adapter (MDA) and a mixture-of-experts-based distillation strategy. We conduct experiments on multiple benchmark datasets, and the results show that UniFed-VLM achieves stronger average performance across diverse tasks compared with existing FL methods. The source code is available at: https://github.com/wangpengyu2004/UniFed-VLM.
Federated instruction fine-tuning enables Large Language Models (LLMs) to adapt to decentralized, privacy-sensitive data without requiring data sharing. Recent Mixture-of-Experts (MoE) LLMs are particularly attractive for federated learning because their sparse activation reduces computation and communication while scaling model capacity. However, existing federated MoE methods primarily focus on parameter aggregation and personalization, overlooking the routing behavior of MoE models as a source of information for client collaboration. Under heterogeneous instruction distributions, indiscriminate aggregation can lead to negative transfer, highlighting the need to identify which clients should collaborate during federated optimization. We propose ClientMorpher, a routing-aware, personalized federated instruction fine-tuning framework that leverages routing signatures from pretrained MoE models to organize client collaboration prior to aggregation. We investigate two complementary clustering strategies: ClientMorpher-C, which directly clusters clients using expert activation profiles, and ClientMorpher-E, which first clusters experts based on their cross-client usage signatures and then derives client collaboration groups. We evaluate ClientMorpher for federated instruction fine-tuning on the Databricks Dolly-15K dataset, using pathological and Dirichlet-based heterogeneous client distributions across multiple instruction-following tasks. Experimental results show that routing-aware collaboration consistently improves personalized performance compared to conventional federated averaging and local training, while maintaining the same communication cost. Furthermore, our study shows that client-centric and expert-centric clustering provides an effective and scalable approach for personalized federated instruction fine-tuning of sparse MoE LLMs.
Instruction tuning aligns large language models, including multimodal ones, with diverse user intents, but scaling to heterogeneous mixtures is hindered by gradient interference and bandwidth-heavy synchronization. We ask whether these two bottlenecks can be addressed jointly by training parts of the mixture independently and reconciling them once in parameter space. We develop a local quadratic theory inside a shared flat basin that yields three results: weight merging produces a curvature-weighted variance reduction; PCA-aligned conflict splitting maximizes this gain along high-curvature directions; and merging additionally acts as spectral filtering with implicit norm regularization. These results directly motivate MERIT, a decentralized merge-ready instruction-tuning pipeline that estimates dataset-level gradient conflicts, partitions the mixture along the top PCA conflict axes, fine-tunes each partition independently with no inter-partition communication, and merges once via token-weighted averaging. On Qwen2.5-VL-3B with 136 Vision-FLAN tasks, MERIT improves the 8-benchmark average from 54.3 (joint training) to 57.0. The same recipe scales to a 7B model on a 1.6M-example, 176-source mixture -- matching or exceeding centralized joint training with minimal cost overhead -- and transfers to text-only FLAN. Our code is available at https://github.com/naver-ai/merit.