Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

You Only Communicate Once: One-shot Federated Low-Rank Adaptation of MLLM

Authors: Binqian Xu, Haiyang Mei, Zechen Bai, Jinjin Gong, Rui Yan, Guosen Xie, Yazhou Yao, Basura Fernando, Xiangbo Shu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that YOCO cuts communication to 0.03% of multi-round FL while surpassing those methods in several multimodal scenarios and consistently outperforming all one-shot competitors. ... 5 Experiments 5.1 Experimental Setup 5.2 Experimental Results
Researcher Affiliation Academia 1Nanjing University of Science and Technology 2Show Lab, National University of Singapore 3Institute of High-Performance Computing, A*STAR
Pseudocode Yes Algorithm 1 YOCO
Open Source Code Yes https://github.com/1xbq1/Fed MLLM ... The code is also included in the supplementary material.
Open Datasets Yes To evaluate YOCO s effectiveness in true one-shot communication for MLLM adaptation, we use five public multimodal datasets: Hateful-Memes [25], Crisis MMD [2], VQA-RAD [27], SLAKE [30], and Med Alpaca [18] (the last three are medical datasets).
Dataset Splits Yes Following [54], we conduct experiments under aligned, missing, cross, and hybrid modal scenarios with non-IID partitioning from a Dirichlet distribution (α = {5.0, 1.0, 0.5}), considering multimodal heterogeneity (β = {30%, 40%, 50%}, I-T={3-7, 5-5, 7-3}, and p = {80%, 70%, 60%}).
Hardware Specification Yes All experiments are performed on NVIDIA A40.
Software Dependencies No As shown in Table 4, we also conduct experiments with different MLLM versions and varying total numbers of clients. ... Mini CPM-V-2_6-int4 [62] as the default version of MLLMs. ... Mini CPM-Llama3-V-2_5-int4 [62].
Experiment Setup Yes Only the last layer s q_proj is fine-tuned using Lo RA [23], with a rank of 8, a lora_alpha of 8, and a dropout rate of 0.05. Under the default setting, the total number of clients is 10, with each client trained for 10 epochs. We report the best performance of different baselines at both the 5th and 10th epochs. For the Hateful-Memes dataset, the initial learning rate is set to 2e-5, while for the other datasets, it is set to 2e-4. A cosine learning rate scheduler is used, with a warmup ratio of 1%. The per-device training batch size is set to 1, and the gradient accumulation steps are set to 4. ... Details of γ and λ are in the appendix.