Mang Ye

Wuhan University
2 previous affiliations
270Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Mang Ye focuses on privacy-preserving federated learning, multimodal analysis, and person re-identification. His work addresses the challenges of distributed model training across heterogeneous clients and the complexities of cross-spectral and lifelong visual retrieval. Recently, he has concentrated on parameter-efficient adaptation and safety alignment for multimodal large language models.

Research Trajectory

In 2025, Ye advanced the capabilities of distributed AI systems, particularly in environments characterized by diverse client architectures and corrupted data. He developed frameworks to address the research direction of Federated Learning with Diverse Model Architectures. This is exemplified by Robust Asymmetric Heterogeneous Federated Learning With Corrupted Clients, which introduced a contrastive learning strategy and an asymmetric communication protocol to maintain high accuracy despite severe data noise. He also investigated the balance between global consensus and client equity in Rethinking Fair Federated Learning from Parameter and Client View, proposing a method to prune redundant parameters that cause inter-domain conflict and dynamically shift aggregation focus to ensure fairness across diverse clients. Extending federated paradigms to non-Euclidean data, he contributed to the area of Managing Topological Heterogeneity in Federated Graph Learning with OASIS: One-Shot Federated Graph Learning via Wasserstein Assisted Knowledge Integration. This framework enables a global model to learn from decentralized graph data in a single communication round by synthesizing synergy graphs and utilizing a topological codebook to transfer intricate structural details.

Building on his expertise in complex data representations, Ye's 2026 research heavily emphasized visual retrieval and person re-identification (ReID) across challenging conditions. To tackle the domain of Bridging the Modality Gap in Cross Spectral Vision, he introduced a novel task for 24-hour person retrieval in Cross-Modal Semantic Decoupling and Transfer for Text-to-Visible-Infrared Person Re-Identification. This work separates color-dependent features from structural ones to match text descriptions to both visible and infrared images. He further expanded ReID to long-term scenarios with Towards Realistic Lifelong Re-identification: Identity Recurrence with Changing Clothes, decoupling permanent body features from transient clothing to prevent representation drift as models evolve over time. Moving beyond static queries, he explored interactive retrieval systems in Interactive Person Retrieval via Multi-Turn Multimodal Conversation, developing a memory-enhanced framework that tracks conversation states to iteratively refine searches based on user feedback.

Ye also applied multimodal learning to human-centric analysis. He addressed the difficulty of generalizing affective computing models to unseen individuals in Enhancing Cross-subject Emotion Recognition via Heterogeneous Distribution Augmentation and Collaborative Learning. Recognizing that real-world multimodal streams are often imperfect, he developed API: Adaptive Prototype Imputation for Incomplete Multimodal Sentiment Analysis, a framework that handles missing modalities by retrieving and fine-tuning stable class-level prototypes, avoiding the computational overhead of generative imputation.

In his most recent work, Ye has bridged his expertise in federated learning and multimodal analysis with the rapid advancement of large foundation models. He investigated parameter-efficient fine-tuning in distributed settings, identifying a "robustness asymmetry" in Low-Rank Adaptation (LoRA). This led to Decoupled Low-Rank Adaptation for Robust Federated Fine-Tuning, which uses a dual-branch architecture to isolate noise and protect clean decision boundaries. He further optimized federated personalization in FedPissa: Towards Federated Personalized Adaptation of Foundation Models via LoRA Subspace Mapping, selectively aggregating LoRA components to significantly reduce communication and computation costs. Concurrently, he addressed the generative capabilities and safety of Multimodal Large Language Models (MLLMs). In Optimizing ID Consistency in Multimodal Large Models: Facial Restoration via Alignment, Entanglement, and Disentanglement, he proposed a training-free framework to preserve facial identity during AI image editing by aligning diffusion trajectories. Finally, he tackled the complex cross-modal couplings that can produce unsafe semantics in SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization, guiding models to reason about potential risks without relying on human supervision.

Current Directions

The most recent papers in the evidence indicate a strong focus on the efficient adaptation and safety alignment of foundation models, particularly Multimodal Large Language Models (MLLMs). Ye is actively investigating parameter-efficient fine-tuning techniques, such as Low-Rank Adaptation (LoRA), to enable robust and personalized federated learning while minimizing communication overhead and mitigating the impact of noisy client data. Concurrently, his work addresses the unique vulnerabilities introduced by complex text-image interactions in generative models. This includes developing training-free methods to maintain identity consistency during diffusion-based image editing and designing self-rewarded policy optimization frameworks to ensure the safety and verifiable reasoning of MLLMs without extensive human supervision.

Profile updated

See something wrong? Report a mistake