Under Construction!

Midori [緑 ] means "Green", a core philosophy behind our research questions and approaches.

RQ1) Decentralized AI Challenge: How can we tackle the complexities associated with collaborative learning environments?

RQ2) Responsible AI Challenge: How do we defend against malicious actors in an age where AI is for all to use?

RQ3) Minimalistic Intelligence: How can we do more with less?

RQ4) Intelligent Applications: Beyond standard benchmarks, how can we solve domain-specific problems with green intelligence?

🤓We particularly value works led by our own team members, or works originated at our lab by visiting students. 😆

For the below selected publications, *denotes our team members (my NTU students/staff)

#denotes our visiting students (work completed during visiting attachment to our team at NTU)

Selected Works

2026

Project page

Jiaming Zhang*, Che Wang#, Yang Cao, Longtao Huang, Wei Yang Bryan Lim. "Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models" ICLR 2026.

TL;DR: We introduce ReasonBreak, a concept-targeted perturbation method that disrupts multimodal models’ hierarchical geolocation reasoning so that malicious users cannot infer where a photo was taken.

Addresses: RQ2, RQ3

Project page coming soon

Zeng Zihao*, Siyi Li*, Xinyu Yan*, Lei Xiao, Wei Yang Bryan Lim. "AdaCache: Adaptive Caching and Context Augmentation for Efficient LLM Serving" ICLR 2026.

TL;DR: We introduce AdaCache, a method that speeds up RAG LLM serving by (i) reusing retrieved-context KV caches more flexibly via attention-aware hierarchical caching with selective token recomputation, and (ii) adaptively stopping context expansion early using a lightweight confidence score, achieving large TTFT reductions while preserving answer quality.

Addresses: RQ3, RQ4

Project page coming soon

Tianxiang Chen#, Wenjie Hou, Feng Wang, Tiantong Wu*, Zhiming Zheng, Shaoting Tang, Wei Yang Bryan Lim. "FedAdamom: Adaptive Momentum for Improved Generalization in Federatedd Optimization" CVPR 2026 (Oral).

TL;DR: We introduce FedAdamom, which adapts server momentum instead of the learning rate, to keep fast saddle-point escape while better selecting flat minima, improving both convergence and test accuracy.

Addresses: RQ1, RQ3

Project Page with Paper + Code

Tiantong Wang*, Yiyang Duan*, Haoyu Chen#, Tiantong Wu*, Wei Yang Bryan Lim. "M-loss: Quantifying Model Merging Compatibility With Limited Unlabeled Data." AAAI 2026.

TL;DR: We introduce M-loss, an unlabeled-data metric that assesses when weight-merged models approximate output ensembling, and use it to guide model-merging strategies.

Addresses: RQ1, RQ3

Project Page with Paper + Code

Fuyao Zhang*, Xinyu Yan*, Tiantong Wu*, Wenjie Li#, Tianxiang Chen#, Yang Cao, Ran Yan, Longtao Huang,
Wei Yang Bryan Lim, Qiang Yang. "Oblivionis: A Lightweight Learning and Unlearning Framework for Federated Large Language Models." AAAI 2026.

TL;DR: We introduce Oblivionis, a lightweight framework that integrates federated learning and targeted unlearning for LLMs, to enable privacy-preserving training and compliance with GDPR's right to be forgotten.

Addresses: RQ1, RQ2, RQ3

2025

Project Page with Paper + Code

Fan Wu*, Xinyu Yan*, Jiabei Liu*, Wei Yang Bryan Lim. "FedRAM: Federated Reweighting and Aggregation for Multi-Task Learning." NeurIPS 2025.

TL;DR: We introduce a three-stage federated multi-task learning method that leverages a small proxy model to learn task and client weights, enabling efficient and accurate aggregation during fine-tuning of larger models under heterogeneity.

Addresses: RQ1, RQ3

Project Page with Paper + Code

Che Wang#, Ziqi Zhang, Yinggui Wang, Tiantong Wang*, Yurong Hao*, Jianbo Gao, Tao Wei, Cao Yang, Zhong Chen,
Wei Yang Bryan Lim. "AegisGuard: RL-Guided Adapter Tuning for TEE-Based Efficient & Secure On-Device Inference." NeurIPS 2025.

TL;DR: AegisGuard protects only the most model-stealing-sensitive LoRA adapters inside a TEE, compressing them to achieve near black-box stealing resistance while reducing latency, memory footprint, and accuracy loss.

Addresses: RQ1, RQ2

Project Page with Paper + Code

Jingyuan Zhang*, Yiyang Duan*, Shuaicheng Niu, Yang Cao, and Wei Yang Bryan Lim. "Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning." ICLR 2025.

TL;DR: We introduce a one-round federated domain adaptation approach that replaces averaging-based model aggregation with server-side fine-tuning of a global adapter on each client’s small prototype set, with differential privacy guarantees.

Addresses: RQ1, RQ2, RQ3

Project Page with Paper + Code

Zhengyi Zhong#, Weidong Bao, Ji Wang, Shuai Zhang, Jingxuan Zhou, Lingjuan Lyu, Wei Yang Bryan Lim. "Unlearning through Knowledge Overwriting: Reversible Federated Unlearning via Selective Sparse Adapter" CVPR 2025.

TL;DR: We first identify only the most “knowledge-sensitive” layers of a model, then overwrite them via sparsely trained adapters while freezing everything else. This achieves effective forgetting (on par with full retraining) yet remains reversible, costs far fewer resources, and preserves other knowledge better than standard methods.

Addresses: RQ2, RQ3

Project Page with Paper + Code

Wenjie Li#, Kai Fan, Jingyuan Zhang*, Hui Li, Wei Yang Bryan Lim, and Qiang Yang. "Enhancing Security and Privacy in Federated Learning using Low-Dimensional Update Representation and Proximity-Based Defense." TKDE 2025.

TL;DR: We propose FLURP, which compresses each client’s updates into a low-dimensional representation, and design a privacy-preserving, proximity-based rule to filter out poisoned updates. Experiments show that this drastically reduces secure-computation overhead while maintaining high accuracy against various Byzantine attacks.

Addresses: RQ2, RQ3

Page updated

Google Sites

Report abuse