Yiyang Duan

Research interests

• LLM-agent safety

  Spotting risk early as multi-step agent interactions unfold.

• Inference-time interventions

  Steering model behavior through representations, without retraining.

• Retrieval-augmented generation

  Making retrieval-based systems safer and easier to control.

• Federated learning

  Prototype-based adaptation and collaboration across decentralized data.

• Model merging and efficiency

  Estimating merging compatibility and pruning models at inference time.

Publications

1. NeurIPS 2026

   WURI: Watching unfolding risk in agent interactions

   Y. Duan, T. Wu, Y. Hao, W. Y. B. Lim

2. KDD 2026 Research Track

   Alignment-free multi-modality large-small model bidirectional collaboration with missing modality

   W. Guo, J. Mao, Y. Tong, C. Fang, X. Zhang, Y. Ban, Z. Hu, Y. Duan, F. Zhuang

3. ICML 2026

   SubspacePath Pruner: Inference-time pruning via probe-based representation–parameter coupling

   Z. Gong, Y. Hou, F. Wu, C. Wang, F. Zhang, T. Wu, Y. Hao, J. Zhang, Y. Duan, T. Wang, F. Huang, C. Yuen, W. Y. B. Lim

4. AAAI 2026

   M-Loss: Quantifying model merging compatibility with limited unlabeled data

   T. Wang, Y. Duan, H. Chen, T. Wu, W. Y. B. Lim

5. KDD 2025 Research Track

   FEZE: Alignment-flexible zero-shot vertical federated learning

   W. Guo, Y. Tong, Y. Duan, F. Zhuang, X. Zhang, Z. Hu, J. Dong

6. ICLR 2025

   Enhancing federated domain adaptation with multi-domain prototype-based federated fine-tuning

   J. Zhang, Y. Duan, S. Niu, Y. Cao, W. Y. B. Lim

Preprints

1. arXiv 2026

   C²U in RAG: Compositional concept unlearning in retrieval-augmented generation

   Y. Duan, F. Wu, T. Wu, C. Wang, Z. Gong, W. Guo, B. Delattre, Y. Cao, W. Y. B. Lim

   arXiv:2610.29223

2. arXiv 2026

   Learning marginal viewpoint utility for budgeted evidence selection in RAG

   Y. Duan, Z. Yu, T. Wu, Y. Wu, W. Y. B. Lim

   arXiv:2610.29234

Experience

1. 2025 – present

   Research on AI safety, RAG and LLM-agent safety

   Nanyang Technological University, with Asst. Prof. Lim Wei Yang Bryan

   Studying inference-time representation interventions for safe and controllable AI systems, with a focus on retrieval-augmented generation and multi-step LLM-agent trajectories. Built WURI, a lightweight prefix-level monitor that detects unfolding risk early in agent interactions (NeurIPS 2026), and C²U, which unlearns concepts on the retrieval side by editing queries in embedding space, without retraining the retriever or rebuilding the index.

2. Aug 2024 – Jun 2025

   Federated learning and edge computing (M.Sc. project)

   Nanyang Technological University, with Asst. Prof. Lim Wei Yang Bryan

   Developed MPFT, a multi-domain prototype-based federated fine-tuning method for federated domain adaptation with single-round communication (ICLR 2025). Proposed M-Loss, a label-free metric for estimating model-merging compatibility in resource-constrained edge environments (AAAI 2026). Extended the prototype-based framework to multimodal continual learning.

3. Jul 2023 – Dec 2024

   Research Assistant, Transfer Learning Laboratory

   Beihang University, with Prof. Zhuang Fuzhen

   Built and evaluated vertical federated learning models, improving both the performance and the security of existing methods.

Education

1. 2026 – present

   Ph.D. in Computer Science

   College of Computing and Data Science, Nanyang Technological University. Advisor: Asst. Prof. Lim Wei Yang Bryan

2. 2025

   M.Sc. in Artificial Intelligence

   College of Computing and Data Science, Nanyang Technological University

3. 2024

   B.Eng. in Artificial Intelligence

   Institute of Artificial Intelligence, Beihang University, Beijing