Tzu-Han Lin

MS Student @ National Taiwan University
Visiting Research Intern @ University of Virginia
Seeking PhD positions for Fall 2026!

I am an M.S. student at the Graduate Institute of Networking and Multimedia, National Taiwan University, fortunate to be advised by Prof. Yun-Nung (Vivian) Chen. I am also currently a visiting research intern at the University of Virginia, advised by Prof. Yu Meng.

I received my bachelor’s degree from National Taiwan University, during which I worked with Prof. Hung-yi Lee and Prof. Yun-Nung (Vivian) Chen.

Research Focus: My research centers on natural language processing, particularly Large Language Models (LLMs). I am interested in developing LLMs and agents that are resource-efficient and reliable.

Nowadays, I think about:

Language Agents:
- Trustworthiness: Beyond task completion, how can we make the behavior of agents more interpretable and faithful to users?
- Efficiency: How can we control or reduce unnecessary search calls in long-horizon search agents (e.g., DeepResearch agents)? How should we quantify and balance the trade-off between reasoning and search?
Self-Improving/Self-Evolving AI Systems:
- Conditions for Autonomous Improvement: What are the necessary conditions for an AI system to reliably and sustainably self-improve?
- Alignment: How can we ensure that self-improving systems remain aligned with human values as they surpass the level at which humans can easily provide direct supervision?
- Verification and Evaluation at Scale: How can we efficiently verify and evaluate progress on long-horizon, open-ended tasks that are inherently difficult to assess (e.g., scientific research agents)?

Misc: When I am not doing research, I like to listen to different genre of musics, especially Blues, Rock, and Grunge. I’m a huge fan of John Mayer and I love to play eletric guitars!

news

Dec 18, 2025	Our new preprint AdaSearch: Balancing Parametric Knowledge and Search in Large Language Models via Reinforcement Learning is out! We investigate how LLM search agents should balance parametric knowledge and search, and propose AdaSearch, a framework that teaches agents to explicitly decide when to search via RL. It achieves superior self-knowledge awareness without complex reward engineering, and provides transparent decision-making rationales.
May 27, 2025	Excited to share that I have joined Prof. Yu Meng’s group as a visiting research intern, and I’m grateful for the opportunity to work with the team!
Feb 19, 2025	Our new preprint Transferring Textual Preferences to Vision-Language Understanding through Model Merging is out! We show that text scalar RMs can be merged into Vision LLMs to build VL-RMs. (Update 05/2025: The paper is accepted to ACL 2025 Main.)
Jul 01, 2024	Our preprint DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging is out! We show that scalar reward models can be merged with intruction-tuned LLMs to derive domain-specific reward models w/o training! (Update 09/2024: The paper is accepted to EMNLP 2024 Main.)
Jan 04, 2024	Our preprint PEFT for Speech: Unveiling Optimal Placement, Merging Strategies, and Ensemble Techniques is out! (Update 02/2024: The paper is accepted to ICASSP 2024 SASB Workshop.)

selected publications

AdaSearch: Balancing Parametric Knowledge and Search in Large Language Models via Reinforcement Learning

Tzu-Han Lin, Wei-Lin Chen, Chen-An Li, Hung-yi Lee, Yun-Nung Chen, and Yu Meng

arXiv preprint arXiv:2512.16883, 2025

Abs PDF Code

Equipping large language models (LLMs) with search engines via reinforcement learning (RL) has emerged as an effective approach for building search agents. However, overreliance on search introduces unnecessary cost and risks exposure to noisy or malicious content, while relying solely on parametric knowledge risks hallucination. The central challenge is to develop agents that adaptively balance parametric knowledge with external search, invoking search only when necessary. Prior work mitigates search overuse by shaping rewards around the number of tool calls. However, these penalties require substantial reward engineering, provide ambiguous credit assignment, and can be exploited by agents that superficially reduce calls. Moreover, evaluating performance solely through call counts conflates necessary and unnecessary search, obscuring the measurement of true adaptive behavior. To address these limitations, we first quantify the self-knowledge awareness of existing search agents via an F1-based decision metric, revealing that methods such as Search-R1 often overlook readily available parametric knowledge. Motivated by these findings, we propose AdaSearch, a simple two-stage, outcome-driven RL framework that disentangles problem solving from the decision of whether to invoke search, and makes this decision process explicit and interpretable. This transparency is crucial for high-stakes domains such as finance and medical question answering, yet is largely neglected by prior approaches. Experiments across multiple model families and sizes demonstrate that AdaSearch substantially improves knowledge-boundary awareness, reduces unnecessary search calls, preserves strong task performance, and offers more transparent, interpretable decision behaviors.
Transferring Textual Preferences to Vision-language Understanding Through Model Merging

Chen-An Li, Tzu-Han Lin, Yun-Nung Chen, and Hung-yi Lee

In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2025

Abs PDF Code

Large vision-language models (LVLMs) perform outstandingly across various multimodal tasks. However, their ability to evaluate generated content remains limited, and training vision-language reward models (VLRMs) with preference data is computationally expensive. This paper explores a training-free alternative by merging text-based reward models (RMs) with LVLMs to create VLRMs. Our approach shows that integrating these models leads to improved performance over LVLMs’ scoring and text-based RMs, offering an efficient method for incorporating textual preferences into LVLMs.
DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging

Tzu-Han Lin, Chen-An Li, Hung-Yi Lee, and Yun-Nung Chen

In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024

Abs PDF Code

Reinforcement learning from human feedback (RLHF) is a popular strategy for aligning large language models (LLMs) with desired behaviors. Reward modeling is a crucial step in RLHF. However, collecting paired preference data for training reward models is often costly and time-consuming, especially for domain-specific preferences requiring expert annotation. To address this challenge, we propose the **Do**main knowled**ge** merged **R**eward **M**odel (**DogeRM**), a novel framework that integrates domain-specific knowledge into a general reward model by model merging. The experiments demonstrate that DogeRM enhances performance across different benchmarks and provide a detailed analysis showcasing the effects of model merging, showing the great potential of facilitating model alignment."
PEFT for Speech: Unveiling Optimal Placement, Merging Strategies, and Ensemble Techniques

Tzu-Han Lin^*, How-Shing Wang^*, Hao-Yung Weng^**, Kuang-Chen Peng^**, Zih-Ching Chen, and Hung-yi Lee

In 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), 2024

Abs PDF

Parameter-Efficient Fine-Tuning (PEFT) is increasingly recognized as an effective method in speech processing. However, the optimal approach and the placement of PEFT methods remain inconclusive. Our study conducts extensive experiments to compare different PEFT methods and their layer-wise placement adapting Differentiable Architecture Search (DARTS). We also explore the use of ensemble learning to leverage diverse PEFT strategies. The results reveal that DARTS does not outperform the baseline approach, which involves inserting the same PEFT method into all layers of a Self-Supervised Learning (SSL) model. In contrast, an ensemble learning approach, particularly one employing majority voting, demonstrates superior performance. Our statistical evidence indicates that different PEFT methods learn in varied ways. This variation might explain why the synergistic integration of various PEFT methods through ensemble learning can harness their unique learning capabilities more effectively compared to individual layer-wise optimization.