Hello! I am an M.S. student in Computer Science at the University of California, San Diego (UCSD). I recently received my B.S. in Computer Science from National Yang Ming Chiao Tung University (NYCU), where I worked with Prof. Tien-Fu Chen.

Most recently I was a visiting scholar at UCSD, advised by Prof. Lily Weng, working on the internal mechanisms behind refusal behavior in large language models.

Research Interests

My research focuses on large language model (LLM) latent reasoning, safety, and efficient model compression. I am particularly interested in understanding the internal mechanisms of LLMs in order to improve their reasoning, reliability, and efficiency in real-world deployments.

• Latent Reasoning: studying how LLMs can reason in latent space rather than through explicit token-level chains, for more efficient and controllable inference.
• LLM Safety and Interpretability: identifying detection heads and refusal heads inside LLMs, and editing them to make models robust to adversarial attacks — DRefA (NeurIPS 2025 Lock-LLM Workshop).
• Model Compression and Efficiency: structured pruning driven by activation-importance estimation, and post-training quantization for resource-constrained hardware.

News

• [2026.08] Our paper “From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling” was accepted to EMNLP 2026 Findings.
• [2025.10] Our paper “How to Make LLMs Safer? Detecting and Editing Key Heads in LLMs” was accepted to the NeurIPS 2025 Lock-LLM Workshop.
• [2025.06] I joined UCSD as a Visiting Scholar, advised by Prof. Lily Weng, working on LLM safety.

Experience

University of California, San Diego Jun 2025 – Sep 2025

Position: Visiting Scholar
Topic: LLM Safety and Interpretability
Advisor: Prof. Lily Weng

Investigated internal attention mechanisms in large language models, focusing on detection heads that identify harmful content and refusal heads that drive the model to reject unsafe requests, and developed systematic methods to identify and analyze these components.

Proposed Detection Refusal Advanced LLM (DRefA), which amplifies detection and refusal heads. With DRefA, the LLaMA-3 safety rate rose from 15% to 99% under ADV-LLM attacks, at a cost of only ~5% average accuracy across benchmark datasets.

AI System Testing Lab, NYCU Sep 2024 – Jun 2025

Position: Research Assistant
Topic: Structured Pruning for LLMs
Advisor: Prof. Tien-Fu Chen

Developed a pruning approach based on activation-importance estimation from calibration data, applied to LLaMA-3.2-3B-Instruct on WikiText at a 5% pruning ratio.

Achieved a perplexity of 12.23, reducing degradation by 56% relative to random pruning (13.72) with minimal loss in output quality.

Micro-Star International (MSI) Jul 2024 – Jun 2025

Position: Research and Development Intern
Topic: Perception and On-Device Inference for Autonomous Mobile Robots
Location: New Taipei City, Taiwan

Designed and implemented a floor segmentation model for Autonomous Mobile Robots (AMRs) using camera-based perception, reaching 85% segmentation accuracy in test environments.

Deployed the trained models to embedded systems with pruning and quantization, enabling real-time inference at 25+ FPS on constrained hardware with a ~15% model size reduction.

Publications

1. Kuan-Lin Chu, Chung-En Sun, and Tsui-Wei Weng. "From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling" EMNLP 2026 Findings, 2026. [Project Page]
2. Kuan-Lin Chu, Chung-En Sun, and Tsui-Wei Weng. “How to Make LLMs Safer? Detecting and Editing Key Heads in LLMs.” NeurIPS 2025 Lock-LLM Workshop, 2025. [Project Page]

Education

Feel free to reach out for collaborations, discussions, or even a casual chat about research!