SAFTAC: Simulation-Augmented Fine-Tuning of Open-Source LLMs for Analog Circuit Design
UIUC, Urbana, IL
Advisor: Huan Zhang
2025 - Present
- Motivation: This project studies post-training open-source LLMs for specification-conditioned analog circuit design, where models must reason over circuit structure, constraints, and simulation feedback.
- Result: We constructed a simulation-grounded dataset of 8,626 design tasks across 10 analog circuit types and fine-tuned Qwen3-8B to achieve 71.4% single-pass pass@1 and 78.2% two-pass pass@1, outperforming strong proprietary and open-source baselines.
- Contribution: I led the project design, built the dataset and training pipeline, designed a three-stage SFT curriculum, and implemented a GRPO-based reinforcement fine-tuning pipeline using topology checks, BO-based sizing, and SPICE simulation outcomes as reward signals.
MAC-Tuning: LLM Multi-Compositional Problem Reasoning with Enhanced Knowledge Boundary Awareness
Remote
Advisor: May Fung
Jun 2024 - Current
- Motivation: Existing evaluations of LLM hallucination mostly focus on single-question settings. This project studies hallucination detection and reasoning when models must answer multiple compositional questions at the same time.
- Result: We proposed Multiple Answers and Confidence Stepwise Tuning (MAC-Tuning), a fine-tuning method that improves answer precision by an average of 25%. The paper was admitted to EMNLP 2025 Main.
- Contribution: I led the data collection and training pipeline, including multi-problem prompt construction, model fine-tuning, and output evaluation. I also designed evaluation metrics including accuracy, AP, and MAP, implemented automatic evaluation with LLM-as-a-judge and keyword-based extraction, and manually verified selected cases for reliability.
- Link: Github Repo and Paper.
Random Augmentations Cheaply Break LLM Safety Alignment
UIUC, Urbana, IL
Advisor: Gagandeep Singh, Jason Vega
Jun 2024 - Dec 2024
- Motivation: Many jailbreak methods are costly or require substantial manual effort. This project studies whether simple random prompt augmentations can weaken safety alignment in LLMs.
- Result: We show that low-resource and unsophisticated attackers can significantly improve their chances of bypassing alignment with just 25 random augmentations per prompt.
- Contribution: I implemented and evaluated multiple low-cost input augmentation strategies, including string-level and character-level perturbations. I also helped design the evaluation metrics, manually labeled model outputs, and conducted case studies comparing LLM-based safety classification with human evaluation.