Evaluation and Self-Improvement of Autonomous Research Agents
Advisor: Chao Zhang Ongoing Research
- Investigating how autonomous research agents can improve their research capabilities through iterative experimentation and feedback.
- Developing evaluation protocols to assess research outcomes, intermediate decision quality, and the generalizability of agent self-improvement.
- Exploring hypothesis evaluation and feedback-driven experiment selection to improve the reliability and efficiency of autonomous research.
SAFTAC: Simulation-Augmented Fine-Tuning of Open-Source LLMs for Analog Circuit Design
Advisor: Huan Zhang
- Led the development of SAFTAC, a simulation-augmented post-training framework for specification-conditioned analog circuit generation and feedback-guided correction.
- Built an automated data curation pipeline integrating LLM annotation, Bayesian optimization, and SPICE simulation, constructing 8,626 design tasks across 10 circuit types.
- Designed a three-stage SFT curriculum followed by GRPO-based reinforcement fine-tuning, leveraging simulation-grounded rewards to train both circuit generation and feedback-guided correction.
- Fine-tuned Qwen3-8B, achieving 71.4% single-pass and 78.2% two-pass pass@1, outperforming the strongest closed-source baselines by 15.9 and 7.8 percentage points, respectively.
- Publication: Accepted to NeurIPS 2026 Main Conference.
MAC-Tuning: LLM Multi-Compositional Problem Reasoning with Enhanced Knowledge Boundary Awareness
Advisor: Yi R. (May) Fung
- Led research on LLM confidence estimation and knowledge-boundary awareness in multi-problem settings, covering both independent and sequentially related questions.
- Proposed MAC-Tuning, a two-stage supervised fine-tuning framework that constructs multi-question QA and confidence data to decouple answer prediction from confidence estimation.
- Evaluated MAC-Tuning across six benchmarks and multiple LLM families and sizes, improving average precision by up to 25% over baselines while reducing confidence calibration error.
- Publication: Accepted to EMNLP 2025 Main Conference.
- GitHub Repository and Paper.
Random Augmentations Cheaply Break LLM Safety Alignment
Advisor: Gagandeep Singh
- Investigated the robustness of LLM safety alignment to low-cost random prompt augmentations across 17 models, examining the effects of model size, quantization, fine-tuning-based defenses, and decoding strategies.
- Implemented character-level perturbations and random string insertion strategies with varying augmentation strengths to evaluate their effectiveness in bypassing safety alignment.
- Co-designed human-calibrated safety evaluation metrics using 1,220 manually annotated cases, demonstrating that just 25 random augmentations can substantially increase attack success rates on safety-aligned LLMs.
- Paper (preprint).