Under review: We submitted one first-authored paper to NDSS 2027.
Published in Transactions on Machine Learning Research (TMLR), 2026
We construct a benchmark dataset for evaluating agentic open-source software compilation methods and propose a multi-agent compilation method enhanced with LLM-assisted retrieval.
Published in AIP Advances, vol. 14, no. 9, 2024
We propose a new masking pretraining method for Force and Energy-Centric Graph Neural Networks to surrogate molecular dynamics, reducing RMSE for water datasets by up to 38%.
Published in EMNLP 2026 Main Conference, 2026
We introduce critique-aware supervision for training more reliable agents on long-horizon tool-calling tasks.
Published in NeurIPS 2025 Reliable ML from Unreliable Data Workshop (non-archival), 2024
We propose a jailbreaking technique that encodes malicious queries with layered novel ciphers, exploiting LLMs ability to interpret complex instructions, increasing jailbreaking success rates from 40% to 78%.
Published in arXiv preprint, 2024
We study the use of Retrieval-Augmented Generation to improve LLM factual accuracy and reduce hallucinations for domain-specific queries in private knowledge bases.