Research directions and publications. My name is highlighted.
Modulo ADCs fold high-dynamic-range signals into a bounded range, and classical unfolding breaks down when the threshold, sampling density or signal model is mismatched. I build reconstruction methods that stay reliable — a multi-output network with a physics-based refinement (6× lower NMSE than the best classical method on a mixed benchmark), integer-aware OMP for compressive unlimited sampling, and a systematic robustness study for radar signals.
A hierarchical diffusion model that denoises scaffold atoms before substituents, with an EGNN that also predicts each atom's role and time-gated classifier guidance — 91.7% molecular stability on QM9 and markedly faster convergence than EDM.
Self-study of Stanford CS336: a byte-level BPE tokenizer and Transformer LM written from scratch, Chinchilla-style scaling-law fitting, a Common Crawl filtering and deduplication pipeline, and SFT/GRPO on Qwen2.5-Math-1.5B that raised MATH accuracy to 63%.
At Tsinghua, delay analysis of wireless federated learning with saddle-point approximation and large deviation theory, and popularity prediction for edge caching. At Huawei, wireless localization and base-station parameter calibration that replaced manual drive tests in a commercial platform.
Have fun and try to win.
Submitted to IEEE Transactions on Signal Processing, 2026
Submitted to IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2027), 2026
IEEE Radar Conference (RadarConf'26), 2026
Manuscript, 2025
IEEE Journal on Selected Areas in Communications, 2021
PCT Patent WO2021159304A1, published Aug 2021, 2021
IEEE International Conference on Communications (ICC), 2021
IEEE Global Communications Conference (GLOBECOM), 2020