I'm a Ph.D. student at UC Merced, advised by Wan Du. I work on LLM post-training, efficient inference, and AI for science.
More capable AI.
Faster, cheaper inference.
New ways to solve scientific problems.
The value of AI depends on model capabilities, deployment constraints, and how it is used. Research across algorithms, inference, and scientific applications connects these concerns and brings different perspectives to the same problem. Working across these areas helps sharpen research judgment and identify questions worth pursuing. Even a highly capable model can fall short when its use does not fit the problem.
Imperfect action is better than perfect inaction.
Recent preprints, publications, and manuscripts under review.
Under review, ACL ARR · arXiv, October 2026
We study OPD from a knowledge-distillation perspective: which teacher corrections help a student reason better? By changing tokens while keeping model weights fixed, we find that the teacher's advantage accumulates across many corrections, and that choosing where to intervene matters.
Under review, ICLR 2027 · arXiv, October 2026
Can agents cooperate under epistemic and communication asymmetry? Existing benchmarks study information asymmetry; SANZARU adds unequal abilities to see, hear, and speak through 120 cooperative bomb-defusal puzzles. Across seven models, teams solve 49% of the puzzles, and trace analysis attributes about 60% of failures to communication.
arXiv preprint · October 2026
Some statements are neither true nor false. We test whether LLMs recognize this in plural descriptions using 1,000 diagnostic examples. Across seven open-weight models, we find a tendency to force these cases into binary judgments.
Under review, ICLR 2027
Orthogonal BackFill (OBF) compresses the KV cache passed between agents by folding information from discarded values into the retained cache. It reduces per-agent relay volume by approximately 4.7× compared with full-cache relay. At the same relay budget, OBF outperforms eviction alone in 31 of 42 comparisons across seven benchmarks and Qwen3 models from 4B to 14B.
Under review, ICLR 2027
Can a weaker teacher still help train a student? cOPD investigates this question through contrastive supervision for on-policy distillation.
Electronics, 13(24), 5042
FedSparse uses historical gradients to identify important parameters and applies stronger sparsity penalties to less important updates. This reduces communication by up to 24% while keeping accuracy nearly unchanged.
IEEE International Conference on Communications (ICC), pp. 2622–2627
When a client's data distribution changes, the global model may not yet reflect that change. HybridCom combines local performance and server-side gradient signals to prioritize these clients, improving accuracy by approximately 1.3% over the evaluated baselines.
2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, pp. 119–124
Federated Gradient Tailor corrects conflicting client gradients before aggregation. On non-IID MNIST and CIFAR-10, it improves accuracy by 3.2–5.8 percentage points over FL without gradient calibration and reaches baseline accuracy in 2–6× fewer communication rounds. The paper also provides a convergence proof.
IEEE HPCC/DSS/SmartCity/DependSys, pp. 928–935
DCD uses feedback from the teacher and student to adjust the order of training samples during distillation. Compared with random sampling, it improves accuracy by 1.7% on CIFAR-10, 2.5% on CIFAR-100 and 1% on CINIC-10.
I am a Research Assistant at UC Merced, advised by Wan Du, since August 2025. Previously, I worked remotely with Yuchao Zhang at the joint laboratory of BUPT and ChuangCache.
LLM post-training and multi-agent systems
AI Hardware Simulation · KV-Cache Compression
LLMs for quantum computing
Joint Laboratory of BUPT and ChuangCache — Research Assistant (Remote), advised by Dr. Yuchao Zhang
Three applied-ML roles before the PhD.
Brix AI — Irvine, CA
Orbex Labs — Remote
Baidu, Inc. — Beijing, China
University of California, Merced — advised by Dr. Wan Du
University of Illinois Urbana-Champaign
University of Washington
I'm always happy to talk about LLM post-training, efficient inference, and AI for science. Feel free to reach out if you work on related questions or would like to collaborate.