I study how knowledge is represented inside models, how it moves between them, and where those internal representations conflict.
I study how knowledge is represented inside models, and how it moves and combines across them. The same knowledge takes different carriers in different settings: gradients in federated learning, KV caches in multi-agent LLM systems, weights when specialized models are merged.
Across all of them the recurring obstacle is the same — conflict, where knowledge from different sources interferes once combined. My work so far spans KV-cache compression for latent multi-agent LLMs, knowledge distillation with curriculum learning, and conflict-aware federated learning under non-IID data.
I am currently working on cross-family on-policy distillation, where a student must re-express what the teacher holds in its own internal states, and on conflict in model steering, merging and fine-tuning.
Peer-reviewed work and preprints. Author order follows each venue; my own name is highlighted in every list.
Under review, NeurIPS 2026 · first author
Adapts KV-cache eviction to multi-agent latent communication and introduces Orthogonal BackFill, which injects a low-rank component of the discarded value states back into the retained cache. Compressed relay matches or beats full KV relay on five of nine benchmarks while keeping 9.9–20.2% of the prompt KV.
Electronics, 13(24), 5042
A Resource Optimization Proximal loss that enforces sparse client updates, cutting communication by up to 24% at near-unchanged accuracy.
IEEE International Conference on Communications (ICC), pp. 2622–2627
Adapts client participation to distribution drift through a hybrid client/server contribution indicator.
2nd ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network, pp. 119–124
Federated Gradient Tailor detects and corrects conflicting client gradients before aggregation — 5% higher accuracy with 6× faster convergence on non-IID data, with a formal convergence proof.
IEEE HPCC/DSS/SmartCity/DependSys, pp. 928–935
A curriculum generator that adapts training batches to teacher–student interaction, mitigating the harm of heterogeneous sampling during distillation.
Two labs. The first four publications came out of BUPT; the current line of work is at UC Merced.
University of California, Merced — Research Assistant, advised by Dr. Wan Du
Joint Laboratory of BUPT and ChuangCache — Research Assistant, advised by Dr. Yuchao Zhang
Brix AI — Irvine, CA
Orbex Labs — Remote
Baidu, Inc. — Beijing, China
If you work on distillation, KV-cache compression, model merging, or anything where
knowledge from two models has to coexist, I would like to hear about it.
yipingli@ucmerced.edu