Yuhan Yan

你好，我是 Yuhan，专注于 AI 基础设施与高效推理系统的设计与落地。

我目前主要关注以下方向：

模型推理优化：长上下文推理、Token 压缩、缓存复用，以及时延与吞吐优化
模型压缩与部署优化：结构化剪枝、量化，以及 ONNX / TensorRT 部署
训练与服务工程化：分布式训练、平台能力建设和生产环境落地

工作之外，我也享受运动和文化生活。普拉提🧘、美式橄榄球🏈和户外徒步🏞️让我保持活力与专注；电影🎬、播客🎧和阅读📖则不断带来新的感受和灵感。我很享受与不同背景的人交流、碰撞想法，也常常因为好奇心驱动，去深入研究一个问题。

publications

ACM MM

DaG: Cross-Modal-Guided Token Compression with Adaptive Retained-Audio Budgeting for AV-LLMs

Yuhan Yan and Fei Luo

Unique first author submission. , 2026

Under review

Bib

@unpublished{dag2026,
  title = {DaG: Cross-Modal-Guided Token Compression with Adaptive Retained-Audio Budgeting for AV-LLMs},
  author = {Yan, Yuhan and Luo, Fei},
  year = {2026},
  note = {Under review},
}

ICML

ActTaylor: Rethinking Taylor-Based Structured Pruning via Activations

Wenwen Hou, Yuhan Yan, Fei Luo, and 1 more author

2026

Under review

Bib

@unpublished{acttaylor2026,
  title = {ActTaylor: Rethinking Taylor-Based Structured Pruning via Activations},
  author = {Hou, Wenwen and Yan, Yuhan and Luo, Fei and Sun, Ying},
  year = {2026},
  note = {Under review},
}

patents

Training method of frame classification model, testing method and device of application performance CN114898265A