Pengjie Wang | Tsinghua University

About Me

I am Pengjie Wang (王鹏杰), a third-year undergraduate at Tsinghua University studying Computer Science and Finance (GPA 3.96/4.00, expected graduation June 2028). I work on algorithm–system co-design for efficient and performant large language models, advised by Prof. Xu Han and Prof. Zhiyuan Liu at THUNLP and mentored by Yuxiang Huang.

Previously, I worked on multimodal dynamic graph learning with Zongyuan Wu and advised by Prof. Wenwu Zhu, and worked on efficient attention for diffusion transformers, advised by Jintao Zhang and Prof. Jianfei Chen.

Research Interests

Efficient and performant model architectures: attention mechanisms for long contexts
ML systems: offload-aware scheduling and scalable infrastructure for evolving AI workloads
Continual learning: as a meta-learning problem, where a model learns to turn its own experience and feedback into updates

News

[Sept. 2026] MM-DyGraph is accepted to the NeurIPS 2026 Evaluations and Datasets Track.
[Aug. 2026] NOSA is accepted to the EMNLP 2026 Main Conference.
[Jul. 2026] I gave an oral presentation on NOSA at the ACM TURC 2026 Student Forum, and our poster received the Best Poster Award.

Publications

EMNLP

NOSA: Native and Offloadable Sparse Attention

Yuxiang Huang*, Pengjie Wang*, Jicheng Han, Weilin Zhao, Zhou Su, Ao Sun, Hongya Lyu, Hengyu Zhao, Yudong Wang, Chaojun Xiao†, Xu Han†, Zhiyuan Liu† (*Equal contribution, †Corresponding authors)

Conference on Empirical Methods in Natural Language Processing (EMNLP), Main Conference, 2026. Also at the ICML AdaptFM Workshop.

PDF Code Blog

NeurIPS

MM-DyGraph: A Dataset, Benchmark, and Model for Multimodal Dynamic Graphs

Zongyuan Wu*, Pengjie Wang*, Lingshan Chen*, Tianhang Wan, Yuan Meng, Xin Wang, Ling Feng, Wenwu Zhu (*Equal contribution)

Conference on Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track, 2026.

OpenReview

Selected Projects

UNVEIL: Unmasking via Efficient Inference Learning (NLP course research, team lead) [Code]
Cast diffusion-LLM unmasking as an MDP and learned a lightweight per-token ranking policy to replace static heuristics. Co-designed UNVEIC, which combines commit-and-freeze KV reuse with selective recomputation.
Towards GUI Agents with World Models in Latent Spaces (DRL course research, with Qinwei Ma et al.) [Code]
Extended Digi-Q with a transition model that predicts future VLM embeddings instead of pixels. As an inference-time look-ahead, it improved a pretrained agent’s task success in the general setting, while model-based RL training remained unstable under rollout error accumulation and distribution shift.
What Transfers and What Couples (AI in Quantitative Finance course research, team lead) [Code]
Extended the TRADES diffusion model to multi-asset limit order book generation. Parameter sharing preserved price-placement accuracy for data-poor assets, while explicit lead–lag coupling and no-arbitrage guidance showed no benefit.
StarryOS Kernel Contributions (Operating Systems course) [Code]
10 merged PRs in Rust, including an eBPF runtime, a loadable kernel module loader, POSIX and OFD file locks, and multi-threaded execve.
FPGA Chat System (Digital Logic Design course) [Code]
One-to-one chat between two FPGA boards over UART, with connection handshake, reconnection, quoting, message recall, and stickers.

Honors and Awards

China National Scholarship, 2026
Best Poster Award, ACM TURC Student Forum, 2026
Ruiqi Scholarship, Tsinghua University, 2025
Freshman Scholarship (Second Class), Tsinghua University, 2024
6th of 93, PaddlePaddle MarTech Challenge (Customer Behavior Forecasting), 2025
7th place, 1st Tsinghua Supercomputer Competition (THUSCC), 2024
Second Prize, College Physics Competition in Selected Regions of China (全国部分地区大学生物理竞赛), Physics Majors Division (as a CS major), 2024

Service

Tutor, Tsinghua University Q&A Center (校答疑坊), 2025–2026: 70+ tutoring sessions and 100 hours across Calculus, Complex Analysis, and Fundamentals of Programming.
Organizer, 30th Tsinghua Agent Competition, 2025–2026: helped develop the AntWar 2 game, supported participants, and helped run the finals. The event was named one of Tsinghua’s Top Ten Competitions.