I am a direct-entry Ph.D. student in Computer Science at the University of California, Santa Barbara (UCSB), advised by Prof. Xifeng Yan. I am currently interning at Tencent Cloud WorkBuddy on post-training, reinforcement learning, and RSI, and I plan to master out with a master’s degree in July 2027. My research focuses on LLM efficiency, model architecture, speculative decoding, sparse and long-context attention, and autonomous agents.

I aim to connect high-level agentic capabilities with low-level algorithmic innovations, treating model architecture and agent frameworks as complementary paths toward scalable, capable AI.

I have experience building agent systems. Previously, I earned my B.S. in Computer Science from Shanghai Jiao Tong University as a member of the ACM Honors Class.

Publications & Manuscripts

Learning When Not to Attend Globally

All-or-Here Attention (AHA); Xuan Luo*, Jiaming Shan*, Wesley Truong, Kailai Zhang, Hanzhe Zhang, Xifeng Yan
Manuscript PDF (anonymous review copy); equal contribution

Proposed a token-level adaptive attention mechanism whose binary router switches each attention head between global and sliding-window attention. On OLMo-2-1B with a 128-token window, 88.4% of token-head decisions chose local attention with comparable aggregate performance across six benchmarks. AHA also improves DuoAttention as an orthogonal routing layer.

Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants

Deepak Nathani, Cheng Zhang, Chang Huan, Jiaming Shan, Yinfei Yang, Alkesh Patel, Zhe Gan, William Yang Wang, Michael Saxon, Xin Eric Wang
arXiv preprint, 2026

Built benchmark and environment abstractions for proactive and mobile agents, including finite-state-machine app modeling, goal inference, and multi-app orchestration evaluation.

Building Constrained Human-AI Cooperation: An Inclusive Embodied Social Intelligence Challenge

Weihua Du, Qiushi Lyu, Jiaming Shan, Zhenting Qi, Hongxin Zhang, Sunli Chen, Andi Peng, Tianmin Shu, Kwonjoon Lee, Behzad Dariush, Chuang Gan
NeurIPS 2024, Datasets and Benchmarks Track; poster

Conceptualized and implemented physically constrained agents and helper bots in ThreeDWorld, using in-context learning to deploy language-based agents for human-assistance tasks and construct a new benchmark dataset.

Building Cooperative Embodied Agents Modularly with Language Models

Hongxin Zhang*, Weihua Du*, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B. Tenenbaum, Tianmin Shu, Chuang Gan
ICLR 2024; poster; equal contribution

Co-developed a prompt-based state-abstraction framework for multi-agent cooperation, led the VirtualHome environment and heuristic-baseline implementation, and conducted user studies comparing LLM agents with planning baselines.

Research Experience

University of California, Santa Barbara

Ph.D. Student in Computer Science; advisor: Prof. Xifeng Yan; September 2024 - Present

• Co-first-authored All-or-Here Attention and contributed to real-world acceleration and inference-speedup code paths.
• Explored end-to-end fine-tuning for DFlash acceleration.
• Built benchmark and environment abstractions for PARE.

Massachusetts Institute of Technology

Summer Research, CSAIL; advisor: Prof. Chuang Gan; August 2023 - June 2024

• Designed physically constrained agents and helper bots in the ThreeDWorld simulation environment.
• Deployed language-based agents for human-assistance tasks and benchmark generation.

MIT / Shanghai Jiao Tong University

Remote Research; advisor: Prof. Chuang Gan; February 2023 - May 2023

• Co-developed a prompt-based state-abstraction framework for multi-agent cooperation.
• Led VirtualHome environment and heuristic-baseline implementation and conducted user studies.

Internship Experience

Tencent Cloud WorkBuddy

Large Language Model Agent Research Intern; September 2026 - Present

• Trained large language model agents from experience data and conducted research on RSI.

RWKV

Agent Intern; March 2026 - June 2026

• Developed high-concurrency local document summarization and retrieval pipelines.
• Used RWKV to summarize tool outputs and extract key information, reducing downstream token consumption.
• Concurrently generated readable HTML views alongside the primary output.

Agent Daily Use

• Auto Research in Scientific Work. Use auto-research workflows for question decomposition, reference search and synthesis, experiment planning, and reusable research notes; summarized practical lessons in this article.
• TA Grading Agent. Converted the Gradescope grading interface into MCP tools and agent workflows for batch grading; dual-pass grading and human-verification queues reduce TA workload by roughly 70%.
• Thought-to-Blog Agent. Built a pipeline that turns high-signal AI chat records into publishable blog drafts through privacy/value screening and staged working, final, and published versions.
• Autonomous Wealth Research Agent. Built a daily investing automation connected to moomoo and a local agent workforce, with portfolio-aware recommendations, supply-chain research, tax and cash-flow constraints, risk gates, and simulated trading.
• OpenClaw-like Personal Agent Framework. Developed and use a personal agent framework integrating memory, sleep and compaction, MCP tool routing, and long-running personal workflows.

Education

Shanghai Jiao Tong University

B.S. in Computer Science; ACM Honors Class; September 2020 - June 2024

• Elite Computer Science program for students selected from the top 5%.
• GPA in core courses: 3.98/4.3, ranked 3/36.
• 2023 National Scholarship.

University of California, Santa Barbara

Ph.D. Student in Computer Science; advisor: Prof. Xifeng Yan; September 2024 - Present

Honors & Awards

• 2023 National Scholarship Award
• Zhiyuan Honorary Scholarship, 2020-2023 (Top 2% at Shanghai Jiao Tong University)
• 2021 Interdisciplinary Contest in Modeling, Honorable Mention
• 2021 National Undergraduate Mathematical Contest in Modeling, Provincial Third Prize
• 2021 International Physics Competition for University Students, Silver Medal

Technical Skills

• AI / ML: PyTorch, Hugging Face, Triton, FlashAttention, vLLM, Weights & Biases
• AI-native workflow: Codex, OpenClaw, Cursor, Claude Code, Tmux
• Programming: Python, C++, Lisp, Java
• Developer tools: Git, Docker, LaTeX