Boyu Tian is a computer systems researcher working on infrastructure for AI agents. His research connects memory-system design with model serving and agent workloads, including concurrency control when many agents share limited accelerator memory.[1][2]
Start with CONCUR →
Education
Links
Website
GitHub Google Scholar ORCID Career
Tian completed a doctorate at Tsinghua University in 2025, following an engineering degree at Shanghai Jiao Tong University. His doctoral work examined near-data processing, memory pooling and disaggregated memory. His current biography describes work at a stealth startup building language-model infrastructure, without naming the employer.[1]
Contributions
Control how many agents run at once
CONCUR studies batch inference for agents whose tool calls and long contexts make their lifetimes difficult to predict. Admitting too many agents can overfill the key-value cache and cause repeated memory movement. The proposed controller uses runtime congestion signals to adjust agent-level concurrency, aiming to keep the system productive without overwhelming its memory capacity. Tian is a co-author of the study.[2]
Build systems for post-training
Tian’s project list includes Weaver, NexRL and the Nex-N1 agentic-model report. These connect agent execution with fine-tuning and reinforcement learning. His earlier systems work provides a useful perspective on the memory, scheduling and coordination costs underneath a model-training interface.[1]
Tools, writing & talks
Writing, research & talks
Start here
CONCURarxiv.orgCongestion-based admission control for high-throughput agentic batch inference.
Read the KV-cache congestion example, then follow how the controller changes concurrency. Basic model-serving and memory concepts help. Treat the reported throughput gains as results for the evaluated workloads, rather than a promise for every agent service.
Co-author[2]- Nex-N1 technical reportarxiv.org
A team report on training agents through large-scale executable environments.
Co-author[3]