Publication
Recursive Synthesis for Long-Horizon Terminal Tasks
Our new work on long-horizon agent workflows is now available on arXiv.
AI Agents · Large Language Models
Anhao Li
Computer Science Undergraduate · University of Minnesota Twin Cities
I work on AI agents and large language models — self-evolving agents, agent harnesses, and reinforcement learning for long-horizon terminal tasks.
I build agent infrastructure and the benchmarks used to evaluate it. Outside research, I have competed in algorithm contests for about ten years, from NOI and USACO to the ICPC World Finals, and now coach the UMN team.
News
Recent updates
Publication
Long-Horizon-Terminal-Bench
We released a benchmark for evaluating agents on extended terminal tasks.
Internship
LLM Agent Infrastructure at VecML
Completed my second software engineering internship at VecML in Bellevue, WA.
ICPC
Qualified for the ICPC World Finals
Ranked 9th in North America at ICPC NAC, becoming the first UMN team to qualify in 10 years.
ICPC
Gold at the North Central NA Regional Contest
Won a second consecutive regional gold medal after our 2024 result.
Internship
AI Data Analysis Platform at VecML
Completed a software engineering internship at VecML in Bellevue, WA.
Research Interests
What I spend my time on
Self-Evolving Agents
Agents that improve from their own execution traces — synthesizing harder tasks, verifying their own outcomes, and feeding what survives back in as training signal.
Agent Harnesses
The scaffolding around the model: tool interfaces, context management, and execution environments. How much of a model's capability actually shows up on a task often depends on the harness as much as the weights.
Post-training
Turning a base model into a dependable agent — supervised fine-tuning, preference optimization, and the data synthesis pipelines that feed them.
Reinforcement Learning
Reward design and credit assignment for agents acting over long horizons, where a single pass/fail at the end says very little about where the run actually went wrong.
Publications
All on Scholar ↗
Recursive Synthesis for Long-Horizon Terminal Tasks
A recursive, verification-driven pipeline that synthesizes progressively harder terminal tasks, giving agents a training and evaluation signal that scales with their capability.
Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading
A benchmark for agents on extended terminal tasks, replacing pass/fail checks with dense reward-based grading so partial progress is measured rather than discarded.
Experience
Software Engineer Intern
LLM Agent Infrastructure & Evaluation
VecML Inc. (AI startup) · Bellevue, WA
Software Engineer Intern
AI Data Analysis Platform
VecML Inc. (AI startup) · Bellevue, WA
Honors & Awards
Selected
ICPC World Finalist 2026
Ranked 9th in North America at ICPC NAC; first UMN team to qualify in 10 years.
Gold Medal — ICPC North Central NA Regional Contest
Gold medalist in two consecutive regional contests.
NOIP / CSP First Prize
Received a cumulative total of six first prizes across NOIP and CSP competitions.
USACO Platinum — Global Top 30
Ranked among the global top 30 competitors.
Bronze Medal — National Olympiad in Informatics (NOI)
Bronze medalist at the 2021 National Olympiad in Informatics.
Silver Medal — National Olympiad in Informatics Winter Camp
Silver medalist at the national winter camp in informatics.
Education
University of Minnesota Twin Cities
Degree
B.S. in Computer Science
GPA
3.96 / 4.0
Activities
Competitive Programming · ICPC Contestant · ICPC Coach, 2026–2027
Teaching & Service
ICPC Coach
Coaching the University of Minnesota competitive programming team — practice contests, training plans, and problem-set design.
Competitive Programming Mentoring
Writing contest editorials and training notes on the blog, and mentoring students moving from introductory algorithms to ICPC-level problems.
Writing
All posts →Contest write-ups, technical notes, and the occasional reflection. Every post is available in English and Chinese.
ICPC
Ten Years of Competitive Programming — A Retirement Memoir
From my first encounter with competitive programming to the ICPC World Finals — a farewell to ten years of contests.
ICPC
ICPC NAC 2026 — A North American Championship Diary
The world was finally gentle with you.
ICPC
2026.2–3 Rehab Training
そうか、大人になったんだね
Get in touch
Happy to talk about agents, evaluation, or contest problems.
Email is the fastest way to reach me.