Long-horizon agents
How do agents stay coherent, recover from failure, and keep making progress across long, open-ended tasks?
↗Currently exploring embodied intelligence @ YesAI Lab
Student researcher · Builder · Community organizer
Hi, I'm Bofan.
I'm a 19-year-old computer science student at UESTC, working on long-horizon agents, agent failure, and embodied intelligence — from evaluation pipelines to real robots.
research mode ✦
UESTC · 2025
LONG-HORIZON AGENTS✦AGENT FAILURE✦VLA✦WORLD-ACTION MODELS✦ROBOTIC MANIPULATION✦
What keeps me curious
Research questions
I care about agents that operate beyond a single prompt — systems that remember, adapt, recover, and eventually act in the real world.
Embodied intelligence
Learning VLA, WAM, world models, and robotic manipulation through hands-on deployment on real robots.
↗Agent evaluation
Building harnesses and benchmarks that reveal what AI agents can actually do — and where they break.
↗Selected work
Research, in progress
LIGpaper
Agent failure · 2026
Let It Go
Narrative Lock-in in Open-ended Real-Life AI Agents
A first-author study of how open-ended agents become trapped by their own narratives and continue acting on stale assumptions.
RESULT / STATUSICML 2026 Workshop on Agent Generalization · EMNLP 2026 submission
ARexperiment
Automated research · EPIC Lab
Autoresearch for Diffusion Cache
Multi-agent workflow for end-to-end research
Designed an automated research workflow that searched, tested, and improved diffusion-cache ideas under a strict compute budget.
RESULT / STATUSBetter than existing baselines · 1× RTX 4090 · 8-hour budget
01exploring
Embodied AI · YesAI Lab
From agents to robots
VLA, WAM, and robotic manipulation
Operating real robotic systems, deploying models, and learning the full manipulation stack from perception to action.
RESULT / STATUSCurrent exploration · July 2026 — now
Built in public
Projects & infrastructure
AGENT EVALUATION01
Growing Bench: Worth It
Runs agents on disposable code and LaTeX workspaces, records their trajectories, and judges whether the work justified its time, scope, and interaction cost.
LEARNING SYSTEM02
MetaCode
An all-in-one learning and practice environment for people getting started with AI coding.
AGENT WORKFLOW03
NeoMyGo
A hands-on workflow for exploring agent-driven cache research and experimentation.
EVALUATION INFRA04
Nexus
An agent harness and LLM evaluation pipeline for repeatable, evidence-driven experiments.
BENCHMARK05
ClawMark
Contributing tasks and evaluation work to a benchmark for long-horizon AI agents.
LANGUAGE
ESSAY 04投稿越来越多，然后呢？投稿、评审与职业回报怎样形成一个越转越快的循环；当六万篇都认真，研究、判断与生活又该怎么继续。阅读文章 ↗
NOTES 03具身智能随想My Thoughts on Embodied Intelligence一些关于 VLA/WAM、coding agent 构建环境、具身 Harness、物理与语义 grounding 的阶段性想法。阅读随想 ↗
BENCH 02做个专测防御性写作和编程的 BenchmarkGrowing Bench: Worth It正确还不够。这些工作真的值得做吗？一个测试 Agent 分寸感、并随前沿模型和真实用户经验不断生长的 Living Benchmark。阅读文章 ↗
CHECKPOINT 01大一年终总结：顿觉天地宽暑假一直在忙，一直想写这样的一个总结。见到很大的世界，思绪很多，所以写给自己，作为我 world model 的一个 checkpoint。阅读文章 ↗Learning by doing
A very early journey
2026.07 — NOW
ShanghaiTech University · YesAI Lab
Research Intern
Embodied intelligence with Shi Ye and Jingya Wang. VLA, WAM, robotic manipulation, and real-robot deployment.
2026.03 — 2026.07
Shanghai Jiao Tong University · EPIC Lab
Research Intern
Agent research with Prof. Linfeng Zhang: automated research for diffusion cache and agent failure in open-ended tasks.
2025.09 — NOW
University of Electronic Science and Technology of China
B.Sc. Computer Science · Honors College
Undergraduate student and president of the UESTC AI Society, building a community around learning and making with AI.
Beyond the lab
Hikaru,
after hours.
When I'm not training agents or deploying robots, I'm probably watching tokusatsu, following a new anime, listening to Stray Kids, or starting yet another sandbox world.
特撮 / TOKUSATSUMyGO!!!!!TerrariaStardew ValleyMinecraftStray KidsAnime
CURRENTLY
Learning how intelligence leaves the screen and enters the physical world.
Let's talk
Have a strange agent problem?
I'm always happy to talk about agents, robots, research, and things worth building.