Environment
GUI + CLI + Code
Hi! I'm Longtao Zheng.
I build real-world environments for computer agents.
I'm Longtao Zheng (郑龙韬), a researcher at Bytedance, working on RL for coding agents. My research focuses on training open-ended and long-horizon agents. I earned my PhD from Nanyang Technological University (NTU) Singapore in 2026, advised by Prof. Bo An. Previously, I received my Bachelor's degree in computer science from University of Science and Technology of China (USTC) in 2022.
Research
I study the full stack of open-domain and long-horizon computer agents since 2023,
including real-world environments, inference-time harnesses, and RL for LLM agents.
observation
action
Harness
Long-horizon autonomy
context
reasoning / action
LLM
RL self-improvement
Blogs
How to build a fully asynchronous black-box agentic RL system
A first-principles design for fully asynchronous RL with any agent harness
Learning when to compact context in long-horizon coding agents
Teaching coding agents to adaptively compress noisy history into useful context
The science of LLM RL
LLM RL = task distribution + rollout policy + credit assignment + stable policy updates
GRPOshared
Global baseline
Unstable gradients
Dr. MASper-agent
A baseline
B baseline
Stable gradients
Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems
Stable training algorithm and open-source codebase for multi-agent LLM RL
GRPOshared
same for every token
High variance
OTBper-token
W: cumulative uncertainty
Low variance
The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL
A token-level baseline prevents RL training collapse and reduces token consumption
KernelGYM
Multi-turn RL
Test-time scaling
Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations
Optimizing Triton kernel generation with multi-turn RL and test-time scaling
void-turn trajectory filtering
score
grad norm
SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning
Simple trajectory filtering stabilizes multi-turn RL and emerges diverse reasoning
Towards Efficient Online Tuning of VLM Agents via Counterfactual Soft Reinforcement Learning
Finetuning VLM agents with online RL
Cradle: Empowering Foundation Agents Towards General Computer Control
A general computer-control agent that acts through visual observations and keyboard/mouse input
MEMO: Memory-Guided Diffusion for Expressive Talking Video Generation
A SOTA and open-weight model for audio-driven talking video generation
Environment
Tools
Benchmarks
AgentStudio: A Toolkit for Building General Virtual Agents
A trinity of environments, tools, and benchmarks for general virtual agents
FinAgent
A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist
The first multimodal agent for financial trading
Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control
One of the earliest web agents with state abstraction, trajectory prompting, and memory
True Knowledge Comes from Practice: Aligning Large Language Models with Embodied Environments via Reinforcement Learning
Finetuning LLM agents with online RL
Baselineconfounded
different actions see different teammates
Amostly T1
Bmostly T2
Biased action choice
CTCATrectified
overrep. ↓underrep. ↑
AT1 + T2 adjusted
BT1 + T2 adjusted
Stable under shifts
Controlling Type Confounding in Ad Hoc Teamwork with Instance-wise Teammate Feedback Rectification
A causality-based solution to deal with type confounding in ad hoc teamwork
MAGENTA
Multi-Agent Multi-Game Entity Transformer: Towards Generalist Models in MARL
A generalist transformer for Honor of Kings, Starcraft II, and Neural MMO
Towards Skilled Population Curriculum for Multi-Agent Reinforcement Learning
Autocurricula for MARL in complex sparse-reward environments like Google Football