AGENTS · LEARNING · MEMORY

Learning from experience.
Building better agents.

I'm Haoting Shi, a Computer Science undergraduate at Shanghai Jiao Tong University. I work on computer-use agents, reinforcement learning, and agent memory—exploring how agents use tools, build on experience, and improve through interaction.

My work spans data synthesis, training pipelines, interactive environments, and evaluation systems, connecting agent research with reproducible engineering.

Computer-use agentsReinforcement learningAgent memoryLLM post-training

Selected projects

01 / RESEARCH

Pixel is All You Need

PIXEL-SPACE REASONING

Reading, reasoning, and writing entirely in pixel space

Studies purely pixel-space generative models without text tokens, tokenizers, or text encoders. With images as both input and output, the models write reasoning steps and answers on a canvas, exploring scaling, frame-by-frame chain of thought, and reinforcement-learning post-training.

Project ↗Repository ↗Paper PDF ↗

MemQ

MEMORY + RL

Self-evolving agent memory with Q-learning and provenance DAGs

Built an RL-based memory pipeline combining episodic memory, Q-learning, skill retrieval, and evaluation harnesses to support improvement during agent execution.

Repository ↗arXiv ↗

CUA-Universe

GUI + CLI

A scalable, dynamic environment for hybrid GUI + CLI agents

Led end-to-end environment development, spanning task generation, data collection, training, and evaluation for agents that coordinate graphical interfaces and command-line tools.

arXiv ↗

FedGUI

FEDERATED LEARNING

Federated GUI agents across platforms, devices, and operating systems

Developed data and evaluation pipelines for cross-platform federated GUI agents, including multimodal preprocessing, synthetic interactions, heterogeneous evaluation, and experiment analysis.

Repository ↗arXiv ↗

MCP-Persona

TOOLS + EVALUATION

Evaluating agents on personal applications through environment simulation

Designed benchmark tasks, evaluation protocols, and analysis workflows for personalized MCP-based agent-tool interactions, with a focus on reliable evaluation.

Repository ↗arXiv ↗

Education & honors

02 / BACKGROUND

Shanghai Jiao Tong University

B.S. in Computer Science

Sep 2024 — Jun 2028 (expected)

  • National Scholarship
  • SJTU Merit Student
  • MCM · Meritorious Winner