🎓 Computer Science Undergraduate @ HUST
🤖 Deputy Team Leader @ Unique Studio · 联创团队

Learning to model the world, one question at a time.

I'm Haoran Qian, a Computer Science undergraduate at Huazhong University of Science and Technology (HUST) and Deputy Team Leader of Unique Studio. Since June 2026, I have also been doing research at Shanghai Jiao Tong University, advised by Zhipeng Zhang.

My interests span world models, generative models, video generation, multimodal learning, and embodied intelligence. I'm especially curious about how models represent change over time, respond to actions, and generate futures that are useful for decision-making.

I learn by reading papers, tracing implementations, and running experiments. I also keep notes on failed hypotheses and negative results: understanding why an idea does not work is part of the research.

• World Models & Video Generation — Action-conditioned dynamics, controllable video prediction, and reliable long-horizon rollouts.
• Embodied Intelligence — Connecting visual world models with action, planning, and policy learning.
• Generative & Multimodal Learning — Learning representations that connect perception, generation, and interaction.
• LLMs & Agents — Tool use, ReAct-style workflows, and parameter-efficient fine-tuning.

Co-author · ICLR 2027 — in submission

Exploring reinforcement learning for more reliable action-conditioned world models and stable embodied simulation.

See my publications page for the current record.

A growing collection of Chinese paper notes, implementation walkthroughs, and research questions on world models. Topics include the Dreamer and LingBot series, action conditioning, embodied world models, and evaluation.

A research study of when action-conditioned world models justify specialist models and distillation. It documents response probes, baseline diagnostics, and negative results across the tested ACWM-Phys environments.

Experiments with ReAct-format data generation and LoRA / QLoRA fine-tuning for language-model tool use.

Learning Transformer and LLM implementations through code, with experiments on training behavior, batch size, loss, and learning rate.

• Research: Python, PyTorch, Transformer models, world-model experiments.
• Adaptation & agents: LoRA / QLoRA, ReAct-style tool use, data preparation.
• Workflow: Git, implementation reading, experiment analysis, and research documentation.

• Research at Shanghai Jiao Tong University · June 2026–present · Advised by Zhipeng Zhang.
• Deputy Team Leader, Unique Studio @ HUST · Learning and building with a student-led technology community.
• National First Prize, RAICOM Robotics Developer Competition · Team award, August 2026.
• Computer Science, HUST · Undergraduate, 2025–present.

• Homepage: aethernoah.github.io
• Email: hrqian06@gmail.com
• Beyond research: Writing, photography, exploring unfamiliar cities, and keeping small moments in Fragments / 浮光.

Read carefully. Build patiently. Keep asking questions.