Junjie Qiu

Memory modeling / Multimodal / Physical AI

I’m a first-year Ph.D. student at HKUST (Guangzhou), advised by Prof. Yuxuan Liang. I study memory modeling for physical AI.

I’m interested in how past experiences shape an agent’s future behavior and decisions through memory, and how we can shape that influence. My work spans temporal pretraining, multimodal video generation, and memory mechanisms for long-form storytelling.

CV · English 中文简历 Scholar GitHub Email

Research interests

Technical question

Memory modeling for physical AI

I study how past experiences shape an agent’s future behavior and decisions through memory, and how we can shape that influence.

Research and publications

Societal question

Beyond the cycle of success and failure

How can people with different personalities and histories of success and failure sustain the ability to face reality, take the actions a successful person would take, and improve their lives, without depending on exceptional willpower or endurance?

Read the full reflection

Publications

FactoST-v2 FactoST-v2 architecture overview

Preprint · 2026

Learning to Factorize and Adapt: A Versatile Approach Toward Universal Spatio-Temporal Foundation Models

Siru Zhong, Junjie Qiu, Yangyu Wu, Yiqiu Liu, Yuanpeng He, Zhongwen Rao, Bin Yang, Chenjuan Guo, Hao Xu, Yuxuan Liang

A unified temporal backbone for variable-length context and probabilistic forecasting, with full pretrained weight reuse during downstream adaptation.

arXiv

I designed Universal Temporal Pretraining (UTP) and ran its pretraining, cross-domain evaluation, and ablation studies.

Method & evaluation

The encoder-only quantile model uses normalized sequence patches and randomly masked history prefixes, with gated attention and partial rotary positional encoding. It was configured for up to 2,048 input steps and 256 forecast steps; fixed-weight zero-shot evaluations used 12→12 and 96→96 context/forecast settings on benchmarks excluded from pretraining.

On PEMS08 at 96→96, the model with history masking achieved an MAE of 91.09 versus 107.22 without masking, a 15% relative reduction. See Tables IX–X in the paper.

FactoST FactoST architecture overview

NeurIPS 2025Spotlight

Learning to Factorize Spatio-Temporal Foundation Models

Siru Zhong, Junjie Qiu, Yangyu Wu, Xingchen Zou, Zhongwen Rao, Bin Yang, Chenjuan Guo, Hao Xu, Yuxuan Liang

Separating transferable temporal knowledge from domain-specific dependencies to make foundation models easier to adapt.

Paper Poster

I evaluated cross-domain transfer through zero-shot and few-shot forecasting, assessing how pretrained temporal representations adapt to new datasets.

Experience

Mar. 2026— Sep. 2026

ByFusion · LuckyShort

Architect

Led video-agent architecture for AI short-drama production, connecting long-form planning, narrative memory, controllable generation, and feedback on visual consistency.

Aug. 2025— Aug. 2026

Hong Kong University of Science and Technology (Guangzhou)

Research Assistant

Designed and pretrained temporal models for cross-domain forecasting, with variable-length context, quantile prediction, and efficient downstream adaptation.

Mar. 2025— Dec. 2025

Huawei · 2012 Laboratories

Research Intern

Adapted pretrained temporal models to Ascend and evaluated them against reproduced baselines on telecommunications and power forecasting datasets.

Mar. 2024— Jul. 2024

SUSTech · Center for Computational Science and Engineering

Intern · High-Performance Computing

Benchmarked and commissioned HPC nodes, then used MPI profiling to optimize the interconnect topology of four newly installed nodes, reducing MPI communication time by approximately 20%.

Projects

Jul. – Sep. 2026

Agentic Video Generation · Cross-Shot Consistency

ByFusion · LuckyShort

Built a reference-conditioned video workflow that checks character and scene continuity across shots, regenerating inconsistent outputs while retaining accepted frames.

How it works

Shared reference assets and multimodal checks help identify character and scene drift. Feedback is routed to regenerate the affected keyframes or shots. Creative direction is managed across generation stages and refined using final-video feedback, with model priors used to warm-start candidate search.

Jan. – Mar. 2026

Rare-Earth Pricing · Action-Conditioned Replanning

China Mobile Research Institute · China Rare Earth Group

Contributed to method design and experiments for long-horizon price modeling, conditioning forecasts on geopolitical and supply-chain observations and pricing actions to compare market scenarios and support replanning.

May – Aug. 2025

Tencent Cloud · Behavior-Aware Resource Allocation

Tencent Rhino-Bird Fund

Contributed to temporal-modeling method design for cloud resource allocation, using user behavior and historical allocation sequences as context for compute assignment decisions.

Apr. 2024

ASC24 · LLM Inference Infrastructure

Student Cluster Competition

Built a four-node, eight-GPU inference system with vLLM, Ray, and RDMA, dispatching requests to idle model replicas to balance workloads under fixed hardware constraints.

10th overall among 300+ teams · First Prize & Group Competition Award

Education

Sep. 2026 — Present

Hong Kong University of Science and Technology (Guangzhou)

Ph.D. in Intelligent Transportation

Advised by Prof. Yuxuan Liang

Sep. 2021 — Jul. 2025

Southern University of Science and Technology

B.S. in Data Science and Big Data Technology

GPA 3.89/4.00 · Rank 4/53 (top 10%)
Outstanding Graduate; Outstanding Undergraduate Thesis; Top Ten Graduate Award, Shude Residential College