Kuan Zhang(张宽)
About Me
I am a Ph.D. Candidate at the College of AI, Tsinghua University.
My research focuses on general game agents and general game world models. I am now working with Prof. Yiming Li at THUSI-Lab on building agents that can play and reason across diverse games, and I am currently a Research Intern at miHoYo (HoYoverse).
Previous: B.Eng. in Software Engineering at Beijing Institute of Technology (BIT); BIT-DataLab with Prof. Chengliang Chai — label noise learning and pretraining data selection.
Outside of research, I enjoy esports, AAA games, music, novels, and animation. Proud Master-tier League of Legends player. 🎮
Education
College of AI, Tsinghua University
2026 – PresentPh.D. Candidate · Intern in THUSI-Lab
Beijing Institute of Technology
2022 – 2026B.Eng. in Software Engineering · Intern in BIT-DataLab
Chengdu No.7 High School
2019 – 2022High School
Research Experience
miHoYo (HoYoverse)
Mar 2026 – PresentResearch Intern
Publications
* denotes equal contribution, † denotes corresponding author
SnapBench: Benchmarking Snap-and-Ask Multimodal Retrieval for Mobile Interactions
The first paired benchmark for robust snap-and-ask multimodal retrieval, with 1,145 queries and 9,085 gallery items under 53 controlled corruption conditions. Image artifacts substantially degrade retrieval, while coarse user text can drag down joint retrieval; we propose MOOR for reliability-aware modality calibration.
Towards Spatial Supersensing in the Wild
A spatial supersensing benchmark of 442 in-the-wild long-form videos and 6,980 QA pairs, showing that even strong multimodal models fail at coherent world-state tracking.
Towards Generalist Game Players: An Investigation of Foundation Models in the Game Multiverse
A systematic survey tracing the full lifecycle of generalist game players across four interdependent pillars — Dataset, Model, Harness, and Benchmark — and charting a five-level roadmap from single-game mastery toward the ultimate creator stage in the game multiverse.
GameVerse: Can Vision-Language Models Learn from Video-based Reflection?
A comprehensive video game benchmark enabling a reflective visual interaction loop for VLMs, with cognitive hierarchical taxonomy spanning 15 globally popular games, dual action space, and milestone evaluation.
Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization
An efficient, hyperparameter-free, instance-level optimization framework for label noise in image classification.
Harnessing Diversity for Important Data Selection in Pretraining Large Language Models
Combine multi-arm bandit with influence function to select data for pretraining process of large language models.