🧭 About

I am a final-year undergraduate in Computer Science and Technology at Northeastern University, China, advised by Prof. Shiqi Zhao since 2024. I have worked with Prof. Xiuying Chen (MBZUAI) since 2025, and am visiting her group from October 2026 to January 2027. I rank 5 / 110 (top 4.55%) in my cohort with a CGPA 4.0481 / 5.00 (90 / 100) and hold the China National Scholarship.

My work on evaluating LLM social reasoning against real human gameplay appears at EMNLP 2026 Main Conference, where I am a co-first author; before that I worked on geolocation from large-scale human gameplay data. Two newer papers are under review at ICLR 2027: EditJudgeBias (co-first author), an audit of the multimodal LLM judges that evaluate image editing, and LS-B (first author), which reads a continual learner through an fMRI model of human visual cortex to decide where it should specialize.

Underneath all of it is one question: whether the benchmarks and automatic judges we rank models with measure what they claim to. I come at it from two sides β€” auditing the judges themselves, and building benchmarks out of real human behavior rather than model self-play β€” and the same result keeps returning. A correct answer is not a correct reasoning process; persuasive language is not strategic correctness; a verdict that looks right is not one that used the right evidence. Surface competence is not underlying competence, and most of our instruments measure only the first. Measuring that gap, explaining it, and closing it is what I want to spend the next several years on.

πŸ”¬ Current focus β†’ Interpretability of bias in multimodal models: identifying the internal semantics that carry a bias, and mechanisms to correct it.

πŸ”₯ News

  • 2026 Β· 10 Joining MBZUAI as a visiting student, advised by Prof. Xiuying Chen.
  • 2026 Β· 10 EditJudgeBias and LS-B posted (arXiv:2610.01670, arXiv:2610.01590); both under review at ICLR 2027.
  • 2026 Β· 08 WereBench accepted to EMNLP 2026 Main Conference.
  • 2025 Β· 10 WereBench preprint posted: arXiv:2510.11389.
  • 2025 Β· 02 Geolocation preprint and the GeoComp dataset released: arXiv:2502.13759.
  • 2024 Awarded the China National Scholarship.

πŸ“„ Publications

EditJudgeBias2026 Β· 10

Do MLLM Judges Judge the Edit? Auditing Bias in Image Editing Evaluation with Verified Quality Preservation

Yuan Huang*, Zirui Song*, Xiuying Chen  (* equal contribution)

Under review at ICLR 2027 · co-first author

PDF arXiv

LS-B2026 Β· 10

Two Routes to the Middle: Placement Search and Brain Readouts Converge on Where Continual Learners Should Specialize

Yuan Huang, Zihan Chen, Runbin Zhang, Hongwei Ding, Changzeng Fu, Shiqi Zhao

Under review at ICLR 2027 · first author

PDF arXiv

WereBench2025 Β· 10

Beyond Survival: Evaluating LLMs in Social Deduction Games with Human-Aligned Strategies

Zirui Song*, Yuan Huang*, Junchang Liu*, et al.  (* equal contribution)

EMNLP 2026, Main Conference · co-first author

PDF arXiv Dataset

Geolocation2025 Β· 02

Geolocation with Real Human Gameplay Data: A Large-Scale Dataset and Human-Like Reasoning Framework

Zirui Song, Jingpu Yang, Yuan Huang, et al.

Preprint

PDF arXiv Dataset Code

πŸŽ“ Education

  • 2026 – 27 Visiting Student, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
    Oct 2026 – Jan 2027 · hosted by Xiuying Chen, Assistant Professor, MBZUAI, my collaborator since 2025
  • 2023 – 27 B.Eng., Computer Science and Technology, Northeastern University, China
    CGPA 4.0481 / 5.00 (90 / 100) · rank 5 / 110 (top 4.55%) · advised by Shiqi Zhao, Assistant Professor, Northeastern University (Brain‑Computer Interface and Hardware Acceleration Lab), since 2024

πŸ† Awards

  • 2023 – 24China National Scholarship (national level)
  • 2024 – 26Northeastern University Comprehensive Scholarship
  • 2024Third Prize, National Finals, Lanqiao Cup Programming Competition

πŸ—£οΈ Why language

Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt.

The limits of my language mean the limits of my world.

ζˆ‘θ―­θ¨€ηš„η•Œι™ζ„ε‘³η€ζˆ‘ηš„δΈ–η•Œηš„η•Œι™γ€‚

Ludwig Wittgenstein, Tractatus Logico-Philosophicus 5.6

I keep that line in view because I read it as a research claim rather than an aphorism. If the limits of a language are the limits of a world, then a system whose entire access to the world runs through language inherits exactly those limits β€” and finding where they fall is a measurement problem, which is the kind of problem I like.

That is why my work sits in natural language processing and why I intend to stay there. Language is at once the object I study, the instrument I have to study it with, and the thing these models are most convincing at; nearly every failure I care about lives in the gap between sounding like understanding and being it.