CV — Yuan Huang

Publications & Research

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

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

EditJudgeBias · Under review at ICLR 2027 · co-first author · arXiv:2610.01670, Oct 2026

Design. 1,196 real editing samples and 13 cues across four evaluation sites. A shift counts as bias only if the cue is verified to preserve edit quality, and is read against the judge’s own zero-dose and re-query noise floors, not against zero.
Findings. Quality-preserving cues move all five MLLM judges beyond their own noise; swapping candidate order alone reverses up to 60.9% of pairwise decisions, against 1.0–10.5% on re-query.

[2] 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

LS-B · Under review at ICLR 2027 · first author · arXiv:2610.01590, Oct 2026

Placement search over ViT depth traces an inverted U (up to 3.5 pp); LS-B reads early tasks through an fMRI encoding model of twelve human visual areas and picks blocks overlapping the peak — on Split ImageNet-R, within 1.5 pp of full-BiLoRA at 60% of the storage, with no labels or backpropagation.

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

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

WereBench · EMNLP 2026, Main Conference · co-first author · arXiv:2510.11389, Oct 2025 · Dataset

Benchmark built from 100+ hours of professional human social-deduction gameplay (32.4M utterance tokens, 30 roles, 15 rule variants), scored against winning-side MVP trajectories rather than LLM self-play; 5 verbal dimensions and 2 decision tasks.
Half of the evaluated models score below 0.50 (best 0.720). Deception reasoning and counterfactual trade-off are weakest while persuasive generation stays strong — fluency ≠ strategic correctness.

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

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

Geolocation · Preprint · arXiv:2502.13759, Feb 2025 · Dataset · Code

Dataset from a public geo-guessing platform — 740K users, 2.7M locations, 25M human response records — with human accuracy calibrating item difficulty. Multi-step geographic reasoning (GeoCoT) improved classic geolocation metrics by up to 25%, and reasoning quality as scored by GeoEval by 9%.