UIUC Postdoc Chen Xusi Joins CUHK: Pioneering Self-Evolving Agents & LLM Alignment

University of Illinois Urbana-Champaign (UIUC) postdoctoral researcher Chen Xusi has accepted a tenure-track assistant professor position in the Department of Computer Science and Engineering at the Chinese University of Hong Kong (CUHK), starting fall 2027. Chen earned his bachelor's degree from Peking University under Professor Jun Gao, completed his PhD at UCLA advised by Professor Wei Wang, and then joined UIUC's Blender Lab as a postdoc working with NLP scholar Professor Heng Ji.

His overarching research goal is to build automated language agents that can assist humans across multiple domains . The article outlines five main research directions:

LLM Reasoning & Alignment: large-model reasoning, reward modeling, constitutional alignment.

Lifelong Learning & Self-Evolving Agents: lifelong learning, memory, world models.

Multimodal Understanding & Generation: multimodal agents, visual reasoning, cross-modal modeling of text, image, and video.

AI for Science / Scientific LLMs: using LLMs to accelerate scientific discovery, literature understanding, scientific reasoning, and interdisciplinary knowledge organization.

Complex Dynamic Systems & Sports Science: modeling complex dynamic systems, planning and strategy learning in sports, player behavior generation, and offline reinforcement learning.

The article traces a consistent thread through Chen's work: replacing hand-crafted training signals, behavioral rules, and evaluation criteria with the model's own reasoning . Early low-resource reasoning work includes:

Gotta (SDM 2023, Best Poster Honorable Mention): designed a fill-in-the-blank auxiliary task so models learn reasoning incidentally during few-shot QA.

MinPrompt (ACL 2024): used graph algorithms to select a minimal, maximum-coverage sentence set for generating QA pairs, turning data augmentation into an optimization problem.

Recent alignment and self-evolution papers push this idea further:

IterAlign (NAACL 2024): lets the model iteratively discover its own issues and automatically adjust behavioral guidelines instead of relying on fixed human-written rules.

RM-R1 (ICLR 2026): frames reward modeling as a reasoning process — the model first determines what constitutes a good output before judging.

DecisionFlow (EMNLP 2025): constructs a structured decision space, infers a latent utility function, and performs trade-offs within that space, yielding answers and explanations that share the same origin and are therefore inherently interpretable.

A parallel line of work applies the same decision-making framework to sports:

ReLiable (CIKM 2022): uses offline reinforcement learning to learn tactical strategies in professional basketball.

PlayBest (CIKM 2024): employs diffusion models for planning and synthesizing player behavior trajectories.

Though seemingly distant from LLMs, the core is identical — planning and decision-making in complex dynamic environments .

Beyond research, Chen co-organized the SURGeLLM Workshop at ACL 2026 and the Sknow-LLM Workshop at KDD 2025, serves as an Area Chair for ACL Rolling Review, and has been a program committee member for ICML, NeurIPS, ICLR, KDD, and WSDM.