Yi Yang

Yi Yang

Auto-generated profile, based on publicly available information

Yi Yang


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
2 April 2026 in ACM Computing Surveys
Verified
Journal Article
Open access
MZMinxing Zhang
Minxing Zhang
YYYi Yang
Yi Yang
RXRoy Xie
Roy Xie
BDBhuwan Dhingra
Bhuwan Dhingra
SZShuyan Zhou
Shuyan Zhou
JPJian Pei
Jian Pei
Large Language Model (LLM)-based agents have recently emerged as a new paradigm that extends the capabilities of LLMs beyond text generation to dynamic interaction with external environments. A critical challenge lies in ensuring their g eneralizability – the ability to maintain consistently high performance across varied instructions, tasks, environments, and domains, especially those different from the agent’s fine-tuning data. Despite growing interest, the concept of generalizability in LLM-based agents remains underdefined, and systematic approaches to measure and improve it are lacking. We provide the first comprehensive review of generalizability in LLM-based agents. We begin by clarifying the definition and boundaries of agent generalizability. We then review existing benchmarks. Next, we categorize strategies for improving generalizability into three groups: methods targeting the backbone LLM, targeting agent components, and targeting their interactions. Furthermore, we introduce the distinction between g eneralizable frameworks and g eneralizable agents and outline how generalizable frameworks can be translated into agent-level generalizability. Finally, we identify future directions, including the development of standardized evaluation frameworks, variance- and cost-based metrics, and hybrid approaches that integrate methodological innovations with agent architecture-level designs. We aim to establish a foundation for principled research on building LLM-based agents that generalize reliably across diverse real-world applications.

8 August 2026
Verified
Conference Paper
Open access
MZMinxing Zhang
Minxing Zhang
YYYi Yang
Yi Yang
ZJZhuofan Jia
Zhuofan Jia
XYXuan Yang
Xuan Yang
JPJian Pei
Jian Pei
XCXianglong Chen
Xianglong Chen

21 August 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
MZMinxing Zhang
Minxing Zhang
YYYi Yang
Yi Yang
JPJian Pei
Jian Pei
Large language models are rapidly evolving from general-purpose assistants into personalized AI systems that collaborate with individuals over long periods of time. Emerging applications increasingly seek not only to answer questions, but also to reason, communicate, and act in ways that faithfully reflect a particular person. Recent systems have begun to distill personal knowledge, workplace experience, communication history, and behavioral traces into AI assistants, skill libraries, memory systems, and personalized agents. Despite this rapid progress, the field remains fragmented. Existing work differs substantially in the evidence it uses, the aspects of a person it preserves, the representations it constructs, the methods it employs, and the way it evaluates fidelity, leaving no common conceptual foundation for person distillation. This paper presents the first comprehensive framework for person distillation: the problem of transforming heterogeneous traces of a particular real person into faithful, evidence-grounded, and revisable computational representations. We organize the field as an end-to-end lifecycle consisting of four tightly coupled components: source evidence, distilled person representations, distillation methods, and evaluation. For each component, we develop a taxonomy that unifies existing research and identifies open design choices. We further organize the methodological landscape into six complementary families, ranging from prompt-based trace-to-profile distillation to memory-based, parametric and adapter-based, preference- and reward-based, behavioral and trajectory-based, and hybrid and revisable distillation. Finally, we propose a comprehensive evaluation framework covering person-fidelity objectives, benchmark construction protocols, quantitative metrics, and diagnostic analyses for measuring evidence grounding, person specificity, boundary awareness, and evaluation reliability. Beyond surveying existing work, this paper establishes a common vocabulary, a conceptual framework, and a research agenda for person distillation. By connecting advances in role-playing LLMs, personalized generation, agent memory, skill learning, and preference modeling, we identify the fundamental challenges that distinguish modeling a particular real person from modeling generic users, roles, or personas. As AI systems become increasingly personalized, collaborative, and long-lived, we believe person distillation will become a foundational capability for the next generation of AI assistants, digital coworkers, and human-centered intelligent systems.