About me

Hi, I'm Guannan Lai.

I am a master's student at the School of Artificial Intelligence, Nanjing University, and a member of the LAMDA Group, supervised by Prof. Han-Jia Ye.

My research focuses on reliable and efficient model reuse. I study how learning systems adapt continuously in dynamic environments, and how they select and reuse suitable capabilities from existing models. My recent work spans LLM routing, continual learning, test-time adaptation, and lifelong agents under resource constraints.

Continual Learning Model Reuse Test-Time Adaptation LLM Routing Lifelong Agents

01From research to tools

Three entry points into my routing research: an evaluation platform, an open-source toolbox, and a multimodal benchmark.

Evaluate / User preferences

RouteJudge

Evaluate LLM routers through the lens of different user preferences, rather than a single shared priority.

RouteJudge evaluation framework

Experiment / Open source

ORBIT

A toolbox for standardized LLM routing research and budget-aware evaluation, bringing routing experiments into a common framework.

ORBIT toolbox framework

Benchmark / Multimodal

MMR-Bench

A benchmark for studying how to route among multimodal LLMs, with a public dataset and code for further research.

MMR-Bench overview

News

02Research directions

I study reliable and efficient model reuse under changing tasks and limited resources. My earlier work helps individual models adapt through continual learning and test-time adaptation; my recent work explores how LLM routing can select, reuse, and reliably evaluate capabilities across multiple models.

Adaptive Model Reuse

Continual learning and test-time adaptation for individual models under changing tasks and environments.

Routing and Evaluation

Capability selection, routing supervision, and reliable evaluation of multiple models under resource constraints and user preferences.

03Selected publications

Conference Papers

EquiRouter query-model interaction encoder and cost-aware ranking framework NeurIPS'26

When LLM Routers Overpay: The Cost of Strong-Model Preference

Guannan Lai, Long Chen, Han-Jia Ye

NeurIPS 2026

Examining why LLM routers systematically favor stronger, more expensive models, and how decision-aware routing can reduce unnecessary over-selection while preserving performance.

DARS teaser EMNLP'26

From Sampled Outcomes to Capability Distributions: Rethinking Supervision for LLM Routing

Guannan Lai, Haoran Hu, Long Chen, Zhenguo Li, Han-Jia Ye

EMNLP 2026

Rethinking LLM routing supervision through capability distributions rather than individual sampled outcomes.

MMR-Bench teaser ECCV'26

MMR-Bench: A Comprehensive Benchmark for Multimodal LLM Routing

Haoxuan Ma*, Guannan Lai*, Han-Jia Ye

ECCV 2026

A comprehensive benchmark for evaluating routing among multimodal large language models.

* Equal contribution.

GOLD teaser CVPR'26

The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation

Guannan Lai, Da-Wei Zhou, Zhenguo Li, Han-Jia Ye

CVPR 2026

Subspace-based continual test-time adaptation that brings efficiency and generalization together.

EDGE teaser ICLR'26

The Lie of the Average: How Class Incremental Learning Evaluation Deceives You?

Guannan Lai, Da-Wei Zhou, Xin Yang, Han-Jia Ye

ICLR 2026

Examining what average performance can conceal in class-incremental learning evaluation.

GDDSG teaser CVPR'25

Order-robust Class Incremental Learning: Graph-driven Dynamic Similarity Grouping

Guannan Lai, Yujie Li, Xiangkun Wang, Junbo Zhang, Tianrui Li, Xin Yang

CVPR 2025

Graph-driven dynamic similarity grouping for class-incremental learning robust to class order.

Workshop Papers

TokenWorth method framework NeurIPS-W'26

Learning What to Remember: Resource-Efficient Self-Improvement for Lifelong Agents

Guannan Lai, Zi-Yang Bai, Jun-Peng Jiang, Han-Jia Ye

NeurIPS 2026 Workshop on Towards Test-Time Continual Learning Agents

TokenWorth learns query-dependent memory worth from sparse execution feedback, guiding experience retrieval and bounded memory management for lifelong agents under constrained generation, interaction, and memory budgets.

RouteJudge teaser ICMLW'26

RouteJudge: Preference-Based Evaluation of LLM Routers under Pluralistic User Preferences

Guannan Lai, Haoran Hu, Han-Jia Ye

Pluralistic Alignment Workshop at ICML 2026

Preference-based evaluation of LLM routers that accounts for diverse user priorities.

Journal Papers

ORBIT framework FCS'26

ORBIT: An Optimal Routing and Budgeted Inference Toolbox

Guannan Lai, Haoran Hu, Hao-Xuan Ma, Han-Jia Ye

Frontiers of Computer Science, 2026

ORBIT is an open-source toolbox for standardized LLM routing research and budget-aware evaluation.

Preprints

RouteFM method framework arXiv'26

Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing

Guannan Lai, Han-Jia Ye

Preprint, 2026

RouteFM learns reusable routing capabilities that transfer across tasks, candidate models, modalities, and deployment conditions through behavioral context.

SaveRouter method framework arXiv'26

Routing Should Pay for Itself: Sparse Supervision for Economical LLM Routing

Guannan Lai, Gelin Bian, Hao-Xuan Ma, Jun-Peng Jiang, Long Chen, Jian-Dong Liu, Zhi-Hao Tan, Han-Jia Ye

Preprint, 2026

SaveRouter selectively acquires informative model feedback and evaluates whether serving-time savings recover the upfront cost of router supervision.

Experience & Education

  • 2026-08
    Research Internship · SinapisAI

    Started research on learnware and LLM routing.

  • 2025 – present
    M.Sc. · Nanjing University

    School of Artificial Intelligence, LAMDA Group. Supervised by Prof. Han-Jia Ye; admitted without entrance examination.

  • 2025-06
    B.Sc. · Southwestern University of Finance and Economics

    Intelligent Finance.

Awards & Honors

  • 2026, National Scholarship for Graduate Students
  • 2025, Outstanding Graduate of Sichuan Province
  • 2024, Bronze Medal, Huawei Cup 2023 ICPC Asia-East Continent Final Contest
  • 2024, Bronze Medal, CCPC Final 2023
  • 2023, Silver Medal, ICPC Asia Hangzhou Regional Contest
  • 2023, Gold Medal, Sichuan Province College Student Programming Competition
  • 2023, ACM Chengdu Chapter Outstanding College Student Award
  • 2022, 2023, 2024, National Scholarship for Undergraduates

Academic Service

  • Conference Reviewer: ICLR, ICML, NeurIPS, CVPR, ACM SIGKDD, COLM
  • Journal Reviewer: IEEE TPAMI, IJCV, Pattern Recognition, TMLR

Writing & Notes

Contact

Email
laign (at) lamda.nju.edu.cn
Address
Guannan Lai, National Key Laboratory for Novel Software Technology, Nanjing University, Xianlin Campus Mailbox 603, 163 Xianlin Avenue, Qixia District, Nanjing 210023, China.
中文地址
南京市栖霞区仙林大道 163 号,南京大学仙林校区 603 信箱,软件新技术国家重点实验室,210023。