Tingji Huang @ LAMDA

黄廷基
Tingji Huang
M.Sc Student, LAMDA Group
School of Artificial Intelligence
National Key Laboratory for Novel Software Technology
Nanjing University, Nanjing 210023, China

Email: huangtj@lamda.nju.edu.cn

Laboratory: Room A304, Shaoyifu Building, Nanjing University Xianlin Campus

• Yi-Kai Zhang, Ting-Ji Huang, Yao-Xiang Ding, De-Chuan Zhan, Han-Jia Ye. Model Spider: Learning to Rank Pre-Trained Models Efficiently. [Spotlight] In: Advances in Neural Information Processing Systems 36 (NeurIPS'23), New Orleans, 2023. [Paper] [Code]

We propose Model Spider, which tokenizes pre-trained models (PTMs) and tasks to enable efficient PTM selection. Leveraging PTMs' approximated performance on historical tasks, Model Spider learns to rank with model-task pairs and generalizes to new downstream tasks. PTM-specific task tokens further improves PTM selection. Model Spider balances efficiency and selection ability, demonstrating promising performance in various configurations of model zoos.

• Ting-Ji Huang, Qi-Le Zhou, Han-Jia Ye, De-Chuan Zhan. Change Point Detection via Synthetic Signals. Advanced Analytics and Learning on Temporal Data - 8th ECML-PKDD Workshop (AALTD@ECML-PKDD'23), Turin, Italy, 2023. [Paper] [Code]

we introduce a novel approach to change point detection that eliminates the requirement for collecting supervised data. We train a discriminant model using artificially generated synthetic signals comprising a combination of intricate patterns and random noise, which is designed to predict the number of change points. Experimental results demonstrate the superiority of our Detection Model via Synthetic Signals (DMSS) on the Human Activity Segmentation dataset.

• Ting-Ji Huang, Xu-Yang Chen, Han-Jia Ye. SeqFusion: Sequential Fusion of Pre-Trained Models for Zero-Shot Time-Series Forecasting. arXiv:2503.02836, 2025. [Arxiv] [Code]

In this paper, we introduce SeqFusion in this work, a novel framework that collects and fuses diverse pre-trained models (PTMs) sequentially for zero-shot forecasting. Instead of collecting diverse pre-training data, SeqFusion selects the most suitable PTMs from a batch of pre-collected PTMs based on the specific temporal characteristics of the target time series, performs sequential predictions, and fuses all the predictions while using minimal data to protect privacy.

• Ting-Ji Huang, Jia-Qi Yang, Chunxu Shen, Kai-Qi Liu, De-Chuan Zhan, Han-Jia Ye. Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens. arXiv:2406.08477, 2024. [Arxiv]

In this paper, we explore how to effectively tokenize users and items in LLM-based recommender systems. We emphasize the role of out-of-vocabulary (OOV) tokens in addition to the in-vocabulary ones and claim the memorization of OOV tokens that capture correlations of users/items as well as diversity of OOV tokens.