Publications

Are Human-generated Demonstrations Necessary for In-context Learning?

Published in ICLR2024, 2024

In this paper, we propose self-contemplation prompting strategy (SEC), a paradigm free from human-crafted demonstrations. The key point of SEC is that, instead of using hand-crafted examples as demonstrations in ICL, SEC asks LLMs to first create demonstrations on their own, based on which the final output is generated.

Recommended citation: Rui Li, Guoyin Wang, and Jiwei Li. ICLR 2024 https://arxiv.org/abs/2309.14681

Similarity-based Neighbor Selection for Graph LLMs

Published in arXiv:2402.03720, 2024

The research introduces Similarity-based Neighbor Selection (SNS) that leverages SimCSE and neighbor selection techniques to enhance graph representations for improved node classification in Text-attributed Graphs (TAGs), demonstrating superior performance over traditional GNNs by addressing challenges like over-squashing and heterophily through LLMs.

Recommended citation: Rui Li, Jiwei Li, Jiawei Han, Guoyin Wang. arXiv:2402.03720 https://arxiv.org/abs/2402.03720