Haolin Yang

Haolin Yang

Master Student, University of Chicago

Research

In-Context Learning as Semantic Clustering plus Fuzzy Copy

Abstract: We consider in-context learning (ICL) as essentially a two-staged process which features a semantic clustering in earlier layers of Large Language Models (LLMs) based on the semantic properties of the in-context demonstrations with different labels, and subsequently a fuzzy copy stage starting from the intermediate and later layers of the models where they develop an increasingly accurate semantic grasp of the in context demonstrations, as well as the relationship between the final query and these demonstrations. We provide evidence that these two phases of ICL can be respectively attributed to different components of the model, and illustrate how the mechanisms which govern the functions of these components could explain various phenomena found in previous studies as concerning the patterns in the layer-wise representations of input sequences found by models under the icl setting.

COVID Impacts on Global Waste Paper Trade: Based on Complex Network Theory

Abstract: This article focuses on China’s two waste paper import restriction policies which came into effect respectively in 2018 and 2021 and investigates how they reshaped the global waste paper trade network from 2017 to 2022, before wrapping our study up with a prediction of the international waste paper trade in 2025. Our findings are: First, the second restriction in particular had a negative influence on the overall connectivity of the global waste paper trade network; Second, the two import bans caused the network to exhibit an increasingly modular structure, in which countries in the network tended to form several distinct communities and trade actively only inside their communities; Finally, they resulted in the marginalization of China in the network and the collapse of the trade community once led by China in the network, and the power vacuum that consequently appeared was filled by the rise of other countries, primarily India, the United States, and several European countries.

How did COVID-19 and the COVID-related policies influence the international energy trade network: Based on complex network theory and regression

Abstract: This article investigates how the global energy trade network has changed with respect to the outbreak of COVID-19 and the policies taken by the countries in the network concerning the pandemic. This paper applies the technique of social network analysis to study the overall structure and features of the global energy trade network between 2019 and 2022. Then, this paper discusses the changes in the relative importance of different countries in the network and examines quantitatively to what extent could those changes be accredited to different COVID-related factors and policies specific to these countries. The findings are: (1)The interconnectedness and the density of connections in the trade network drastically declined in 2020, but then rebounded in 2021 and 2022. (2) The United States, China, India, and Netherlands were consistently the four most important countries in the global energy trade network from 2019 to 2022, whilst the relative importance of the remaining countries changed rapidly. (3)The number of COVID-19 deaths only has a minor influence on a country’s relative importance in the energy trade network. (4)The stringency of the containment and closure policies a country implements as well as the economic support policies it enacts could significantly affect its relative importance in the network. (5)Developed and developing countries, as well as energy-importing and exporting countries, are affected by the pandemic in largely the same ways.