2

The Importance of Prompt Tuning for Automated Neuron Explanations

Recent advances have greatly increased the capabilities of large language models (LLMs), but our understanding of the models and their safety has not progressed as fast. In this paper we aim to understand LLMs deeper by studying their individual neurons. We build upon previous work showing large language models such as GPT-4 can be useful in explaining what each neuron in a language model does. Specifically, we analyze the effect of the prompt used to generate explanations and show that reformatting the explanation prompt in a more natural way can significantly improve neuron explanation quality and greatly reduce computational cost. We demonstrate the effects of our new prompts in three different ways, incorporating both automated and human evaluations.

COVID-19 Increased Censorship Circumvention And Access To Sensitive Topics In China

We study the impact of crisis on information seeking in authoritarian regimes. Using digital trace data from China during the COVID-19 crisis, we show that crisis motivates citizens to seek out crisis-related information, which subsequently exposes them to unrelated and potentially regime-damaging information. This gateway to both current and historically sensitive content is not found for individuals in countries without extensive online censorship. While information seeking increases during crisis under all forms of governance, the added gateway to previously unknown and sensitive content is disproportionate in authoritarian contexts.

The Effect of Streaming Chat on Perceptions of Debates

Broadcast political media is becoming more social. Technology has advanced to the point where now many online video “livestreams” come with embedded livestreaming chatboxes, uniting the on-screen and social components into one realtime, integrated …

Using Facebook Data to Predict the 2016 U.S. Presidential Election

We use 19 billion likes on the Facebook posts of U.S. fan pages from Jan 2015 to Nov 2016 to measure the dynamic ideological positions for politicians, media and users at the state and local level. We then use these measures to derive support rates …

Connective Effervescence and Streaming Chat During Political Debates

Recent advancements in online streaming technologies have re-centered the audience as an important part of live broadcasts, including live political events. In fall 2020, each of the U.S. presidential and vice presidential debates were streamed on a …