<ul data-eligibleForWebStory="true"><li>Chain of Thought (CoT) has been utilized in large language models (LLMs) to enhance output quality.</li><li>Recent studies have shown that transformers have limits in expressive power but can effectively solve complex problems when coupled with CoT.</li><li>Existing works on CoT rely on assumptions like identical training and testing data distributions and corruption-free training data, which may not hold in real-world scenarios.</li><li>A new study is the first to rigorously examine the negative effects of data shifts on CoT performance, especially focusing on the $k$-parity problem.</li><li>The study explores the joint impact of distribution shifts and data poisoning on models trained using CoT decomposition.</li><li>Surprisingly, the research indicates that CoT can lead to decreased performance on learning parity compared to direct prediction generation.</li><li>The technical findings offer a detailed explanation for the reasons behind this phenomenon.</li><li>Title: Data Shifts Hurt CoT: A Theoretical Study</li></ul>

Data Shifts Hurt CoT: A Theoretical Study

Discover more