Reproduce “Understanding In-Context Learning from Repetitions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jianhao Yan, Jin Xu, Chiyu Song, Chenming Wu, Yafu Li, Yue Zhang
The paper proposes and empirically establishes "token co-occurrence reinforcement" as a fundamental mechanism driving in-context learning (ICL) in large language models. This mechanism explains how surface repetitions in demonstrations both beneficially constrain output space and detrimentally lead to spurious connections and ICL failures.
Key insights
- Token co-occurrence reinforcement is a fundamental principle where the connection between any two tokens strengthens with their contextual co-occurrences.
- ICL can be viewed as a combination of various token-level reinforcements that either cooperate to guide outputs or create conflicts.
- This reinforcement phenomenon is an inherent result of LLMs maximizing likelihood during pre-training on natural language data, which naturally contains repetitions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.