Official Review · ICLR 2024

This paper quantitatively investigates in-context learning in terms of surface patterns. It shows that there is an inherent correlation between surface patterns, self-reinforcement, iterative generation and their important role in text generation. In particular, it shows the role of surface surface patterns in text generation and the existence of token co-occurrence reinforcement that strengthens the relationship between two tokens based on their contextual co-occurrences. The experiments on MMLU and GSM8K show that the reinforcement helps to constrain the output space and format the output according to a demonstration such as ‘Let’s think step by step’ . This paper provide a novel framework to understanding in-context learning via the notion of token co-occurrence reinforcement. Through various experiments, the authors have shown how token reinforcement causes spurious correlations in in-context learning. Althogh there is a novelty in showing experimentally that token reinforcement can cause some problems in in-context learning, the paper lacks the important perspective of analyzing why token reinforcement exists and causes problems. For example, the following paper, which is only briefly mentioned in this paper, analyzes the impact of repetition structures in a corpus on in-context learning from an information-theoretic perspective. A Theory of Emergent In-Context Learning as Implicit Structure Induction Michael Hahn, Navin Goyal They showed that the performance of in-context learning is represented by a complexity that repetition structures can be represented by a small PCFG tree, and experimentally investigated theoretical finindings. Is it possible to provide hypothesis about the reason for the token reinforcement phenomenon and check it, even experimentally? The authors may argue that analyzing reasons is the next step, i.e., outside the scope of this paper; however, it is fandamental in machine learning research.