CoT-ICL Lab: A Synthetic Framework for Studying Chain-of-Thought Learning from In-Context Demonstrations

TL;DR

CoT-ICL Lab provides a synthetic framework to systematically study how chain-of-thought in-context learning improves language model performance, revealing the roles of model depth, example quantity, and diversity in training.

Contribution

It introduces a controllable synthetic dataset framework for analyzing CoT-ICL, enabling detailed study of factors influencing in-context learning in language models.

Findings

CoT accelerates accuracy transition across model sizes.

Model depth is crucial for leveraging CoT with few examples.

Limited diversity in training improves causal structure learning.

Abstract

We introduce CoT-ICL Lab, a framework and methodology to generate synthetic tokenized datasets and systematically study chain-of-thought (CoT) in-context learning (ICL) in language models. CoT-ICL Lab allows fine grained control over the complexity of in-context examples by decoupling (1) the causal structure involved in chain token generation from (2) the underlying token processing functions. We train decoder-only transformers (up to 700M parameters) on these datasets and show that CoT accelerates the accuracy transition to higher values across model sizes. In particular, we find that model depth is crucial for leveraging CoT with limited in-context examples, while more examples help shallow models match deeper model performance. Additionally, limiting the diversity of token processing functions throughout training improves causal structure learning via ICL. We also interpret these transitions by analyzing transformer embeddings and attention maps. Overall, CoT-ICL Lab serves as a simple yet powerful testbed for theoretical and empirical insights into ICL and CoT in language models.