To Think or Not to Think: The Hidden Cost of Meta-Training with Excessive CoT Examples

TL;DR

This paper investigates the negative impact of excessive chain-of-thought examples during meta-training on large language models' reasoning abilities and proposes a method to optimize the mix of CoT and non-CoT examples to improve performance on novel tasks.

Contribution

It introduces CoT-Recipe, a formal approach to balance CoT and non-CoT examples during meta-training, enhancing reasoning performance on unseen tasks.

Findings

Meta-training with excessive CoT examples can degrade performance.

CoT-Recipe improves accuracy on novel tasks by up to 300%.

Applying techniques to pretrained LLMs yields up to 130% accuracy gains.

Abstract

Chain-of-thought (CoT) prompting combined with few-shot in-context learning (ICL) has unlocked significant reasoning capabilities in large language models (LLMs). However, ICL with CoT examples is ineffective on novel tasks when the pre-training knowledge is insufficient. We study this problem in a controlled setting using the CoT-ICL Lab framework, and propose meta-training techniques to learn novel abstract reasoning tasks in-context. Although CoT examples facilitate reasoning, we noticed that their excessive inclusion during meta-training degrades performance when CoT supervision is limited. To mitigate such behavior, we propose CoT-Recipe, a formal approach to modulate the mix of CoT and non-CoT examples in meta-training sequences. We demonstrate that careful modulation via CoT-Recipe can increase the accuracy of transformers on novel tasks by up to 300% even when there are no CoT examples available in-context. We confirm the broader effectiveness of these techniques by applying them to pretrained LLMs (Qwen2.5 series) for symbolic reasoning tasks and observing gains of up to 130% in accuracy.