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

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Kothapalli, Vignesh, Fatahibaarzi, Ata, Firooz, Hamed, Sanjabi, Maziar
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909945643925504
author Kothapalli, Vignesh
Fatahibaarzi, Ata
Firooz, Hamed
Sanjabi, Maziar
author_facet Kothapalli, Vignesh
Fatahibaarzi, Ata
Firooz, Hamed
Sanjabi, Maziar
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2512_05318
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle To Think or Not to Think: The Hidden Cost of Meta-Training with Excessive CoT Examples
Kothapalli, Vignesh
Fatahibaarzi, Ata
Firooz, Hamed
Sanjabi, Maziar
Computation and Language
Artificial Intelligence
Machine Learning
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.
title To Think or Not to Think: The Hidden Cost of Meta-Training with Excessive CoT Examples
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.05318