Logic Consistency Makes Large Language Models Personalized Reasoning Teachers
Abstract: Large Language Models (LLMs) have advanced natural language processing significantly with Chain-of-Thought (CoT) reasoning and In-Context Learning (ICL), but their deployment is limited by high computational and operational costs. This paper introduces Personalized Chain-of-Thought Distillation (PeCoTD), a novel approach to transfer reasoning capabilities from LLMs to smaller, more deployable models. Recognizing the comprehension difficulties small LMs face with LLM-generated rationales, we first develop a metric called Self Logic Consistency (SLC) to assess rationale quality. This refinement process ensures the maintenance of semantic equivalence with the original LLM rationales, facilitating more effective fine-tuning and avoiding distribution shifts.
This approach, focusing on data quality in Knowledge Distillation (KD), mitigates comprehension variability in small LMs and extends the applicability of CoT KD strategies.
Our experiments show that PeCoTD significantly improves the reasoning abilities of small models across diverse datasets.
Paper Type: Long
Research Area: Efficient/Low-Resource Methods for NLP
Research Area Keywords: Large Language Model; Chain-of-Thought;Knowledge Distillation;
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency
Languages Studied: English
Submission Number: 5024
Loading