Transferring Reasoning Capabilities Between LLMs Operating via Curriculum Learning Policy

Abstract

In-context reasoning methods, exemplified by Chain-of-Thought (CoT) (et alia.,) empower the reasoning abilities of large language models (LLMs), eliciting them to solve complex reasoning tasks step-by-step. Nevertheless, the capacities to deliver robust CoT explanations arise only in models with billions of parameters, representing a barrier to entry for many users forced to operate on a smaller model scale, i.e., Small Language Models (SLMs). Even though many companies are releasing LLMs of the same family with a reduced number of parameters, these models sometimes produce misleading answers and are unable to deliver accurate step-wise reasoned answers. This paper proposes a method to transfer step-wise reasoning over SLMs by operating via Instruction-tuning (IT) on synthetic demonstrations delivered in a pedagogically motivated manner. In particular, firstly, we propose aligning step-wise reasoning capabilities via IT using Demonstrations "taught" by LLMs teacher to SLMs students. Then, we operate via Curriculum Learning, a pedagogically motivated learning method that improves the IT phase. We analyse the impact on the downstream performances of four question-answering benchmarks. The results show that SMLs can be instructed to reason via Demonstrations delivered by LLMs. We move a step further in research: conceiving SLMs as human learners, we expose them to a CL teaching-based approach, obtaining better results on downstream performances.