Chain-of-Thought (CoT) engineering has proven to be an effective way to improve the logical reasoning ability of Large Language Models (LLMs). The traditional CoT method prompts the LLM to think about the problem step by step in the input, and the LLM outputs the answer step by step according to the dismantling steps. This process is a chain of thought processes that shows reasoning, which increases the latency and complexity of the LLM, making it slower and slower to respond. Instruction fine-tuning is a technique for optimizing pre-trained models to suit specific tasks. In this paper, we propose an innovative paradigm Pre-Cognitive Inductive Reasoning (PCIR), which transforms the explicit inference of LLM into implicit inference by embedding CoT into the instruction fine-tuning process, to improve the response speed and logical reasoning ability of LLMs. We conducted experiments on three datasets: GSM8K, AQuA, and the Multi-Role Task Dataset (MRTD). Among them, MRTD is the dataset we built to validate the role cognitive ability of LLMs. The results show that the PCIR method is superior to the existing methods. Compared with the traditional CoT methods, the accuracy of GSM8K, AQuA, and MRTD is improved by 1.69%, 4% and 7%, respectively.
Zhang et al. (Fri,) studied this question.