Leveraging Prompt Tuning-Based Cognitive Attention to Enhance Logical Inference in Large Language Models

Xiaoyan Li, Cuicui Jiang

Abstract: Large language models (LLMs) have demonstrated remarkable problem-solving abilities, but the impact of attention on their log- ical reasoning capabilities remains underexplored. This study in- vestigates the intersection of cognitive neuroscience and LLMs, focusing on prompt fine-tuning techniques to analyze how human- like cognitive abilities and disabilities affect the problem-solving performance of these models. Two GPT-4 based models were devel- oped using prompt fine-tuning and retrieval-augmented generation (RAG). The models were evaluated using the Criteria Cognitive Ap- titude Test (CCAT) dataset, which assesses cognitive abilities such as problem-solving, critical thinking, and information processing. Results showed that the prompt-tuned GPT-4 model achieved the highest accuracy (81.2%), while the model lacking attention per- formed poorly on questions requiring long-term inference. GPT-4’s analysis highlighted the importance of attention in solving prob- lems that demand long-term reference and identified the deficien- cies in the attention-deficient model. This study sheds light on the mechanisms of problem-solving in the brain and the potential of AI to approximate human-like cognition, paving the way for future research at the intersection of cognitive neuroscience and artificial intelligence.