Attention needs intelligence: a human cognitive inspired approach to enhance the reasoning capabilities of large language models

Research on reasoning capabilities of Large Language Models (LLMs) has been significantly improved throughout the past couple of years with the need of achieving human level intelligence in LLMs. Reasoning is the core skill that humans have which makes humans superior to other living beings in the activities of problem solving, critical thinking and decision making. Chain-of-Thoughts (CoT), Reinforcement Learning (RL) in LLM training pipeline and agent scratchpads were some of the approaches taken to pull out the reasoning capabilities of LLMs. Even though these methodologies show promising results, following a human cognitive inspired mechanism for language model reasoning can be seen as a significant gap in the domain, and yet the supervised fine-tuned behavior of these methodologies acts as a barrier for exploration and stochastic behavior of language models. A novel RL framework which aligns human cognitive principles and probabilistic reasoning to the LLM training pipeline will be introduced in this research study with the aim of achieving better reasoning capabilities through LLMs which have small number of parameters. The proposed solution is supplemented by a philosophically inspired novel CoT architecture and a Bayesian rewarding mechanism which gradually leads the LLM to follow probabilistic reasoning behavior. The trained policy model or the base LLM was successfully able to adhere to the introduced CoT structure by maintaining 89.5% accuracy for the gsmk8k and dolphin-r1 mixed test dataset. The reward distribution over training steps evidenced the expected stochastic reasoning pattern of the policy model without reward overoptimization.