Adaptive Dual-Inference Framework for Self-Training in Large Language Models to Enhance Cognitive Task Performance

The need for improved task performance and adaptability in models designed to handle complex reasoning tasks has driven significant efforts toward more efficient learning mechanisms. The introduction of a dual-inference self-training framework presents a novel approach to enhancing cognitive task performance through iterative self-assessment and refinement without external supervision. This framework, applied to the Llama model, demonstrated the ability to improve accuracy, generalization, and reasoning efficiency across tasks of varying complexity. The dual-inference mechanism allowed the model to engage in recursive refinement cycles, enabling it to reduce errors and optimize output coherence in a manner that traditional fine-tuning techniques do not achieve. Experiments conducted across a range of cognitive challenges revealed substantial improvements in task complexity resolution, reasoning speed, and multi-domain generalization, further showing the framework’s efficacy. The findings suggest that autonomous learning frameworks, such as the one presented here, offer significant advantages in terms of scalability, cost-efficiency, and performance, providing new avenues for the development of intelligent systems capable of handling real-world cognitive tasks.