Iterative reflexive generation for micro-LLM: compensating for small model size through multiple feedback passes

An Iterative Reflexive Generation (IRG) approach for small-size language models (micro-LLM) is proposed. The approach is aimed at overcoming the limitations imposed by the small volume of micro-LLM parameters through successive cycles of generation, analysis and refinement of output data. Micro-LLM performs rough response generation; then the same model instance or specialized simple mechanisms analyze the generated text for compliance with the task, format, logical integrity, and identify weaknesses.; Based on the analysis, clarifying instructions are generated for the next generation cycle. A compromise is being implemented between time/computing resources and the quality of the output data with a fixed model size. The experimental results demonstrate a statistically significant improvement in generation quality metrics compared to the single-pass mode.