Recursive Self-Improving LLM Agents with Meta-Cognitive Feedback Loops
Even though AI is getting better quickly, Large Language Models (LLMs) are still very good at multi-step reasoning and structured problem solving. However, their performance during inference often relies heavily on initial prompts and set strategies. When outputs are not ideal, fixing them usually requires outside help instead of changing internal strategies. This paper presents a self-improving recursive LLM agent framework that improves reasoning performance without retraining the model. The system works through a meta-cognitive feedback loop with three parts: an Executor, a Critic, and a Planner. The Executor creates task responses using an initial reasoning strategy. The Critic examines outputs to find logical errors or inconsistencies. The Planner adjusts the strategy based on this feedback. Through repeated cycles of execution, evaluation, and strategy refinement, the agent gradually improves the quality of the solution. Experimental tests on reasoning, coding, and math tasks show consistent progress over iterations. The results indicate that the accuracy of reasoning increases from 85% to 95%, the factual correctness increases from 88% to 97%, and the coherence improves from 75% to 94% after five refinement cycles. These findings suggest that structured prompt engineering combined with meta-cognitive feedback allows LLM agents to independently refine reasoning strategies, reduce errors, and achieve higher success rates in tasks without changing the model’s underlying parameters.