A central obstacle to the practical deployment of Reinforcement Learning (RL) is the prevalence of sparse rewards, which often necessitates task-specific dense signals crafted through costly trial-and-error. Automated reward decomposition and return–redistribution methods can reduce this burden, but they are largely semantically agnostic and may fail to capture the multifaceted nature of task performance, leading to reward hacking or stalled exploration. Recent work uses Large Language Models (LLMs) to generate reward functions from high-level task descriptions, but these specifications are typically static and may encode biases or inaccuracies from the pretrained model, resulting in a priori reward misspecification. To address this, we propose the Metacognitive Introspective Reward Architecture (MIRA), a closed-loop architecture that treats LLM-generated reward code as a dynamic object refined through empirical feedback. An LLM first produces a set of computable reward factors. A dual-loop design then decouples policy learning from reward revision: an inner loop jointly trains the agent’s policy and a reward-synthesis network to align with sparse ground-truth outcomes, while an outer loop monitors learning dynamics via diagnostic metrics and, upon detecting pathological signatures, invokes the LLM to perform targeted structural edits. Experiments on MuJoCo benchmarks show that MIRA corrects flawed initial specifications and improves asymptotic performance and sample efficiency over strong reward-design baselines.