LLM Self-Refinement

Large language model self-refinement asks whether models can improve their answers, code, plans, or other outputs by evaluating and revising their own work. Common approaches use iterative critique-and-revision, structured rubrics or judges, and external signals such as execution results, formal verifiers, retrieved evidence, or environment feedback; some methods also train models to internalize reflection through reinforcement learning. These techniques are applied in settings from mathematical reasoning and code generation to clinical extraction, writing, and agentic tasks, with adaptive stopping and memory used to manage repeated attempts. A central tension is that revision can fix errors but also damage correct answers: gains depend on task, model, feedback quality, and compute, while externally grounded verification is often more reliable than self-assessment alone.