Optimize LLM Self-Refinement by Asymmetrically Allocating Model Capacity

A recent study investigates the optimal allocation of model capacity within Large Language Model (LLM) self-refinement pipelines, which typically involve generation, critique, and revision stages. Contrary to common practice, which often treats model size uniformly, this research systematically examines how different model sizes affect each stage's performance. The study utilized various model sizes from Qwen3 and Gemma 3 across five benchmarks. The key findings indicate that larger models for the generation and revision stages generally lead to improved pipeline performance, with an undersized refiner potentially harming overall results. Interestingly, the performance is largely insensitive to the size of the critic model; even a small critic consistently outperforms omitting the critique stage entirely. These insights suggest that computational resources should be allocated unevenly across the self-refinement pipeline, providing practical guidance for designing more efficient multi-stage LLM systems.