Self-refinement pipelines don't need equally capable models at each stage—you can save compute by using smaller critics while keeping generators and revisers large.
This paper studies how to allocate model sizes efficiently across self-refinement pipelines (generation → critique → revision). Testing multiple model sizes on different benchmarks, the authors find that larger generators and revisers help, small critics still work well, and undersized revisers hurt performance. The key insight: don't use the same model size for all three stages.