Stage-Wise Study Finds the Critic in a Generate-Critique-Revise Loop Barely Needs Any Model Capacity

arXiv 2608.21345·medium signalXBlueskyLinkedInCopy link

Asymmetric Capacity Allocation in Self-Refinement Pipelines (arXiv 2608.21345, Aug 21) runs the first systematic stage-wise model size study of self-refinement across 5 benchmarks using 6 sizes of Qwen3 and 4 sizes of Gemma 3. Larger generators and refiners generally improve the pipeline and an undersized refiner can actively hurt performance, but results are highly insensitive to critic size, though including even a small critic consistently beats omitting critique entirely. For anyone paying per token across a three-stage agent loop, that is a direct instruction to spend capacity on generation and revision and route the critique step to the cheapest model that still produces a critique.