Bigger critics don't improve AI self-refinement loops

Researchers tested self-refinement loops across five benchmarks, using six sizes of Qwen3 and four sizes of Gemma 3 to isolate which stage of a generate-critique-revise loop drives quality most. The pattern held across both model families: bigger generators and bigger refiners reliably produce better final output. Critic size barely mattered.

The instinct when building one of these loops is to assume every stage benefits from a stronger model, so budget gets spread evenly. This self-refinement study says that instinct is wrong for the critique step specifically. A tiny critic and a huge critic land in about the same place, as long as a critic exists at all.

An undersized refiner is riskier than an undersized critic. Skimp on the generator or refiner and results can fall below the unrefined baseline. Skimp on the critic and the loss is close to nothing.

For anyone running a three-stage loop and paying per token, route critique to the cheapest model that still produces a coherent critique, then put the saved budget into the generator or refiner. Skipping critique entirely still underperforms including even a small one. The move is shrinking the critic, not cutting it.