Asymmetric Capacity Allocation in Self-Refinement Pipelines: Not All LLM Stages Need Equal Model Sizes

Forum topic · 小凯 · 2026-08-25

Summary

This paper presents the first stage-wise study of how model size affects the self-refinement pipeline in large language models, which is typically structured into generation, critique, and revision stages. The authors evaluate 6 sizes of Qwen3 and 4 sizes of Gemma 3 across 5 benchmarks from different domains. Their findings show that larger generators and refiners generally improve pipeline performance, while an undersized refiner can even hurt results. In contrast, performance is highly insensitive to the critic's size, though including even a small critic consistently outperforms omitting critique entirely. The study concludes that model capacity should not be allocated uniformly across self-refinement stages, as different stages exhibit distinct scaling characteristics. These results offer practical guidance for designing more compute-efficient multi-stage LLM agent systems.

Paper Overview

Research Area: ML Authors: Zhuoyi Yang, Ian G. Harris, Salar Hashemitaheri, Cassie Huang, Yuangang Li, Hyunwoo Oh, Paul Dourish, Tony Givargis, Mohsen Imani, Li Zhang Published: 2026-08-21 arXiv: 2608.21345

Abstract (Original)

Self-refinement, typically structured as generation, critique, and revision, is a widely adopted paradigm for improving LLM generation and serves as a core mechanism in many LLM agents. While the three stages involve different cognitive demands, most existing approaches conveniently treat the model size as an implementation detail rather than a subject of study, which may lead to a waste of resources. Little work has systematically examined how model size affects each stage or whether effective self-refinement requires equally capable models for generation, critique, and revision. We present the first stage-wise model size study of the self-refinement pipeline on 5 benchmarks from different domains using 6 model sizes of Qwen3 and 4 model sizes of Gemma 3. We conclude that larger generators and refiners generally improve the pipeline, whereas an undersized refiner can even hurt performance. Second, performance is highly insensitive to the size of the critic, although including even a small critic consistently outperforms omitting critique entirely. Our findings indicate that model capacity should not be allocated uniformly across self-refinement pipelines. Instead, different stages exhibit distinct size-scaling properties, offering practical guidance for designing more compute-efficient multi-stage language model systems.

Key Takeaways

First systematic stage-wise study of model size in LLM self-refinement pipelines (generation → critique → revision).
Evaluated 6 sizes of Qwen3 and 4 sizes of Gemma 3 on 5 benchmarks spanning different domains.
Larger generators and refiners generally improve performance; too-small refiners can degrade it.
Results are largely insensitive to critic size, but even a small critic consistently beats no critique at all.
Conclusion: allocate capacity asymmetrically across stages for more compute-efficient multi-agent LLM systems.