Asymmetric Capacity Allocation in LLM Self-Refinement Pipelines: First Stage-Wise Model Size Study

Forum topic · 小凯 · 2026-08-25

Summary

Self-refinement—typically structured as generation, critique, and revision—is a widely adopted paradigm for improving LLM outputs and a core mechanism in many LLM agents. However, most existing work treats model size as an implementation detail rather than a research subject, potentially wasting resources. This paper presents the first stage-wise model size study of the self-refinement pipeline, evaluating 6 sizes of Qwen3 and 4 sizes of Gemma 3 across 5 benchmarks from different domains. The authors find that larger generators and revisers generally improve pipeline performance, while an overly small reviser can even hurt results. In contrast, performance is highly insensitive to critic size, though even a small critic consistently outperforms omitting critique entirely. These findings indicate that model capacity should not be distributed uniformly across self-refinement stages, offering practical guidance for designing compute-efficient multi-stage LLM systems.

Overview

Self-refinement, typically structured as generation → critique → 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 treat model size as an implementation detail rather than a subject of study, which may lead to a waste of resources.

This paper asks: does effective self-refinement actually require equally capable models for each stage? The authors present the first stage-wise model size study of the self-refinement pipeline.

Experimental Setup

Models: 6 model sizes of Qwen3 and 4 model sizes of Gemma 3
Benchmarks: 5 benchmarks from different domains
Method: Systematically vary model size per stage (generator, critic, reviser) and measure pipeline performance

Key Findings

1. Larger generators and revisers generally improve pipeline performance. Scaling up helps at both ends of the pipeline. 2. An overly small reviser can hurt performance rather than merely provide no benefit. 3. Performance is highly insensitive to critic size. However, even including a small critic consistently outperforms completely omitting critique.

Implications

Model capacity should not be allocated uniformly across self-refinement stages. Different stages exhibit distinct size-scaling characteristics, offering practical guidance for designing more compute-efficient multi-stage LLM systems.