# Asymmetric Capacity Allocation in Self-Refinement Pipelines

Source: [arXiv](https://arxiv.org/abs/2608.21345v1)  
Feed7 permalink: https://feed7.dev/p/2608-21345v1-0octema  
Published: 2026-08-21T17:52:17.000Z  
Trust: Needs Review (needs_review)

## Why Included

Self-refinement pipelines need not use equally capable models: invest capacity in generation and revision, while a small critic may preserve gains at lower compute cost.

## Source Summary

Across **5 benchmarks**, the study varies **6 Qwen3 sizes** and **4 Gemma 3 sizes** across generation, critique, and revision. Larger generators and refiners generally help, while an undersized refiner can reduce performance.

## Practical Implication

For agent pipelines, allocate the strongest affordable models to generation and revision. Keep critique in the loop, but test a smaller critic: results were largely insensitive to critic size, and even a small critic beat omitting critique.

## Agent-Ready Context

Across **5 benchmarks**, the study varies **6 Qwen3 sizes** and **4 Gemma 3 sizes** across generation, critique, and revision. Larger generators and refiners generally help, while an undersized refiner can reduce performance.

For agent pipelines, allocate the strongest affordable models to generation and revision. Keep critique in the loop, but test a smaller critic: results were largely insensitive to critic size, and even a small critic beat omitting critique.

These are stage-level findings across two model families, not a universal routing formula. Builders still need workload-specific evaluations before turning asymmetric allocation into a fixed policy.

## Connected Context

Feed7 judgment across 551 accumulated Signals:

This narrows model routing from choosing one model per task or turn to allocating capacity by pipeline stage: generation and revision are the sensitive positions, while critique may remain useful with a cheaper model. It supports asymmetric deployment, but only as a workload-tested policy because the evidence spans two model families and does not capture production costs or switching behavior.

- [How Cursor Router chooses the right model for the task](https://feed7.dev/p/how-cursor-router-works-1fhejk3) — Cursor supplies the production routing context that this study lacks: its per-turn correction signals and economics could determine when the paper’s stage-level capacity pattern is worth applying.
- [Guide, Verify, Solve — Anirban Chatterjee, Sonar](https://feed7.dev/p/guide-verify-solve-anirban-chatterjee-sonar-1igfmbm) — The finding that even a small critic improves over no critique reinforces Sonar’s case for keeping verification inside each iteration, while adding that the verifier may not need the strongest model.
- [TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning](https://feed7.dev/p/2608-04007v1-0gjxuy2) — Both make pipeline stages independently actionable: this study assigns inference capacity by role, while TurnSight assigns training signals to individual tool-use turns.

## Context Map

- Layer: agent
- Domains: coding, research
- Topics: harness-engineering, model-selection, agent-reliability

## Uncertainty

- These are stage-level findings across two model families, not a universal routing formula. Builders still need workload-specific evaluations before turning asymmetric allocation into a fixed policy.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.