Propose-and-Amplify Strategy

Summary: A cost-effective approach where an expensive, high-capability model generates seed tasks or examples, which are then used to train a cheaper model to amplify the generation process at scale. This strategy optimizes the trade-off between quality and computational cost in large-scale task generation.

Overview

The Propose-and-Amplify Strategy leverages the complementary strengths of different model capabilities to achieve scalable task generation. An expensive, frontier model with superior reasoning and creativity capabilities creates high-quality seed examples that establish patterns, templates, and quality standards. These seeds then guide a more cost-effective model to generate similar content at much larger scale.

This approach is particularly valuable in scenarios requiring thousands or tens of thousands of examples where using the expensive model for all generation would be prohibitively costly, but where quality standards must remain high. The strategy relies on the assumption that once good examples are established, a cheaper model can learn to replicate the patterns and maintain acceptable quality levels.

The effectiveness depends on the seed examples capturing the full diversity and complexity needed for the target domain, and the amplification model's ability to generalize from limited examples while maintaining quality standards.

Key Details

• Two-Phase Process: Initial seed generation by expensive model, followed by large-scale amplification by cheaper model
• Cost Optimization: Dramatically reduces computational costs compared to using expensive models for all generation
• Quality Control: Seed examples establish quality standards and provide templates for the amplification phase
• Scale Achievement: Enables generation of thousands of examples while maintaining acceptable quality levels
• Pattern Transfer: Relies on the cheaper model's ability to learn and replicate patterns from seed examples
• Domain Coverage: Seed examples must represent the full scope and diversity of the target domain
• Quality-Cost Trade-off: Balances between maintaining high standards and achieving scalable production

Relationships

• Multi-Agent Environment Creation — implements propose-and-amplify through creation-audit agent pairs
• Trajectory Distillation — uses similar principles of transferring capability from larger to smaller models
• GDP-Grounded Benchmarking — benefits from propose-and-amplify for generating economically representative task sets
• Computer-Use Agents — training data generation often employs propose-and-amplify strategies
• Automated Verification — quality control mechanisms essential for validating amplified outputs
• Cross-Software Generalization — seed examples help establish patterns that transfer across different software domains

Sources

• sources/arxiv-260406126 — demonstrates propose-and-amplify in CUA-World benchmark creation where expensive models generate seed tasks across software categories