Prompt Engineering at Scale: Provably Effective Multi-Agent Cascades for Attribute Generation in E-Commerce

ICLR 2026 Conference Submission20309 Authors

Keywords: Prompt Engineering, CascadeAgent, Product Attribute Generation, Multi-pass Prompt Generation (MPG), Textual Gradient Optimization, Industrial Scale, Theoretical Analysis

TL;DR: CascadeAgent is a multi-agent framework that automatically generates and optimizes domain-specific LLM prompts, showing significant improvements in e-commerce product attribute enrichment.

Abstract:

Developing specialized Large Language Model (LLM) prompts for domain-specific tasks at scale remains a significant hurdle, particularly for e-commerce applications managing tens of thousands of distinct product attributes. We introduce CascadeAgent, a novel multi-agent framework that automates prompt adaptation and specialization through semantic gradient-based refinement. CascadeAgent employs a hierarchical architecture where a central Prompting Agent orchestrates four specialized counterparts—Writing, Generation, Evaluation, and Flaw Detection—that collaboratively analyze domain metadata, construct attribute-specific prompts, and enhance performance through iterative feedback. Our approach combines Multi-pass Prompt Generation (MPG) for modularity with textual gradient optimization that refines instructions based on detected error patterns. We provide formal theoretical analysis demonstrating provable convergence towards reduced loss under defined conditions. In a large-scale e-commerce case study on product attribute enrichment, CascadeAgent generated and optimized over 27,000 distinct prompts, achieving improvements of +21% to +33% in precision and +12% to +14% in coverage across multiple LLMs.
These results highlight CascadeAgent's capacity for robust, automated prompt engineering at industrial scale, while making more affordable models viable for deployment. The framework's modular design, iterative improvement mechanism, and theoretical guarantees make it a strong candidate for applications requiring principled refinement of vast numbers of task-specific prompts.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 20309