Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM-driven Constrained Copy Generation through Iterative Refinement

Varun Vasudevan, Faezeh Akhavizadegan, Abhinav Prakash, Yokila Arora, Jason Cho, Tanya Mendiratta, Sushant Kumar, Kannan Achan

TL;DR

The paper tackles constrained copywriting for e-commerce banners, where multiple simultaneous constraints (length, topics, keywords, tone, lexical order, etc.) make first-pass instruction following by LLMs unreliable and manual curation costly. It presents an end-to-end framework with a generator, an evaluator block, and a refiner that iteratively improves copies until all constraints are met or refinement limits are reached, operating in batches to scale production. Across three use cases of increasing complexity, the approach yields substantial gains in constraint satisfaction (16.25–35.91 percentage points) and, when deployed in pilot studies, significant CTR improvements (38.5–45.21%). The work demonstrates the practical impact of scalable, iterative refinement for personalized marketing copy, enabling broader exploration of copy variants and faster optimization of engagement.

Abstract

Crafting a marketing message (copy), or copywriting is a challenging generation task, as the copy must adhere to various constraints. Copy creation is inherently iterative for humans, starting with an initial draft followed by successive refinements. However, manual copy creation is time-consuming and expensive, resulting in only a few copies for each use case. This limitation restricts our ability to personalize content to customers. Contrary to the manual approach, LLMs can generate copies quickly, but the generated content does not consistently meet all the constraints on the first attempt (similar to humans). While recent studies have shown promise in improving constrained generation through iterative refinement, they have primarily addressed tasks with only a few simple constraints. Consequently, the effectiveness of iterative refinement for tasks such as copy generation, which involves many intricate constraints, remains unclear. To address this gap, we propose an LLM-based end-to-end framework for scalable copy generation using iterative refinement. To the best of our knowledge, this is the first study to address multiple challenging constraints simultaneously in copy generation. Examples of these constraints include length, topics, keywords, preferred lexical ordering, and tone of voice. We demonstrate the performance of our framework by creating copies for e-commerce banners for three different use cases of varying complexity. Our results show that iterative refinement increases the copy success rate by $16.25-35.91$% across use cases. Furthermore, the copies generated using our approach outperformed manually created content in multiple pilot studies using a multi-armed bandit framework. The winning copy improved the click-through rate by $38.5-45.21$%.

LLM-driven Constrained Copy Generation through Iterative Refinement

TL;DR

The paper tackles constrained copywriting for e-commerce banners, where multiple simultaneous constraints (length, topics, keywords, tone, lexical order, etc.) make first-pass instruction following by LLMs unreliable and manual curation costly. It presents an end-to-end framework with a generator, an evaluator block, and a refiner that iteratively improves copies until all constraints are met or refinement limits are reached, operating in batches to scale production. Across three use cases of increasing complexity, the approach yields substantial gains in constraint satisfaction (16.25–35.91 percentage points) and, when deployed in pilot studies, significant CTR improvements (38.5–45.21%). The work demonstrates the practical impact of scalable, iterative refinement for personalized marketing copy, enabling broader exploration of copy variants and faster optimization of engagement.

Abstract

Crafting a marketing message (copy), or copywriting is a challenging generation task, as the copy must adhere to various constraints. Copy creation is inherently iterative for humans, starting with an initial draft followed by successive refinements. However, manual copy creation is time-consuming and expensive, resulting in only a few copies for each use case. This limitation restricts our ability to personalize content to customers. Contrary to the manual approach, LLMs can generate copies quickly, but the generated content does not consistently meet all the constraints on the first attempt (similar to humans). While recent studies have shown promise in improving constrained generation through iterative refinement, they have primarily addressed tasks with only a few simple constraints. Consequently, the effectiveness of iterative refinement for tasks such as copy generation, which involves many intricate constraints, remains unclear. To address this gap, we propose an LLM-based end-to-end framework for scalable copy generation using iterative refinement. To the best of our knowledge, this is the first study to address multiple challenging constraints simultaneously in copy generation. Examples of these constraints include length, topics, keywords, preferred lexical ordering, and tone of voice. We demonstrate the performance of our framework by creating copies for e-commerce banners for three different use cases of varying complexity. Our results show that iterative refinement increases the copy success rate by % across use cases. Furthermore, the copies generated using our approach outperformed manually created content in multiple pilot studies using a multi-armed bandit framework. The winning copy improved the click-through rate by %.

Paper Structure

This paper contains 17 sections, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Proposed framework: Copy generation through iterative refinement
  4. Problem definition
  5. Generator and formatter
  6. Evaluator
  7. Refiner
  8. Use cases and evaluation criteria
  9. Use cases
  10. Evaluation Criteria
  11. Results and Discussion
  12. Effectiveness of iterative refinement
  13. Impact on user engagement
  14. Practical considerations
  15. Conclusion
  16. ...and 2 more sections

Figures (2)

  • Figure 1: LLM-based end-to-end framework for scalable copy generation. It has three main blocks: a generator, an evaluator, and a refiner. The generated copy undergoes formatting and evaluation. If any evaluation fails, it is sent to the refiner for revision. This process continues until the copy passes all evaluations or the maximum number of refine attempts is reached. Once a copy passes all evaluations, it goes to human reviewers.
  • Figure 2: Example of iterative refinement. The generated copy in (1) is rejected by a tone of voice evaluator with feedback shown in (2). This feedback is used in the refine prompt in (3) to get the refined copy in (4). The tone of voice evaluator accepts the refined copy with feedback shown in (5). {tone} contains tone of voice keywords. {campaign_name} is the name for the campaign and {service_qualifier} is any specific attribute such as "free trial".