Control Large Language Models via Divide and Conquer

TL;DR

The Divide and Conquer Generation strategy is introduced, effective for both white-box and black-box LLMs, to enhance LLMs performance in LCG tasks, which demonstrates over 90% improvement on success rate in the most challenging LCG task.

Abstract

This paper investigates controllable generation for large language models (LLMs) with prompt-based control, focusing on Lexically Constrained Generation (LCG). We systematically evaluate the performance of LLMs on satisfying lexical constraints with prompt-based control, as well as their efficacy in downstream applications. We conclude that LLMs face significant challenges in consistently satisfying lexical constraints with prompt-based control. We identified three key limitations of LLMs for LCG, including (1) position bias, where LLMs tend to satisfy constraints that appear in specific positions within the input; (2) low responsiveness to decoding parameters, which render minimal impact on control of LLMs; and (3) struggle with handling the inherent complexity of certain constraints (e.g., compound words). To address these issues, we introduce a Divide and Conquer Generation strategy, effective for both white-box and black-box LLMs, to enhance LLMs performance in LCG tasks, which demonstrates over 90% improvement on success rate in the most challenging LCG task. Our analysis provides valuable insights into the performance of LLMs in LCG with prompt-based control, and our proposed strategy offers a pathway to more sophisticated and customized text generation applications.

Figures (13)

• Figure 1: Sub-figure (a) illustrates that modern LLMs struggle to consistently meet complex lexical constraints in real-world scenarios. As shown in sub-figure (b), the Divide and Conquer Generation strategy divides the constraints into two parts (satisfied and missed), then generates a response with the missed constraints - i.e. response (2) - and merges them with the satisfied ones, enhancing the LLMs' ability to meet all specified constraints.
• Figure 2: Experimental results showing instance success rates by the number of concepts. GPT models consistently outperform LLaMA models, though performance declines across all models as the number of keywords increases.
• Figure 3: Experiment results on instance success rate by number of keywords. As the number of keywords increases, LLMs’ performance decreases dramatically with prompt-based control.
• Figure 4: Experimental results on the position sensitivity of LLaMA3-8b, presented as keyword coverage rate (y-axis) for constraints placed at different positions (x-axis). Across varying numbers of keywords, the performance of LLMs exhibit similar trends.
• Figure 6: Comparison experiment between Rejection Sampling (RJ) and Divide-and-Conquer Generation (DnC). The x-axis represents the maximum number of allowed iterations, while the y-axis shows the error rate for each approach in satisfying all lexical constraints. The Divide-and-Conquer Generation strategy improves LLM performance by over 40% compared to the baseline.