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.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the control problems of large - language models (LLMs) in lexically constrained generation (LCG) tasks. Specifically, the author focuses on achieving precise control over the text generated by LLMs through prompt - based control methods, ensuring that the generated text can meet specific lexical constraints. #### Main problems and challenges: 1. **Position Bias**: - LLMs tend to satisfy constraints that appear in specific positions in the input, which leads to inconsistency in the generation results. For example, when certain keywords appear at the beginning or end of the prompt, the performance of the model will be different. 2. **Low Responsiveness to Decoding Parameters**: - Decoding parameters (such as temperature, top - k, top - p, etc.) have little impact on LLMs in LCG tasks and cannot significantly improve the model's ability to control lexical constraints. 3. **Struggle with Complex Constraints**: - For some complex constraints (such as compound words), LLMs are difficult to handle effectively and may split compound words into sub - words, thus changing or misinterpreting the intention of the output. #### Proposed solutions: To solve the above problems, the author proposes a method called "Divide and Conquer Generation (DnC)". This strategy decomposes complex lexical constraint tasks into multiple simple subtasks and gradually generates text that satisfies all constraints. The DnC strategy is applicable to white - box and black - box LLMs and significantly improves the performance of the model in LCG tasks. #### Experimental verification: The author verifies the effectiveness of the DnC strategy through a series of experiments, including but not limited to: - **Simple - constraint experiments**: Use the CommonGen benchmark dataset to evaluate the ability of LLMs to generate coherent sentences. - **Complex - constraint experiments**: Increase the number of concepts that need to be incorporated into the generated text and evaluate the model's ability to handle complex constraints. - **Decoding - parameter sensitivity analysis**: Systematically adjust the decoding parameters to study their impact on the performance of LCG tasks. - **Practical - application tests**: Evaluate the performance of LLMs in practical application scenarios such as recipe generation, table - to - text conversion, and customer profile writing. #### Conclusions: By introducing the DnC strategy, the author successfully solves the main challenges faced by LLMs in LCG tasks and significantly improves the success rate of the model in meeting lexical constraints. Especially in the most complex LCG tasks, the DnC strategy increases the success rate of the model by more than 90%, providing an effective solution for more complex text - generation applications. ### Summary This paper conducts an in - depth analysis of the performance of LLMs in LCG tasks, identifies the main existing problems, and proposes an effective solution - the Divide and Conquer Generation (DnC) strategy. This strategy not only improves the performance of LLMs in LCG tasks but also provides new ideas for future research and applications.

Control Large Language Models via Divide and Conquer

Adaptable Logical Control for Large Language Models

Evaluating, Understanding, and Improving Constrained Text Generation for Large Language Models

Prompt-Based Length Controlled Generation with Reinforcement Learning

Prompt-Based Length Controlled Generation with Multiple Control Types

Controllable Text Generation for Large Language Models: A Survey

A Comprehensive Evaluation of Constrained Text Generation for Large Language Models.

Towards Understanding the Effectiveness of Large Language Models on Directed Test Input Generation

Self-controller: Controlling LLMs with Multi-round Step-by-step Self-awareness

Large Language Models are Good Multi-lingual Learners : When LLMs Meet Cross-lingual Prompts

Length Controlled Generation for Black-box LLMs

Benchmarking Large Language Models on Controllable Generation under Diversified Instructions

Large Language Model-Aware In-Context Learning for Code Generation

Large Language Models can Contrastively Refine their Generation for Better Sentence Representation Learning

On the Effectiveness of Large Language Models in Domain-Specific Code Generation

Unlocking Anticipatory Text Generation: A Constrained Approach for Large Language Models Decoding

Enabling Controllable Table-to-text Generation Via Prompting Large Language Models with Guided Planning

LLMGA: Multimodal Large Language Model based Generation Assistant

Supervised Knowledge Makes Large Language Models Better In-context Learners

Aligning Large Language Models for Controllable Recommendations