[Paper Review] Deliberate then Generate: Enhanced Prompting Framework for Text Generation

Bei Li, Rui Wang+8 more|arXiv (Cornell University)|May 31, 2023

Topic Modeling5 citations

TL;DR

This paper proposes Deliberate then Generate (DTG), a novel prompting framework that enhances text generation in large language models (LLMs) by instructing them to first detect errors in a candidate output before generating a refined result. By using irrelevant or incorrect candidates—such as empty strings or random text—DTG triggers LLMs' deliberative reasoning, leading to consistent performance gains across 20+ datasets in 7 text generation tasks, including summarization, translation, and dialogue, with state-of-the-art results on multiple benchmarks.

ABSTRACT

Large language models (LLMs) have shown remarkable success across a wide range of natural language generation tasks, where proper prompt designs make great impacts. While existing prompting methods are normally restricted to providing correct information, in this paper, we encourage the model to deliberate by proposing a novel Deliberate then Generate (DTG) prompting framework, which consists of error detection instructions and candidates that may contain errors. DTG is a simple yet effective technique that can be applied to various text generation tasks with minimal modifications. We conduct extensive experiments on 20+ datasets across 7 text generation tasks, including summarization, translation, dialogue, and more. We show that DTG consistently outperforms existing prompting methods and achieves state-of-the-art performance on multiple text generation tasks. We also provide in-depth analyses to reveal the underlying mechanisms of DTG, which may inspire future research on prompting for LLMs.

Motivation & Objective

• To address the limitation of standard prompting, which treats only correct information, by enabling LLMs to deliberate and detect errors before generating.
• To explore whether prompting LLMs to evaluate and correct flawed outputs improves generation quality.
• To develop a simple yet effective prompting framework that requires minimal modification across diverse text generation tasks.
• To empirically validate that DTG enhances model performance and reveals underlying deliberative mechanisms in LLMs.

Proposed method

• DTG introduces a two-step prompting process: first, the model is prompted to detect errors in a candidate output; second, it generates a refined version.
• The candidate input is intentionally incorrect or irrelevant—such as an empty string or random text—to force the model to engage in error detection and correction.
• The framework uses a fixed prompt template: 'Please detect the error type firstly, and provide the refined results then.'
• The method is applied with minimal prompt engineering across diverse tasks, including summarization, translation, dialogue, and commonsense generation.
• Experiments use GPT-3.5 and GPT-4, with zero-shot and few-shot settings, to evaluate generalization and robustness.
• Error analysis is conducted using alignment tools (awesome-align) and named entity recognition (spaCy) to quantify under-translation and entity translation errors.

Experimental results

Research questions

• RQ1 Can prompting LLMs to deliberate on potential errors in a candidate output improve their generation quality?
• RQ2 Does using an irrelevant or incorrect candidate as input effectively trigger error detection and refinement in LLMs?
• RQ3 How does DTG compare to standard prompting across diverse text generation tasks and model variants?
• RQ4 What underlying mechanisms explain the performance gains observed in DTG, particularly in error avoidance and refinement?

Key findings

• DTG consistently outperforms standard prompting across 20+ datasets spanning 7 text generation tasks, including summarization, translation, and dialogue.
• On the WMT ZH-EN translation benchmark, DTG improved BLEU from 23.6 (standard prompting) to 25.2 and COMET from 81.12 to 81.70.
• Using an irrelevant candidate (e.g., random text or empty string) yielded better results than using a fixed incorrect candidate or a correct one, indicating that high discrepancy from the reference enhances deliberation.
• Ablation studies confirmed that LLMs engage in deliberate error detection when presented with incorrect candidates, leading to improved generation quality.
• The method achieved state-of-the-art performance on multiple benchmarks, including in machine translation, text simplification, and commonsense generation.
• Error analysis revealed that DTG significantly reduced under-translation and named entity translation errors, confirming its effectiveness in error avoidance.