Improving Minimum Bayes Risk Decoding with Multi-Prompt.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Instruction fine-tuned large language models (LLMs) demonstrate strong text generation capabilities but suffer from performance instability due to prompt sensitivity.
• A single prompt may not encompass all optimal strategies for a given generation task, leading to sub-optimal outcomes.

Purpose of the Study:

• To introduce and evaluate a novel multi-prompt decoding strategy to enhance the stability and performance of LLM text generation.
• To investigate if generating multiple candidate outputs from a prompt bank improves downstream task performance.

Main Methods:

• Proposing multi-prompt decoding, which generates numerous candidate text outputs from a curated bank of prompts at inference time.
• Employing Minimum Bayes Risk (MBR) decoding to ensemble these candidates, selecting the final output based on a trained value metric.

Main Results:

• Multi-prompt decoding significantly improves MBR decoding performance across a wide range of conditional text generation tasks.
• The enhanced performance is attributed to the creation of a more diverse and higher-quality candidate solution space compared to single-prompt methods.
• Further experiments validate the effectiveness of multi-prompt decoding across different LLM architectures, tasks, and evaluation metrics.

Conclusions:

• Multi-prompt decoding offers a robust method to overcome prompt sensitivity issues in instruction fine-tuned LLMs.
• This technique leads to more stable, optimal, and diverse text generation, improving overall LLM utility in conditional generation scenarios.