Improving Minimum Bayes Risk Decoding with Multi-Prompt
This addresses the practical issue of prompt sensitivity in text generation for users of instruction-tuned LLMs, though it appears incremental as it builds on existing MBR decoding.
The paper tackles the problem of instruction-tuned LLMs being sensitive to prompt construction, which leads to unstable and sub-optimal text generation. It proposes multi-prompt decoding with Minimum Bayes Risk (MBR) decoding, showing improvements across a comprehensive set of conditional generation tasks, models, and metrics.
While instruction fine-tuned LLMs are effective text generators, sensitivity to prompt construction makes performance unstable and sub-optimal in practice. Relying on a single "best" prompt cannot capture all differing approaches to a generation problem. Using this observation, we propose multi-prompt decoding, where many candidate generations are decoded from a prompt bank at inference-time. To ensemble candidates, we use Minimum Bayes Risk (MBR) decoding, which selects a final output using a trained value metric. We show multi-prompt improves MBR across a comprehensive set of conditional generation tasks, and show this is a result of estimating a more diverse and higher quality candidate space than that of a single prompt. Further experiments confirm multi-prompt improves generation across tasks, models and metrics.