Improving Minimum Bayes Risk Decoding with Multi-Prompt.

Gespeichert in:
Bibliographische Detailangaben
Titel: Improving Minimum Bayes Risk Decoding with Multi-Prompt.
Autoren: Heineman D; School of Interactive Computing, Georgia Institute of Technology., Dou Y; School of Interactive Computing, Georgia Institute of Technology., Xu W; School of Interactive Computing, Georgia Institute of Technology.
Quelle: Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing [Proc Conf Empir Methods Nat Lang Process] 2024 Nov; Vol. 2024, pp. 22525-22545.
Publikationsart: Journal Article
Sprache: English
Info zur Zeitschrift: Publisher: ACL Country of Publication: United States NLM ID: 101669294 Publication Model: Print Cited Medium: Print NLM ISO Abbreviation: Proc Conf Empir Methods Nat Lang Process Subsets: PubMed not MEDLINE
Imprint Name(s): Publication: 1997- : Somerset, N.J. : ACL
Original Publication: Philadelphia, Pa. : University of Pennsylvania, 1996-
Abstract: 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 (Figure 1), 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.
Grant Information: R01 LM014600 United States LM NLM NIH HHS
Entry Date(s): Date Created: 20250704 Latest Revision: 20250705
Update Code: 20250705
PubMed Central ID: PMC12226151
DOI: 10.18653/v1/2024.emnlp-main.1255
PMID: 40612446
Datenbank: MEDLINE