Files and links (2)
Published, Version of Record (VoR)Open Access (License Unspecified), Open
Conference proceeding
Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, v 1, pp 5050-5063
2024
Abstract
Large Language Models (LLMs) show promising results in language generation and instruction following but frequently "hallucinate", making their outputs less reliable. Despite Uncertainty Quantification's (UQ) potential solutions, implementing it accurately within LLMs is challenging. Our research introduces a simple heuristic: not all tokens in auto-regressive LLM text equally represent the underlying meaning, as "linguistic redundancy" often allows a few keywords to convey the essence of long sentences. However, current methods underestimate this inequality when assessing uncertainty, causing tokens with limited semantics to be equally or excessively weighted in UQ. To correct this, we propose Shifting Attention to more Relevant (SAR) components at both token- and sentence-levels for better UQ. We conduct extensive experiments involving a range of popular "off-the-shelf" LLMs, such as Vicuna, WizardLM, and LLaMA-2-chat, with model sizes extending up to 33B parameters. We evaluate various free-form question-answering tasks, encompassing domains such as reading comprehension, science Q&A, and medical Q&A. Our experimental results, coupled with a comprehensive demographic analysis, demonstrate the superior performance of SAR. The code is available at https://github.com/jinhaoduan/SAR.
Metrics
9 Record Views
Details
- Title
- Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models
- Creators
- Jinhao Duan - Drexel UniversityHao Cheng - University of Hong KongShiqi Wang - Art Institute of WisconsinAlex Zavalny - Drexel UniversityChenan Wang - Drexel UniversityRenjing Xu - University of Hong KongBhavya Kailkhura - Lawrence Livermore National LaboratoryKaidi Xu (Corresponding Author) - Drexel University
- Contributors
- L W Ku (Editor)A Martins (Editor)Srikumar (Editor)
- Publication Details
- Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, v 1, pp 5050-5063
- Conference
- Annual Meeting of the Association for Computational Linguistics, 62 (Bangkok, Thailand, 11 Aug 2024–16 Aug 2024)
- Publisher
- Assoc Computational Linguistics-Acl
- Number of pages
- 14
- Grant note
- 2319242 / National Science Foundation; National Science Foundation (NSF) 23-ERD-030 (LLNL-CONF-851171) / LLNL-LDRD Program DE-AC52-07NA27344 / U.S. Department of Energy by Lawrence Livermore National Laboratory; United States Department of Energy (DOE)
- Resource Type
- Conference proceeding
- Language
- English
- Academic Unit
- Computer Science
- Web of Science ID
- WOS:001356729805012
- Scopus ID
- 2-s2.0-85204438471
- Other Identifier
- 991022035263804721
InCites Highlights
Data related to this publication, from InCites Benchmarking & Analytics tool:
- Collaboration types
- Domestic collaboration
- International collaboration
- Web of Science research areas
- Computer Science, Artificial Intelligence
- Computer Science, Interdisciplinary Applications
- Computer Science, Theory & Methods