arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2411.13697cs.CV
收藏

分解并利用专家模型的偏好以提升多模态大语言模型的可信度

Decompose and Leverage Preferences from Expert Models for Improving Trustworthiness of MLLMs

University of Cambridge(剑桥大学);Nanyang Technological University(南洋理工大学);Queen Mary University of London(伦敦玛丽女王大学)

Rui Cao, Yuming Jiang, Michael Schlichtkrull, Andreas Vlachos

AI总结:

本文提出DecompGen框架,通过将MLLMs响应分解为原子验证任务并分配给多个开源专家模型生成细粒度评估,自动构建偏好数据集DGPref,从而提升多模态大语言模型的可信度。

摘要:

多模态大语言模型(MLLMs)可以通过与人类偏好对齐来增强可信度。由于人类偏好标注十分费力,近期工作采用评估模型来评价MLLMs的响应,利用基于模型的评估来自动化构建偏好数据集。然而,这种方法面临MLLMs响应冗长且具有组合性的挑战,这些响应通常需要多样化的推理能力,而单一评估模型可能并不完全具备这些能力。此外,大多数现有方法依赖闭源模型作为评估器。为解决这些局限性,我们提出了DecompGen,一个使用开源专家模型集成的可分解框架。DecompGen将每个响应分解为原子验证任务,并将每个任务分配给合适的专家模型以生成细粒度评估。DecompGen的反馈被用于自动构建我们的偏好数据集DGPref。通过偏好学习与DGPref对齐的MLLMs在可信度方面表现出提升,证明了DecompGen的有效性。

英文摘要:

Multimodal Large Language Models (MLLMs) can enhance trustworthiness by aligning with human preferences. As human preference labeling is laborious, recent works employ evaluation models for assessing MLLMs' responses, using the model-based assessments to automate preference dataset construction. This approach, however, faces challenges with MLLMs' lengthy and compositional responses, which often require diverse reasoning skills that a single evaluation model may not fully possess. Additionally, most existing methods rely on closed-source models as evaluators. To address limitations, we propose DecompGen, a decomposable framework that uses an ensemble of open-sourced expert models. DecompGen breaks down each response into atomic verification tasks, assigning each task to an appropriate expert model to generate fine-grained assessments. The DecompGen feedback is used to automatically construct our preference dataset, DGPref. MLLMs aligned with DGPref via preference learning show improvements in trustworthiness, demonstrating the effectiveness of DecompGen.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑