SHAPE : Self-Improved Visual Preference Alignment by Iteratively Generating Holistic Winner

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chen, Kejia, Zhang, Jiawen, Hu, Jiacong, Yang, Jiazhen, Lou, Jian, Feng, Zunlei, Song, Mingli
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929745688526848
author Chen, Kejia
Zhang, Jiawen
Hu, Jiacong
Yang, Jiazhen
Lou, Jian
Feng, Zunlei
Song, Mingli
author_facet Chen, Kejia
Zhang, Jiawen
Hu, Jiacong
Yang, Jiazhen
Lou, Jian
Feng, Zunlei
Song, Mingli
contents Large Visual Language Models (LVLMs) increasingly rely on preference alignment to ensure reliability, which steers the model behavior via preference fine-tuning on preference data structured as ``image - winner text - loser text'' triplets. However, existing approaches often suffer from limited diversity and high costs associated with human-annotated preference data, hindering LVLMs from fully achieving their intended alignment capabilities. We present \projectname, a self-supervised framework capable of transforming the already abundant supervised text-image pairs into holistic preference triplets for more effective and cheaper LVLM alignment, eliminating the need for human preference annotations. Our approach facilitates LVLMs in progressively enhancing alignment capabilities through iterative self-improvement. The key design rationale is to devise preference triplets where the winner text consistently improves in holisticness and outperforms the loser response in quality, thereby pushing the model to ``strive to the utmost'' of alignment performance through preference fine-tuning. For each given text-image pair, SHAPE introduces multiple visual augmentations and pairs them with a summarized text to serve as the winner response, while designating the original text as the loser response. Experiments across \textbf{12} benchmarks on various model architectures and sizes, including LLaVA and DeepSeek-VL, show that SHAPE achieves significant gains, for example, achieving +11.3\% on MMVet (comprehensive evaluation), +1.4\% on MMBench (general VQA), and +8.0\% on POPE (hallucination robustness) over baselines in 7B models. Notably, qualitative analyses confirm enhanced attention to visual details and better alignment with human preferences for holistic descriptions.
format Preprint
id arxiv_https___arxiv_org_abs_2503_04858
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SHAPE : Self-Improved Visual Preference Alignment by Iteratively Generating Holistic Winner
Chen, Kejia
Zhang, Jiawen
Hu, Jiacong
Yang, Jiazhen
Lou, Jian
Feng, Zunlei
Song, Mingli
Computer Vision and Pattern Recognition
Artificial Intelligence
Large Visual Language Models (LVLMs) increasingly rely on preference alignment to ensure reliability, which steers the model behavior via preference fine-tuning on preference data structured as ``image - winner text - loser text'' triplets. However, existing approaches often suffer from limited diversity and high costs associated with human-annotated preference data, hindering LVLMs from fully achieving their intended alignment capabilities. We present \projectname, a self-supervised framework capable of transforming the already abundant supervised text-image pairs into holistic preference triplets for more effective and cheaper LVLM alignment, eliminating the need for human preference annotations. Our approach facilitates LVLMs in progressively enhancing alignment capabilities through iterative self-improvement. The key design rationale is to devise preference triplets where the winner text consistently improves in holisticness and outperforms the loser response in quality, thereby pushing the model to ``strive to the utmost'' of alignment performance through preference fine-tuning. For each given text-image pair, SHAPE introduces multiple visual augmentations and pairs them with a summarized text to serve as the winner response, while designating the original text as the loser response. Experiments across \textbf{12} benchmarks on various model architectures and sizes, including LLaVA and DeepSeek-VL, show that SHAPE achieves significant gains, for example, achieving +11.3\% on MMVet (comprehensive evaluation), +1.4\% on MMBench (general VQA), and +8.0\% on POPE (hallucination robustness) over baselines in 7B models. Notably, qualitative analyses confirm enhanced attention to visual details and better alignment with human preferences for holistic descriptions.
title SHAPE : Self-Improved Visual Preference Alignment by Iteratively Generating Holistic Winner
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2503.04858