Decoupling Prompt Tuning for Long-Tailed Visual Recognition

ICLR 2026 Conference Submission19033 Authors

Keywords: Long-tailed Data Learning, Visual Recognition, Decoupling Prompt

TL;DR: Decoupling Prompt Tuning for Long-Tailed Visual Recognition

Abstract: Long-tailed visual recognition encounters difficulties in feature learning, particularly for tailed classes. Recently, fine-tuning visual prompt of pre-trained large models with powerful feature extraction capabilities for long-tailed data has been investigated. Although the result is promising, visual prompts are easily affected by semantically irrelevant parts of images, resulting in diminished effectiveness. To address this issue, we propose a Long-tailed Decoupling Prompt Tuning (LTDPT) method for long-tailed visual recognition. LT-DPT explicitly decouples visual prompts into foreground-related prompts and background-related prompts, respectively. Specifically, foreground-related prompts emphasize saliency regions of the image which includes the most discriminative information for classification while background-related prompts capture common background features shared across classes, regardless of their category. In comparison with the state-of-the-art methods, extensive experiments demonstrate that the proposed method achieves better performance on long-tailed visual recognition benchmark datasets.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 19033