PulseAugur
实时 13:02:00

New research tackles LVLM efficiency and hallucination problems

Two new research papers address efficiency and hallucination issues in large vision-language models (LVLMs). One paper introduces LRCP, a training-free method that uses low-rank compressibility to prune visual tokens, significantly reducing computational cost while maintaining high performance. The other paper proposes HalluScope, a benchmark and fine-tuning framework (HalluVL-DPO) to combat prompt-induced hallucinations by reducing the models' reliance on textual priors and improving visual grounding. AI

影响 New methods for pruning visual tokens and reducing hallucinations could improve the efficiency and reliability of large vision-language models.

排序理由 Two distinct research papers published on arXiv and highlighted by Hugging Face, addressing core technical challenges in large vision-language models.

在 Hugging Face Daily Papers 阅读 →

AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →

New research tackles LVLM efficiency and hallucination problems

报道来源 [4]

  1. Hugging Face Daily Papers TIER_1 English(EN) ·

    LRCP: Low-Rank Compressibility Guided Visual Token Pruning for Efficient LVLMs

    Large vision-language models (LVLMs) achieve strong multimodal understanding, but their inference cost grows rapidly with the number of visual tokens, especially for high-resolution images and long videos. Existing attention-based methods estimate token importance from attention …

  2. Hugging Face Daily Papers TIER_1 English(EN) ·

    When Prompts Override Vision: Prompt-Induced Hallucinations in LVLMs

    Despite impressive progress in capabilities of large vision-language models (LVLMs), these systems remain vulnerable to hallucinations, i.e., outputs that are not grounded in the visual input. Prior work has attributed hallucinations in LVLMs to factors such as limitations of the…

  3. arXiv cs.CV TIER_1 English(EN) · Jiawei Li ·

    LRCP: Low-Rank Compressibility Guided Visual Token Pruning for Efficient LVLMs

    Large vision-language models (LVLMs) achieve strong multimodal understanding, but their inference cost grows rapidly with the number of visual tokens, especially for high-resolution images and long videos. Existing attention-based methods estimate token importance from attention …

  4. arXiv cs.CV TIER_1 English(EN) · Matthieu Cord ·

    When Prompts Override Vision: Prompt-Induced Hallucinations in LVLMs

    Despite impressive progress in capabilities of large vision-language models (LVLMs), these systems remain vulnerable to hallucinations, i.e., outputs that are not grounded in the visual input. Prior work has attributed hallucinations in LVLMs to factors such as limitations of the…