{"slug": "why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information", "title": "Why and When Visual Token Pruning Fails? A Study on Relevant Visual Information Shift in MLLMs Decoding", "summary": "A new study identifies Relevant Visual Information Shift (RVIS) during decoding as the primary cause of performance degradation in visual token pruning methods for Multimodal Large Language Models (MLLMs) on complex reasoning tasks. The authors propose Decoding-stage Shift-aware Token Pruning (DSTP), a training-free add-on framework that aligns visual tokens with shifting reasoning requirements, significantly mitigating performance loss and yielding gains across benchmarks and architectures.", "body_md": "Recently, visual token pruning has been studied to handle the vast number of visual tokens in Multimodal Large Language Models. However, we observe that while existing pruning methods perform reliably on simple visual understanding, they struggle to effectively generalize to complex visual reasoning tasks, a critical gap underexplored in previous studies. Through a systematic analysis, we identify Relevant Visual Information Shift (RVIS) during decoding as the primary failure driver. To address this, we propose Decoding-stage Shift-aware Token Pruning (DSTP), a training-free add-on framework that enables existing pruning methods to align visual tokens with shifting reasoning requirements during the decoding stage. Extensive experiments demonstrate that DSTP significantly mitigates performance degradation of pruning methods in complex reasoning tasks, while consistently yielding performance gains even across visual understanding benchmarks. Furthermore, DSTP demonstrates effectiveness across diverse state-of-the-art architectures, highlighting its generalizability and efficiency with minimal computational overhead.", "url": "https://wpnews.pro/news/why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information", "canonical_source": "https://research.nvidia.com/publication/2026-09_why-and-when-visual-token-pruning-fails-study-relevant-visual-information-shift", "published_at": "2026-08-11 07:00:49+00:00", "updated_at": "2026-08-11 07:08:50.821122+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "computer-vision", "large-language-models"], "entities": ["Multimodal Large Language Models", "Relevant Visual Information Shift", "Decoding-stage Shift-aware Token Pruning"], "alternates": {"html": "https://wpnews.pro/news/why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information", "markdown": "https://wpnews.pro/news/why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information.md", "text": "https://wpnews.pro/news/why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information.txt", "jsonld": "https://wpnews.pro/news/why-and-when-visual-token-pruning-fails-a-study-on-relevant-visual-information.jsonld"}}