{"slug": "mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy", "title": "MI-Pruner Cuts 89% of Image Tokens While Keeping 97% AI Accuracy", "summary": "Researchers at KU Leuven and Weill Cornell Medicine released MI-Pruner, a training-free method that scores projected image and text features by mutual information to prune visual tokens before a multimodal model's decoder, removing 88.9% of LLaVA1.5-7B's visual tokens (leaving 64 of 576) while preserving more than 97% of the unpruned model's benchmark performance. On video, MI-Pruner averages 93.19% accuracy using only 114 visual tokens, or 5.57% of the original, and it runs on Qwen3VL, which [CLS]-based pruners like VisPruner cannot handle. The authors acknowledge limits: Qwen3VL baselines were built by the authors themselves, MI-Pruner beats cosine scoring by only 1-2 points, and outside reviewers flagged under-testing; code is on GitHub.", "body_md": "# MI-Pruner Cuts 89% of Image Tokens While Keeping 97% AI Accuracy\n\nMI-Pruner decides which image tokens to drop by measuring how much each one shares with the text prompt. It never reads attention maps, so it can be added to off-the-shelf multimodal models without changing them.\n\n- [MI-Pruner](https://arxiv.org/html/2604.03072v2?utm_source=alphasignal) drops image tokens based on how much each shares with the text prompt (mutual information), with no attention maps needed.\n- On LLaVA1.5-7B it keeps over 97% of full performance while removing 88.9% of visual tokens.\n- On video it averages 93.19% accuracy with only 114 visual tokens, or 5.57% of the original.\n- It is training-free, works with FlashAttention, and runs on Qwen3VL, which [CLS]-based pruners like VisPruner cannot handle.\n- Limits: Qwen3VL baselines were built by the authors, MI beats cosine scoring by only 1-2 points, and outside reviewers flagged under-testing.\n- Code is available on [GitHub](https://github.com/renaissanceee/MI-Pruner?utm_source=alphasignal) .\n\nResearchers at KU Leuven and Weill Cornell Medicine have introduced [MI-Pruner](https://arxiv.org/html/2604.03072v2?utm_source=alphasignal), a training-free method that removes visual tokens before a multimodal large language model processes them. It scores projected image and text features with mutual information, requiring no decoder attention maps or model retraining. On LLaVA1.5-7B, the authors report preserving more than 97% of the unpruned model’s benchmark performance while removing 88.9% of its visual tokens, leaving 64 of 576.\n\nLLaVA-style models convert each image into hundreds of tokens that enter the language-model decoder alongside the prompt. Those tokens increase prefill computation, attention work and key-value cache memory, even when many image patches contribute little to the answer. Pruning them before decoding can reduce latency and memory use while leaving the base model unchanged.\n\n## Attention maps are a fragile dependency\n\nExisting visual-token pruners generally rank image patches through attention scores from either the vision encoder or the language-model decoder. FastV reads decoder attention, while VisionZip and VisPruner use signals from the vision encoder.\n\n| Signals and integration requirements of representative pruning methods |  |  | \n|---|---|---|\n| Method | Ranking signal | Integration constraint | \n|---|---|---|\n| FastV | LLM decoder attention | Requires decoder attention scores | \n| VisionZip | Vision-encoder attention | Depends on encoder attention outputs | \n| VisPruner | Vision-encoder attention | Uses a `[CLS]` summary token | \n| MI-Pruner | Projected image and text features | Requires no attention matrix | \n\nDecoder-side ranking complicates integration with optimized inference kernels such as FlashAttention, which computes attention without materializing the full attention matrix. Encoder-side methods can depend on architecture-specific outputs. The authors note that VisPruner expects a `[CLS]` token, while recent Qwen vision-language encoders do not provide one.\n\n## A score for relevance and repetition\n\nMutual information measures how much knowledge of one variable reduces uncertainty about another. Pointwise mutual information, or PMI, applies that idea to a specific pair, such as an image patch and a prompt token. MI-Pruner constructs PMI-based scores from embedding similarities in the projector’s output space, where image and text tokens share the language model’s representation format but have not entered the decoder.\n\n## This story is for Pro members\n\nYou've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.", "url": "https://wpnews.pro/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy", "canonical_source": "https://alphasignal.ai/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy", "published_at": "2026-11-18 18:01:33+00:00", "updated_at": "2026-11-18 18:16:22.221536+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "computer-vision", "ai-research"], "entities": ["MI-Pruner", "KU Leuven", "Weill Cornell Medicine", "LLaVA1.5-7B", "Qwen3VL", "VisPruner", "FastV", "VisionZip"], "also_reported_by": [], "takeaway": "MI-Pruner's mutual-information scoring of projected image and text features prunes 88.9% of LLaVA1.5-7B's visual tokens (64 of 576 remain) while retaining over 97% of full benchmark performance, and on video it holds 93.19% accuracy with 114 visual tokens (5.57% of the original).", "recommendation": {"advice": "Prune visual tokens by mutual-information scoring in the projector's output space instead of decoder attention maps, so pruning stays compatible with FlashAttention; the authors kept over 97% of LLaVA1.5-7B performance while dropping 88.9% of image tokens.", "playbook": "inference-serving", "playbook_url": "https://wpnews.pro/playbooks/inference-serving"}, "alternates": {"html": "https://wpnews.pro/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy", "markdown": "https://wpnews.pro/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy.md", "text": "https://wpnews.pro/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy.txt", "jsonld": "https://wpnews.pro/news/mi-pruner-cuts-89-of-image-tokens-while-keeping-97-ai-accuracy.jsonld"}}