AIBRIEF
GMC, a training free pruning method, lets a vision language model discard most of the image pieces it ingests while keeping 97–99% of its answer quality on author reported benchmarks.
Every high-resolution image fed into a vision-language model is split into thousands of small pieces called visual tokens — the reason multimodal AI is slow and expensive to run. A team at the Institute of Automation, Chinese Academy of Sciences (the 紫东太初 / Zidong Taichu group) released a training-free method called GMC that, on its own benchmarks, lets the model discard roughly 80% of those tokens while keeping 97–99% of its answer quality.
On Qwen2.5-VL-7B-Instruct, the full model uses 1,296 visual tokens per image. Compressing to 256 tokens (an 80.2% cut) preserved 97.78% of the full model's average score, the team reports in the arXiv preprint; pushing to 128 tokens (a 90.1% cut) still kept 99.11%. The method also transferred to LLaVA-1.5-7B, where 64- and 128-token compression retained 99.82% and 99.76% of full-model accuracy.
GMC picks survivors with a three-criterion rule: a token must be attended to by the question text, visually similar to the image, and grounded in the original image coordinates. Dropped tokens' hidden states are transported into the survivors. On a 15,876-token long-document task, the team reports a 1.26× end-to-end speedup and a 73.94% cut in key–value cache memory.
The paper is an arXiv preprint dated 2026-08-03, not peer-reviewed, with no independent replication or comparison to competing pruning methods (FastV, SnapKV, or ToMe) in the source.