Vision-language models are expensive to run, and visual token pruning has emerged as a way to cut memory and inference costs by discarding less informative image tokens. Most research so far has centered on designing clever selectors that pick which tokens to drop, then showing small improvements over previous methods.
A new paper on arXiv takes a different angle. Instead of adding another selector, it asks a more fundamental question: what are the actual performance limits of visual token pruning? The authors argue that the field's focus on incremental, method-specific gains has left this broader question underexplored.
The study's full results are not detailed in the abstract, but the framing suggests a move toward understanding how aggressively tokens can be pruned before model accuracy collapses. That kind of boundary-mapping could help future work set realistic expectations and guide more principled pruning strategies.