Visual token pruning aims to reduce the number of image tokens a model must process. A new preprint on arXiv observes that different pruning strategies can retain different subsets of tokens while still achieving similar task performance. That observation motivates a combinatorial question: how many distinct visual subsets can preserve model behavior?
To address this, the paper proposes "visual sufficiency landscapes," a framing for understanding query-conditioned reliability in pruning. Instead of treating a pruned token set as universally good or bad, the landscape view ties reliability to the specific query the model must answer.
The abstract is truncated, so the full method and results are not available from the summary alone. The preprint is posted as arXiv:2610.24968v1.