Hi, thank you for the great work! I would like to clarify the “64 retained tokens” setting in the comparison table. For multi-stage methods such as SparseVLM and PyramidDrop, does 64 refer to the final number of visual tokens or the average token budget across all LLM layers?
Since ZOO-Prune retains only 64 tokens before entering the LLM, while these multi-stage methods may process more tokens in earlier layers but retains fewer than 64 tokens in later layers, could this put it at a disadvantage in deep-layer reasoning compared with ZOO-Prune, which consistently keeps 64 tokens? Could you clarify how this difference was considered in the comparison?
Hi, thank you for the great work! I would like to clarify the “64 retained tokens” setting in the comparison table. For multi-stage methods such as SparseVLM and PyramidDrop, does 64 refer to the final number of visual tokens or the average token budget across all LLM layers?
Since ZOO-Prune retains only 64 tokens before entering the LLM, while these multi-stage methods may process more tokens in earlier layers but retains fewer than 64 tokens in later layers, could this put it at a disadvantage in deep-layer reasoning compared with ZOO-Prune, which consistently keeps 64 tokens? Could you clarify how this difference was considered in the comparison?