Given the high computational cost of ViTs, token compression has emerged as an effective technique to eliminate low-importance tokens. A common strategy for estimating token importance is to use the [CLS] attention scores averaged across all heads in ...
Given the high computational cost of ViTs, token compression has emerged as an effective technique to eliminate low-importance tokens. A common strategy for estimating token importance is to use the [CLS] attention scores averaged across all heads in each block. Notably, prior studies have shown that the [CLS] token in early blocks often attends to arbitrary regions and local features, whereas in the final block, it focuses on task-relevant tokens. This discrepancy lead to the unintended removal of task-relevant tokens during the compression stages in early blocks. Motivated by this issue, we hypothesized that the head average token scoring approach obscures task-relevant signals. We empirically discover that certain heads in early blocks show high similarity to the final block's attention, capturing key task-relevant representations. Based on this insight, we propose a simple and effective method for token compression called Head-Aware Token Selection (HATS). HATS selects heads within each block whose [CLS] attention distributions are similar to that of the final block and then groups these heads to form a combined [CLS] attention map. Consequently, token importance estimates become more stable, and tokens required for the final prediction are better preserved during compression, without requiring any additional training or architectural modifications. Various experiments demonstrate that HATS achieves more reliable estimation of token importance, mitigates accuracy degradation effectively, and can be applied in an off-the-shelf manner to existing token compression frameworks.