Efficient Tree Cross Attention for Token-Efficient Inference

Tree Cross Attention (TCA) offers a logarithmic number of tokens for efficient inference, outperforming Perceiver IO while using the same number of tokens.