The field of vision and language models is moving towards more efficient processing techniques to reduce computational costs and improve performance. One of the main directions is the development of token pruning methods that can adaptively allocate computation based on image complexity or user query relevance. These methods aim to balance the trade-off between efficiency and accuracy, allowing for faster inference times without compromising essential visual information. Another area of focus is the improvement of attention mechanisms, which are a critical component of foundation models. The introduction of sparse attention and pyramid sparse attention is mitigating the information loss associated with high sparsity, enabling more efficient video understanding and generation. Noteworthy papers in this area include: MambaScope, which proposes a coarse-to-fine scoping approach for efficient vision processing. Script, which introduces a graph-structured and query-conditioned semantic token pruning method for multimodal large language models. VLM-Pruner, which presents a centrifugal token pruning paradigm that balances redundancy and spatial sparsity for efficient vision-language models. Teaching Old Tokenizers New Words, which proposes continued BPE training and leaf-based vocabulary pruning for efficient tokenizer adaptation. PSA, which presents a pyramid sparse attention module for efficient video understanding and generation.