Large Vision-Language Models (LVLMs) achieve remarkable performance in multimodal tasks but suffer from high computational costs due to the large number of visual tokens. Existing pruning methods either apply after visual tokens enter the LLM or perform pre-pruning based solely on visual attention. Both fail to balance efficiency and semantic alignment, as post-pruning incurs redundant computation, while visual-only pre-pruning overlooks multimodal relevance.To address this limitation, we propose Dual-Cue Pruning (DCP), a novel cross-modal pruning framework that jointly considers textual semantics and visual self-attention. DCP consists of a text-aware computation module, which employs a gradient-weighted attention mechanism to enhance text-visual alignment, and an image-aware computation module, which utilizes deep-layer self-attention distributions to retain essential structural information. By integrating both cues, DCP adaptively selects the most informative visual tokens, achieving efficient inference acceleration while maintaining strong task performance. Experimental results show that DCP can retain only 25% of the visual tokens, with a minimal performance degradation of only 0.063% on LLaVA-1.5-13B, demonstrating its effectiveness in balancing efficiency and accuracy.

Neural text generation models are typically trained by maximizing log-likelihood with the sequence cross entropy (CE) loss, which encourages an exact token-by-token match between a target sequence with a generated sequence. Such training objective is sub-optimal when the target sequence is not perfect, e.g., when the target sequence is corrupted with noises, or when only weak sequence supervision is available. To address the challenge, we propose a novel Edit-Invariant Sequence Loss (EISL), which computes the matching loss of a target n-gram with all n-grams in the generated sequence. EISL is designed to be robust to various noises and edits in the target sequences. Moreover, the EISL computation is essentially an approximate convolution operation with target n-grams as kernels, which is easy to implement and efficient to compute with existing libraries. To demonstrate the effectiveness of EISL, we conduct experiments on a wide range of tasks, including machine translation with noisy target sequences, unsupervised text style transfer with only weak training signals, and non-autoregressive generation with non-predefined generation order. Experimental results show our method significantly outperforms the common CE loss and other strong baselines on all the tasks. EISL has a simple API that can be used as a drop-in replacement of the CE loss: https://github.com/guangyliu/EISL.

The impression section of a radiology report summarizes the most prominent observation from the findings section and is the most important section for radiologists to communicate to physicians. Summarizing findings is time-consuming and can be prone to error for inexperienced radiologists, and thus automatic impression generation has attracted substantial attention. With the encoder-decoder framework, most previous studies explore incorporating extra knowledge (e.g., static pre-defined clinical ontologies or extra background information). Yet, they encode such knowledge by a separate encoder to treat it as an extra input to their models, which is limited in leveraging their relations with the original findings. To address the limitation, we propose a unified framework for exploiting both extra knowledge and the original findings in an integrated way so that the critical information (i.e., key words and their relations) can be extracted in an appropriate way to facilitate impression generation. In detail, for each input findings, it is encoded by a text encoder and a graph is constructed through its entities and dependency tree. Then, a graph encoder (e.g., graph neural networks (GNNs)) is adopted to model relation information in the constructed graph. Finally, to emphasize the key words in the findings, contrastive learning is introduced to map positive samples (constructed by masking non-key words) closer and push apart negative ones (constructed by masking key words). The experimental results on two datasets, OpenI and MIMIC-CXR, confirm the effectiveness of our proposed method, where the state-of-the-art results are achieved.