The success of Vision-Language Models (VLMs) often relies on high-resolution schemes that preserve image details, while these approaches also generate an excess of visual tokens, leading to a substantial decrease in model efficiency. A typical VLM includes a visual encoder, a text encoder, and an LLM. Recent studies suggest pruning visual tokens based on visual and textual priors to accelerate VLMs without additional training costs. However, these methods often overlook prompt semantics or suffer from biased self-attention in the LLM. Inspired by the efficient mechanisms of the human brain for multimodal understanding, we introduce AdaV, a novel training-free visual token pruning method. By emulating the neural pathways that preprocess visual and auditory information before the reasoning stage, we shift text-guided visual attention redirection to the pre-LLM stage, which reduces biased token pruning and enhances model robustness with a limited visual token budget. A Self-adaptive Cross-modality Attention Redirection (SCAR) module is further proposed that effectively merges and redirects visual attention with text-to-image attention. Extensive experiments on seven challenging benchmarks demonstrate that our AdaV achieves SOTA performance in training-free VLM acceleration and can be plug-and-play on various VLMs. We plan to open-source the code upon publication.

Despite the recent efforts from the NLP community, balancing the training budget, downstream performance, and general capabilities of large language models (LLM) remains a challenge in many applications. Training the entire model for downstream tasks is expensive, and could easily result in catastrophic forgetting. Parameter-efficient fine-tuning (PEFT) could reduce the training cost, but it still suffers from forgetting, and limits the learning on the downstream tasks. To address the aforementioned issues, we propose a novel mixture of expert (MoE) framework based on Soft LoRA and Identity Mixture (SLIM). SLIM allows dynamic routing between LoRA adapters and identity layers, thus enabling the bypass of LoRA adapters to suppress forgetting of general capacity. We adopt weight yielding with sliding clustering for better out-of-domain distinguish to enhance the routing. We also convert the mixture of LoRA adapters to the model merging formulation and introduce dynamic merging with its fast implementation for LoRA adapters to keep the general capabilities. Extensive experiments demonstrate that the proposed SLIM is comparable to the state-of-the-art PEFT approaches on the downstream tasks while achieving the leading performance in mitigating catastrophic forgetting. We plan to open-source the code upon publication.