Stuffed Mamba: Oversized States Lead to the Inability to Forget

A new retrieval paradigm called State-Centric Retrieval has been proposed, which integrates embedding models and rerankers through reusable states, enhancing the efficiency of Retrieval-Augmented Generation (RAG) systems. This approach involves fine-tuning an RWKV-based large language model to create EmbeddingRWKV, a unified model that optimizes the retrieval process by minimizing redundant computations.

The introduction of SfMamba marks a significant advancement in source-free domain adaptation (SFDA), addressing the challenges of adapting models to unlabeled target domains without access to source data. This framework enhances the selective scan mechanism of Mamba, enabling efficient long-range dependency modeling while tackling limitations in capturing critical channel-wise frequency characteristics for domain alignment.

The introduction of HiFi-Mamba, a dual-stream Mamba-based architecture, aims to enhance high-fidelity MRI reconstruction from undersampled k-space data by addressing key limitations of existing Mamba variants. The architecture features stacked W-Laplacian and HiFi-Mamba blocks, which separate low- and high-frequency streams to improve image fidelity and detail.