Brief

Researchers have developed a new method for class-incremental learning (CIL) in audio-visual settings, addressing the challenge of acquiring new knowledge without losing previously learned information. The approach integrates the SAM-Audio multimodal model by using its audio features to guide visual representations through a novel attention strategy. To further combat catastrophic forgetting, the method incorporates dual-level distillation objectives at both feature and logit levels, demonstrating superior performance on audio-visual CIL benchmarks compared to existing state-of-the-art techniques. AI

Researchers have introduced several new frameworks and benchmarks for multimodal retrieval tasks. Dynamic Adapter Routing (DAR) addresses continual multimodal retrieval by using prototype-based routing for adapter selection. V-SPLADE offers an inference-free sparse retriever for visual documents, improving lexical grounding with caption-gated token supervision. HiKEY proposes a hierarchical retrieval framework for document question answering, leveraging document structure for better routing and evidence integration. Additionally, DeepImageSearch frames image retrieval as an autonomous exploration task within visual histories, introducing a new benchmark (DISBench) to evaluate agentic reasoning. AI