arXiv cs.CL
·
18 hours ago
Researchers introduced TTARAG, a test-time adaptation method that adjusts language model parameters during inference to improve Retrieval-Augmented Generation system performance in specialized domains. The method was evaluated across six specialized domains and showed substantial performance improvements over baseline RAG systems. This approach enables RAG systems to automatically adapt to target domains when distribution shifts occur, reducing the performance gap between general and specialized applications.
arXiv cs.CL
·
18 hours ago
Researchers evaluated continual learning methods for medical visual question answering systems that must adapt to new clinical tasks without forgetting previous knowledge. The study tested existing continual learning approaches across five different task types—classification, multi-label classification, detection, cell counting, and report generation—measuring catastrophic forgetting, task ordering sensitivity, and weight drift patterns. Current continual learning methods showed difficulties maintaining the balance between learning new tasks and retaining old knowledge when tasks had different objectives and supervision formats.
arXiv cs.CL
·
18 hours ago
Researchers introduced KnowAct-GUIClaw, a framework that enhances GUI automation agents by adding cross-platform support and self-improving memory systems that learn from user interactions. The system achieved 64.1% accuracy on the MobileWorld benchmark across Android, iOS, HarmonyOS, and Windows platforms, outperforming existing agents and models. The framework enables continuous performance improvement through accumulated interaction experience and transferable skills that work across different base models with an 8.5% performance gain using Kimi-2.6.
