The field of retrieval-augmented generation (RAG) is moving towards more efficient and effective methods for integrating external context into large language models (LLMs). Recent developments have focused on improving the faithfulness and accuracy of RAG systems, with a particular emphasis on addressing the challenges of noisy and long-tail queries. Researchers are exploring new approaches to context management, such as condensation and normalization, to reduce the computational costs and improve the performance of RAG systems. Additionally, there is a growing interest in understanding the internal mechanisms of LLMs and how they integrate retrieved evidence with parametric memory. Notable papers in this area include: RECON, which introduces a framework for compressing evidence within the reasoning loop, improving training speed and inference latency. DyKnow-RAG, a dynamic noisy-RAG framework that adaptively reweights the contributions of different rollout groups to improve the accuracy of query-item relevance modeling. Probing Latent Knowledge Conflict for Faithful Retrieval-Augmented Generation, which proposes a framework for localizing conflicting knowledge and guiding the model to accurately integrate retrieved evidence. The Role of Parametric Injection, a systematic study of parametric retrieval-augmented generation that clarifies the role of parametric injection and recommends jointly using parameterized and textual documents. Grounding Long-Context Reasoning with Contextual Normalization, which introduces a lightweight strategy for adaptively standardizing context representations before generation, improving robustness to order variation and strengthening long-context utilization.