Researchers have developed a new approach called GRIP, which aims to improve the performance of retrieval-augmented generation (RAG) models. These models can struggle when the query dominates the latent state, making retrieved evidence irrelevant. GRIP introduces capacity asymmetry by limiting the decoder's access to retrieved evidence, forcing it to encode only information not available from the query. The authors claim that GRIP outperforms strong iterative baselines on five reasoning benchmarks and reduces hallucination by 73%. However, these claims are based on a preprint paper and have not been peer-reviewed. --- Why it matters: This matters because RAG models are widely used in applications such as question-answering and text generation. Improving their performance can lead to more accurate and reliable results, which is crucial for tasks that require precise information retrieval. Source: https://arxiv.org/abs/2608.16776