Amazon AGI researchers have shown that a reinforcement learning technique called GRPO can improve long-context model performance by up to 20 points. The study, published on arXiv, introduces the RetrievalAttention method enabling models to retain critical information despite KV-cache compression. Models trained on 32,000-token contexts generalized to up to 1 million tokens. On HotpotQA and 2WikiMultihopQA benchmarks, improvements reached up to 20 points in-domain and 9 points out-of-domain compared to RAG systems. The authors made their code publicly available on GitHub.
Source: Read the original article
