DAIR.AI highlights week's top AI papers

This week's selection underscores a massive research pivot toward making AI agents more autonomous, self-correcting, and computationally efficient over long context windows. Self-improving agent frameworks like Self-Harness shift the prompt and memory optimization burden from humans directly to the models. Meanwhile, efficient context architectures like MiniMax Sparse Attention resolve the quadratic compute bottleneck of standard attention, unlocking repository-scale reasoning. Finally, the focus on evaluation in Agents' Last Exam shows that the community requires new frameworks to accurately measure frontier agent capabilities.