Meta AI has introduced ReasonIR-8B, a retriever model specifically trained for reasoning-intensive information retrieval tasks. Developed using LLaMA3.1-8B, ReasonIR-8B employs a synthetic data generation pipeline that creates challenging queries with plausible but misleading negatives, enhancing its reasoning capabilities. The model outperforms existing retrievers and rerankers on the BRIGHT benchmark, achieving a 36.9 nDCG@10 score while being 200 times more compute-efficient than leading large language model rerankers. Additionally, research on reinforcement learning with verifiable reward (RLVR) demonstrates that large language models can significantly improve mathematical reasoning performance using just one training example. For instance, the Qwen2.5-Math-1.5B model's accuracy on the MATH500 benchmark rose from 36.0% to 73.6% with one-shot RLVR. This approach also shows generalizability across tasks. These advancements highlight progress in efficient training methods and retrieval performance for reasoning tasks in AI models.