[Interest in long context language models (LLMs) grows with 1M token windows. Multi-Needle + Reasoning benchmark evaluates LLM retrieval performance, showing degradation.]