Ten papers covering the gap between advertised context window and the context window models can actually use. Lost in the Middle set the agenda with the U-shaped position curve. LongBench, InfBench, RULER, BABILong, NoCha, and Loong each push the harness in a different direction: broader suites, longer contexts, position robustness, reasoning, books, and multi-doc synthesis. LongBench v2 (December 2024, 503 questions from 8K to 2M words) and NoLiMa (2025) close the loop. NoLiMa especially is the paper to cite when a vendor claims a number you do not believe. LongBench Pro (January 2026) is the newest check: 1,500 bilingual samples from 8K to 256K tokens across 46 models, with the same verdict that effective context length falls short of the claimed window, and cross-lingual gaps on top.