Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation

Mufei Li, Dongqi Fu, Limei Wang +10

Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased re…