Long context tasks can be thought of as a 2-stage problem: retrieving necessary information from context + carrying out the actual task Existing benchmarks only make the 1st retrieval step hard because it is easy to pad the input search space. The second stage is often a very