Why use a synthetic key-value retrieval task in addition to multi-document QA to evaluate positional effects?

The multi-document QA task is insufficient on its own because one could argue that middle chunks are simply harder to reason about. The synthetic key-value task acts as a control by removing reasoning requirements. When both tasks show the same positional performance degradation, it elevates the finding from a benchmark-specific issue to a fundamental property of the model’s context-reading mechanism.