Longer context windows do not straightforwardly improve accuracy

The marketing logic of large context windows is that more input means more grounding. The published evaluation record is less generous than that.

The most cited result in this area is the observation that models retrieve information placed at the start and end of a long context considerably more reliably than information placed in the middle. The effect has been reproduced across model families and has not disappeared as windows have grown.

A second, quieter finding matters as much for practitioners: adding weakly relevant passages can reduce answer quality even when the correct passage is present. Precision in retrieval is worth more than recall once the correct passage is in the set.

The practical reading is that a context window is a budget rather than a container. Filling it because it is available raises cost on every request, raises latency, and can lower accuracy. The teams that measure this generally end up retrieving fewer passages and re-ranking them harder.

None of this argues against long-context models. It argues against the assumption that supplying more material is a substitute for deciding which material matters.