The finding
The authors vary the position of relevant information in multi-document question answering and key-value retrieval. Performance often drops when the needed information sits in the middle of a long input, including for models built to accept long contexts. [1]
What it means for pingpong
A sequential chain accumulates material. Evaluate whether its final recommendation preserves the original constraints and relevant objections. A detail can remain in the transcript yet disappear from the final answer.
The limit
This study does not show that context compression is lossless or that shorter inputs always perform better. Any packing strategy needs direct tests of retained facts, sources, and unresolved objections.
Source
Lost in the Middle: How Language Models Use Long Contexts
Nelson F. Liu et al. (2023)