Language-model agents often compact their context to handle a current question efficiently. But the same compacted context may need to support follow-up questions later. A new arXiv paper argues that compaction should be judged not only on the immediate query but also on these later questions.
The authors evaluate context compaction on later questions using four corpora that annotate the relevant information. This shifts the evaluation focus from single-turn performance to multi-turn usefulness, where the cost of losing details may only become apparent after the first exchange.
The abstract does not report specific results, so the main contribution at this stage is the evaluation framework itself. By measuring how well compacted context serves future questions, the work highlights a potential blind spot in current compaction methods.