Distill the retrieved set before it hits the model: a context refiner fixes turn-by-turn interference in search agents
arXiv 2608.10743·medium signal
Search agents accumulate irrelevant documents every turn, and that retrieved noise actively distracts the model rather than merely wasting tokens. The authors add a distill-based context refiner into the RL training pipeline and improve both reliability and efficiency on long-horizon search. The framing is the useful part for builders: treat per-turn retrieval output as something to refine before it enters context, not as free-to-append evidence.