IterSynth splits deep-search agents into a Planner and a Synthesizer that share one evolving summary, and its 8B model scores 50.7

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.29444 replaces the single ReAct policy with a Planner that identifies information needs and a Synthesizer that folds evidence into a running summary, which serves as the only persistent state instead of the growing search history. Trained with Role-Decoupled Policy Optimization (outcome rewards plus turn-level rubrics), IterSynth-8B averages 50.7 across five long-horizon benchmarks including BrowseComp and Xbench-DS, 4.2% above the strongest prior agent at 8B or smaller. The authors say the split also works as a prompting pattern without the training.