Deep-search agents built on the ReAct recipe face two intertwined bottlenecks, according to a new paper titled IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis. A single policy must simultaneously plan, use retrieved evidence, and synthesize answers — a problem the authors call role coupling — while its context keeps growing until useful information is buried in noise. On BrowseComp, the paper reports that over 59% of ReAct trajectories never terminate within a 64K token budget.
IterSynth addresses both issues with a role-decoupled, summary-based paradigm that alternates between two roles instantiated by a single shared parameter set. A Planner identifies unresolved information needs and formulates the next sub-query from a compact, persistent state, while a Synthesizer integrates newly retrieved evidence into that summary. Crucially, summary updates become a learned part of the search policy rather than an external compression module, and each iteration reasons over a bounded, reconstructed workspace.
To train the system, the authors introduce Role-Decoupled Policy Optimization (RDPO), a reinforcement learning method that combines a terminal outcome reward with turn-level rubric rewards and computes group-relative advantages independently per role. The paper describes this as more precise credit assignment with no algorithmic change beyond GRPO.
Results place IterSynth-8B at an average score of 50.7 across GAIA-text-only, xBench-2505/2510, BrowseComp and BrowseComp-ZH — a gain of 4.2 points over the strongest prior agent at or below 8B parameters, and competitive with several 30B-scale agents at less than one third of the parameters. An ablation traces the training recipe from SFT at 44.1 to outcome-only GRPO at 48.9 and RDPO at 50.7, while replacing the Planner with an untrained model costs roughly 41 points.
The paradigm also works as a model-agnostic prompting scheme, delivering zero-shot gains over ReAct of +5.5 for Claude-4.5-Opus and +4.5 for DeepSeek-V3.1 on average, reaching up to +10.0 on BrowseComp-ZH. The bounded context translates into fewer search rounds (7–8 versus 12+), a context-exhaustion rate below 5%, and 13–15% lower inference time than an equivalent dual-model deployment. Code is available at github.com/Tencent/IterSynth.