Separating planning from synthesis improves long-horizon search agents

IterSynth is a new paradigm for deep search agents that alternates between a Planner role for identifying information needs and a Synthesizer role for integrating evidence into an evolving summary. This decouples the capabilities of planning and synthesis, reducing context noise and capability coupling. The model is trained with Role-Decoupled Policy Optimization (RDPO), which combines terminal rewards with turn-level rubric evaluations and role-specific advantages for credit assignment. In experiments across five long-horizon benchmarks, IterSynth-8B achieved an average score of 50.7, surpassing the strongest prior agent under 8B parameters by 4.2%. Additionally, IterSynth serves as a model-agnostic prompting paradigm that provides zero-shot gains over ReAct on frontier proprietary models. The work is interesting because it directly addresses key limitations of ReAct-style agents by separating roles to prevent premature termination and redundant searches, and by maintaining a bounded summary state rather than a growing search history.