IterSynth addresses two limitations of ReAct-style deep-search LLM agents — role coupling (one policy handling planning, evidence use, and synthesis) and context accumulation (noisy, growing search histories) — by alternating between a dedicated Planner that identifies information needs and a Synthesizer that folds evidence into an evolving summary state, which serves as the agent’s persistent memory of the search rather than raw history. The authors introduce Role-Decoupled Policy Optimization (RDPO), a reinforcement-learning method for training the two roles jointly.
