IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

Deep-search agents that follow the ReAct recipe inherit two coupled bottlenecks: one policy has to plan, use evidence, and synthesize at the same time, while its context keeps growing until it suffocates — over 59% of ReAct trajectories on BrowseComp never even terminate within a 64K budget.

Figure 1: multi-agent specialization, summarization-based context refreshment, and IterSynth — capability decoupling on a single shared policy.

We propose IterSynth, a role-decoupled, summary-based paradigm that alternates two roles instantiated by a single shared parameter set:

• Planner — identifies unresolved information needs and formulates the next sub-query from a compact, persistent state;
• Synthesizer — integrates newly retrieved evidence into that summary, making summary updates a learned part of the search policy instead of an external compression module.

Figure 2: the Planner–Synthesizer loop; each iteration reasons over a bounded, reconstructed workspace.

To train it we introduce Role-Decoupled Policy Optimization (RDPO): a terminal outcome reward augmented with turn-level rubric rewards, with group-relative advantages computed independently per role — more precise credit assignment with no algorithmic change beyond GRPO.

Figure 3: RDPO blends the terminal outcome reward with per-turn rubric scores and normalizes advantages role by role.

Results:

• IterSynth-8B reaches 50.7 avg on GAIA-text-only, xBench-2505/2510, BrowseComp and BrowseComp-ZH — +4.2 over the strongest prior ≤8B agent, and competitive with several 30B-scale agents at less than one third of the parameters.
• Training recipe ablation: SFT 44.1 → outcome-only GRPO 48.9 → RDPO 50.7. Replacing the Planner with an untrained model costs ~41 points, showing that role decoupling is what turns the architecture into reliable long-horizon capability.
• It also works as a model-agnostic prompting paradigm: zero-shot gains over ReAct of +5.5 (Claude-4.5-Opus) and +4.5 (DeepSeek-V3.1) on average, up to +10.0 on BrowseComp-ZH.
• The bounded, reconstructed context translates into fewer search rounds (7–8 vs 12+), a <5% context-exhaustion rate, and 13–15% lower inference time than an equivalent dual-model deployment.

Code: https://github.com/Tencent/IterSynth

Feedback very welcome!