IterSynth Rethinks Deep-Search Agents Through Role Decoupling

Introduction

Deep search is more than issuing a sequence of web queries. An agent must break down a difficult question, determine what information is still missing, formulate useful follow-up searches, and turn scattered evidence into a grounded answer. ReAct-style systems commonly ask one policy to perform all of these jobs. As the trajectory grows, the entire interaction history also becomes part of the context, increasing noise and making useful evidence harder to recover. IterSynth revisits both the division of labor and the way search state is maintained.

Two roles, one shared policy

IterSynth alternates between two explicitly defined roles while using a single shared parameter set:

• Planner: reads a compact current state, identifies unresolved information needs, and proposes the next sub-query;
• Synthesizer: processes newly retrieved evidence and incorporates useful findings into an evolving summary.

The summary is not merely an output compression step at the end of search. It becomes the persistent state that carries the investigation forward. Each new planning turn operates on a bounded, reconstructed workspace rather than the complete raw trajectory. This reduces context noise and gives planning and evidence integration distinct responsibilities without requiring two separately trained models.

RDPO and more targeted credit assignment

A different interaction pattern still needs a suitable training objective. The authors therefore introduce Role-Decoupled Policy Optimization, or RDPO. It combines a terminal outcome reward with turn-level rubric evaluations, then computes group-relative advantages independently for the Planner and Synthesizer roles. A useful query and a beneficial summary update can consequently receive feedback that is closer to their actual function, instead of relying entirely on whether the final answer succeeded.

Results

Across tasks including GAIA text-only, xBench variants, BrowseComp, and BrowseComp-ZH, IterSynth-8B reports an average score of 50.7, 4.2 points above the strongest previous agent at or below the 8B scale. The training ablation moves from 44.1 with SFT to 48.9 with outcome-only GRPO and then to 50.7 with RDPO. The reported drop caused by replacing the trained Planner with an untrained model further suggests that the role structure is contributing real capability rather than serving as a superficial prompt template.

The bounded-summary design is also associated with roughly seven to eight search rounds instead of more than twelve, a context-exhaustion rate below 5%, and 13%–15% lower inference time than an equivalent dual-model deployment. As a prompting paradigm, IterSynth also produces zero-shot gains over ReAct on the evaluated versions of Claude-4.5-Opus and DeepSeek-V3.1.

Why it matters

The broader contribution is a practical view of agent memory: preserve a revisable working summary instead of replaying every search event. This could make research agents less sensitive to context limits and easier to train or deploy. The design also introduces a clear risk: summaries can omit details or carry forward mistakes. Future evaluation will therefore need to examine evidence traceability, summary fidelity, and robustness beyond the reported benchmarks.

Source: Hugging Face Daily Papers