Tencent and ZJU Open-Source IterSynth: Decoupled Deep-Search Agent Architecture Outperforms Existing 8B Baselines

Core Background & Industry Pain Points Deep search requires LLM agents to decompose open-ended queries, gather scattered evidence, and produce grounded syntheses. Prevailing ReAct designs conflate planning, tool invocation, and summarization into a single policy, while unbounded raw HTML accumulation saturates context windows and corrupts reasoning attention with irrelevant boilerplate. ### Architecture Highlights & Internals Tencent and ZJU introduce IterSynth, an alternating dual-role paradigm: a Planner identifies information gaps and directs retrieval actions, while a Synthesizer distills newly observed evidence into an evolving summary state. Rather than accumulating monolithic page dumps, the evolving summary acts as a compact persistent state, slashing token bloat by 70%. Role-Decoupled Policy Optimization (RDPO) computes role-specific advantages by blending terminal task rewards with intermediate rubric scoring. ### Authoritative Benchmarks & Measured Scores Evaluated on BrowseComp, Xbench-DS, and multi-hop corpora, IterSynth-8B registers an average score of 50.7 (+4.2% over existing <=8B agents). Used as a model-agnostic prompting framework, it lifts grounded accuracy on proprietary models by 18.6% over standard ReAct while reducing factual hallucination by 53.4%. ### Developer Hands-on Guide Developers can clone https://github.com/Tencent/IterSynth, plug in standard search APIs (Bing, Google, Tavily), and load pre-trained checkpoints from Hugging Face.