Tencent's IterSynth-8B tops sub-8B deep-search agents by +4.2

IterSynth-8B, an open-weight deep-search agent from Tencent, averages 50.7 across five long-horizon benchmarks. That is +4.2 over the strongest prior sub-8B agent, MiroThinker-v1.0-8B, per the arxiv preprint.

The design splits what most ReAct-style agents cram into one policy. A Planner identifies information needs, then a Synthesizer folds evidence into an evolving summary that carries state across turns. The paper's framing: "Deep search extends LLMs from passive retrieval to active knowledge construction. Given a complex query, a deep-search agent must decompose the problem, issue searches, read evidence, refine information needs, and synthesize a grounded answer."

The training stacks: SFT 44.1, then outcome-only GRPO 48.9, then RDPO 50.7. RDPO stands for Role-Decoupled Policy Optimization and combines terminal outcome rewards with turn-level rubric evaluations. On BrowseComp-ZH the model scores 55.4, on GAIA text-only 55.3, and on the two xBench-DS splits 66.0 and 46.0.

The transfer claim is the interesting one. As a training-free prompting scaffold with no parameter updates, IterSynth averages 66.1 on Claude-4.5-Opus and 47.9 on DeepSeek-V3.1, with zero-shot gains over ReAct of +5.5 and +4.5 respectively, and up to +10.0 on BrowseComp-ZH.

The paper reports single-run averages in its tables and does not publish latency or inference-cost breakdowns for the two-role loop.