IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
Xingyu Wu, Yuchen Yan, Zhengxi Lu, Siqi Chen, Xin ZHANG, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Jian Shao, Jun Xiao, Yongliang Shen
cs.CL, cs.AI
2026-09-24
IterSynth splits Planner and Synthesizer over a summary. The 8B agent averages 50.7 on five benches, +4.2 over the best ≤8B prior, +5.5 over ReAct on Claude.
Deep search still asks one ReAct policy to plan, retrieve, and write the answer. Those skills fight each other, and the context fills with noisy pages. In the appendix, ReAct with a 64K window fails to finish before context exhaustion on more than 59% of BrowseComp traces. Multi-model crews are expensive. External summarizers often drop the evidence that later steps need.
IterSynth runs one shared policy under two role prompts. The Planner sees only the question and the current global summary, then searches or answers. It never reads raw hits. The Synthesizer may only rewrite the summary. The persistent state is that summary, not the growing transcript.
Training has two stages. Qwen3.5-397B-A17B rolls out Planner-Synthesizer traces in a live search environment. After repairing bad tool calls and dropping hallucinated steps, about 10K correct traces remain for full-parameter SFT on Qwen3-8B. Role-Decoupled Policy Optimization (RDPO) then adds RL: a 0/1 outcome reward is broadcast to every turn, plus rubric scores (Planner: gap, specificity, termination; Synthesizer: faithfulness, coverage, consolidation). Advantages are normalized inside each role before a standard GRPO update. Mixing roles in one baseline entangles dense Synthesizer rewards with sparse Planner rewards.
IterSynth-8B averages 50.7 across five long-horizon benches, above MiroThinker-v1.0-8B at 46.5. BrowseComp-ZH is 55.4, +15.2 over the best small agent. xBench-DS-2510 is 46.0 versus 34.0. English BrowseComp is 30.9, a hair below MiroThinker-8B at 31.1. GAIA text-only is 55.3 versus 66.4. The average win is concentrated in Chinese deep search and the newer xBench split.
| Method | Average | BrowseComp-ZH |
| MiroThinker-v1.0-8B | 46.5 | 40.2 |
| IterSynth-8B | 50.7 | 55.4 |
| Tongyi-DR-30B | 58.2 | 46.7 |
As a prompt, IterSynth scores 66.1 on Claude-4.5-Opus versus 60.6 for ReAct and 63.2 for IterResearch. Ablations: SFT 44.1, outcome-only GRPO 48.9, same composite reward without role grouping 47.2, RDPO 50.7. Replacing the Planner with untrained Qwen3-8B drops a three-bench average from 58.9 to 17.8; replacing the Synthesizer drops it to 41.5. Planning is the brittle role.
The workflow already helps as a prompt. The 8B training just puts that structure into weights. The piece to copy is "summary as state plus per-role advantages," not another 30B search model. Gains are largest on Chinese deep search. English BrowseComp and GAIA still trail the strongest same-size baseline.
There is no standalone limitations section. GAIA and English BrowseComp do not beat MiroThinker-8B, so the headline average leans on the Chinese subset. Turn rewards come from an LLM judge. RL uses cached retrieval while evaluation uses live tools. Planner and Synthesizer still share one 8B backbone; two separate models were not tested.