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 and 4 others
Zhejiang University · Tencent
Published September 24, 2026 · on the feed September 25, 2026 · ▲ 9 · arXiv 2609.29444
The abstract
Deep search requires LLM agents to decompose complex queries, search for evidence, and synthesize grounded answers, yet existing ReAct-style agents suffer from two limitations: role coupling, where one policy must handle planning, evidence use, and synthesis; and context accumulation, where growing search histories introduce noise and obscure useful information. To address these issues, we propose IterSynth, a role-decoupled and summary-based paradigm that alternates between a Planner for identifying information needs and a Synthesizer for integrating evidence into an evolving summary state. This design separates planning from synthesis while using the summary as the persistent state of search, reducing both capability coupling and context noise. To train IterSynth effectively, we further introduce Role-Decoupled Policy Optimization (RDPO) for reinforcement learning, which combines terminal outcome rewards with turn-level rubric evaluations and computes role-specific advantages for more precise credit assignment. Experiments on five long-horizon deep-search benchmarks such as BrowseComp and Xbench-DS show that IterSynth-8B achieves an average score of 50.7, surpassing the strongest prior leq8B agent by +4.2\%. Moreover, IterSynth serves as a model-agnostic prompting paradigm, delivering substantial zero-shot gains over ReAct and similar prompting paradigms on frontier proprietary models.
The Brief
2:27A short podcast summary of the paper. Written and voiced by AI.
Transcript
IterSynth. Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis. By Xingyu Wu and eleven others, from Zhejiang University and Tencent. Published on the twenty-fourth of September. Its code repository is public and real: about thirty-five source files.
Deep search asks a language model agent to break a complex query apart, go and find evidence, and synthesize an answer grounded in what it found. Agents in the ReAct style handle this with one policy and one growing history, and the authors name two limitations that follow. The first is role coupling: a single policy has to do the planning, the use of evidence, and the synthesis. The second is context accumulation: as the search history grows, it brings in noise and buries the information that was actually useful.
IterSynth answers both with a role-decoupled paradigm built around a summary. It alternates between two roles. A Planner identifies what information is still needed. A Synthesizer integrates the evidence that comes back into an evolving summary state. Planning is kept apart from synthesis, and the summary, rather than the raw history, is the persistent state of the search. The first choice reduces capability coupling. The second reduces context noise.
Training a system with two roles raises its own question: when an episode goes well or badly, which role deserves the credit? For that the authors introduce Role-Decoupled Policy Optimization, or RDPO, a reinforcement learning method. It combines rewards for the terminal outcome with rubric evaluations at the level of individual turns, and it computes advantages that are specific to each role, aiming at more precise credit assignment.
The evaluation covers five long-horizon deep-search benchmarks, including BrowseComp and Xbench-DS. IterSynth-8B reaches an average score of fifty point seven, which the abstract puts at four point two percent above the strongest prior agent of eight billion parameters or fewer. A second result concerns the paradigm without any training. Used purely as a prompting scheme, IterSynth is model-agnostic, and on frontier proprietary models it delivers substantial zero-shot gains over ReAct and similar prompting paradigms.
Two separations carry the design. One is between the roles, so planning and synthesis are no longer one job carried by one policy. The other is between what the agent has read and what it keeps, so the working state is a summary rather than a transcript. The prompting result suggests the structure helps on its own, before RDPO trains anything into it.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is an agent drowning in its own search history.
You are building an agent that answers hard questions by searching the web. It must break the question apart, search, read evidence and write a grounded answer.
The road ahead
- 1 · The problem ~7 min A single search agent that plans, searches, reads and writes in one growing history gets overloaded and loses useful evidence in the noise.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min One shared model takes turns as a planner and a synthesizer, and a running summary replaces the raw history as the state of the search.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each round, the Planner names a gap, a search fills it, the Synthesizer folds the evidence into the summary, and finally the Planner answers.
- 6 · The experiment ~6 min You’ll guess: On average across long-horizon deep-search benchmarks, by how much did the trained small agent beat the strongest prior agent of its size?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.