Deep Search Agents Should Rewrite a Summary, Not Pile Up History: Zhejiang University and Tencent's IterSynth Scores 50.7 at 8B

Researchers from Zhejiang University and Tencent posted "IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis" (arXiv 2609.29444) on September 24, 2026, proposing a deep search agent in which one model alternates between a Planner role and a Synthesizer role and treats a continuously rewritten summary, rather than the raw search history, as its memory. IterSynth-8B, trained on Qwen3-8B, averaged 50.7 across five long-horizon benchmarks including BrowseComp, 4.2 points above the strongest prior agent at 8B or below, MiroThinker-v1.0-8B (46.5). Applying the same two-role structure as a prompt, with no training at all, lifted Claude-4.5-Opus from an average of 60.6 to 66.1. ASAP summarizes the design and results from the paper itself and then reads what it means for teams building agents.

ReAct agents failed to finish more than half of BrowseComp runs even with 64K of context

IterSynth targets two weaknesses of the ReAct recipe that most open-source deep search agents follow. The first is role coupling: a single policy has to plan, write queries, filter evidence, resolve conflicts, and synthesize the final answer. The second is context accumulation: retrieved passages, intermediate thoughts, and partial conclusions keep piling into one context, burying useful evidence and letting early mistakes propagate.

A figure from the paper's appendix shows how large the problem is. Even with a 64K context, over 59% of ReAct trajectories on BrowseComp failed to produce an answer before the context ran out. In the trajectory statistics, ReAct used more than 12 search rounds and more than 15 search calls per query and still hit a 59% context-exhaustion rate, while IterSynth used 7 to 8 rounds and 8 to 10 search calls and held exhaustion to 5%. The diagnosis is that a large share of deep search failures come not from weak reasoning but from failing to manage memory.

One model plays two roles, and the summary becomes the memory

Each IterSynth round has a Planner sub-step and a Synthesizer sub-step. The Planner sees only the original question and the current global summary, and decides whether to issue a new query or return a final answer; it never sees earlier queries, raw retrieved documents, or the full reasoning history. When the Planner searches, the Synthesizer reads the returned evidence, filters irrelevant passages, extracts findings, resolves inconsistencies, and rewrites the summary. The Synthesizer can only update memory; it cannot search or answer.

The key point is that the two roles are not separate models. One set of weights plays both, distinguished only by role prompts and allowed actions. Because each round's context is rebuilt from the question and the summary, it stays bounded at 32K. The paper argues this delivers the specialization benefit of multi-agent systems and the context control of summary-based agents within a single model. It also tests the alternative: two separately fine-tuned 8B models, one per role, averaged 40.6 after SFT, 3.5 points below the shared model (44.1), and splitting the same 8 GPUs across two models (4+4) added 13 to 15% to inference time.

Rewards are split by role too: RDPO

IterSynth-8B is trained in two stages, supervised fine-tuning on Qwen3-8B followed by a new reinforcement learning method called RDPO. Qwen3.5-397B-A17B first rolled out the Planner and Synthesizer loop in a live search environment; the team repaired tool-call errors, kept only trajectories ending in correct answers, and ran full-parameter SFT. For the second stage, the authors introduced Role-Decoupled Policy Optimization (RDPO). A final-answer correctness reward is broadcast to every turn, and an LLM judge adds a turn-level score for the Planner and Synthesizer outputs separately against five-item rubrics. Claude-4.6-Sonnet distilled those rubrics from pairs of successful and failed trajectories, and Gemini-2.5-flash-lite scored turns during training.

The crux of RDPO is that advantages are normalized separately for each role rather than across a mixed pool of Planner and Synthesizer samples. The ablation shows why that matters. SFT alone averaged 44.1, outcome-only GRPO reached 48.9, and RDPO reached 50.7. A variant with the same composite reward but mixed-role normalization fell to 47.2, below plain GRPO. The paper's reading is that dense Synthesizer rewards and sparse Planner rewards entangle their credit signals when they share one baseline. The authors estimate the judge cost for the entire RDPO stage at about $192.

The 8B leaderboard needs to be read benchmark by benchmark

The 50.7 average is only accurate when broken down. IterSynth-8B's biggest lead came on the Chinese benchmark BrowseComp-ZH, where it scored 55.4, 15.2 points above MiroThinker-v1.0-8B (40.2). It was also the best small agent on xBench-DS-2510 at 46.0. On English BrowseComp, however, it scored 30.9, just under MiroThinker-v1.0-8B (31.1), and on text-only GAIA it scored 55.3, below both MiroThinker-v1.0-8B (66.4) and the 4B AgentCPM-Explore-4B (63.9). A large part of the average advantage comes from a single Chinese benchmark.

The training data offers a plausible reason. The paper says English questions came mainly from public datasets, while Chinese questions came from REDSearcher and the team's own synthesized real-world instances. A strong Chinese result from a Tencent-affiliated study is not surprising, but whether the gains carry over to languages absent from training, such as Korean, needs separate testing.

The size comparison also deserves care. The paper stresses that IterSynth-8B beats ReSum-30B, AgentFold-30B-A3B, and OpenSeeker-30B-SFT on average, yet the same table lists Tongyi-DR-30B at 58.2 and MiroThinker-v1.7-mini at 69.4. The authors themselves frame the Tongyi-DR-30B comparison as lower accuracy but faster inference on identical hardware, 204.7 versus 312.5 minutes on BrowseComp-ZH. "Better accuracy per unit of compute" is a more accurate reading than "beats 30B models."

The structure's effect shows most cleanly in the training-free prompting test

The result closest to practice is the prompting experiment that touches no parameters. Applying only IterSynth's two-role prompts to Claude-4.5-Opus raised its four-benchmark average from 60.6 under ReAct to 66.1, and on BrowseComp-ZH from 60.2 to 70.2. DeepSeek-V3.1 rose from 43.4 to 47.9. Both models also averaged higher than with IterResearch, an earlier summary-based context reconstruction workflow.

The result matters because it removes a confound: training data quality. The authors also report a controlled comparison in which ReAct trajectories were built from the same questions, teacher model, and pipeline. ReAct scored 32.6 versus IterSynth's 44.1 after SFT, and 38.1 versus 50.7 after reinforcement learning. Evidence that the structure itself drives much of the gain comes from two directions.

There is an exception in the prompting results. On GAIA with Claude-4.5-Opus, IterResearch scored 66.0, above IterSynth's 61.2. That hints that repeatedly rewriting a summary can cost accuracy on tasks that chain many facts, which ties into the failure analysis below.

When the summary is the memory, the summary's mistakes are memory too

The two failure modes the authors report pinpoint the weakness of this design. The first is summary information loss: on GAIA, the Synthesizer sometimes recorded two facts correctly but dropped the relationship linking them, so the agent answered wrongly despite having found everything it needed. The second is incorrect evidence consolidation: on BrowseComp, when search results conflicted, the Synthesizer sometimes committed a superficially authoritative but wrong source to the summary, after which later rounds sought corroborating rather than contradicting evidence, a form of confirmation bias.

ReAct drags every raw document along, so in principle it can revisit them. IterSynth discards the raw text and keeps only the summary, so once the summary is wrong, the evidence to correct it is gone. In the role-swap ablation, replacing the Planner with untrained Qwen3-8B dropped the three-benchmark average from 58.9 to 17.8, and replacing the Synthesizer dropped it to 41.5. Beyond the paper's reading that planning is the harder role, this also shows a structural fragility: if summary quality collapses, the whole agent does. The authors propose adding contradiction detection to the Synthesizer's update step.

Design lessons for teams building agents in Korea

For Korean teams building RAG chatbots or internal research agents, the paper's main value is a design principle rather than a training recipe. Because the gains were shown with prompting alone, teams can immediately try splitting "decide where to search" and "fold new evidence into the summary" into separate calls on the commercial model they already use, rebuilding each round's context from the question and summary. The code is on GitHub at Tencent/IterSynth.

The paper also flags what to watch. A summary-only design needs a source URL attached to every claim in the summary, and a rule that conflicting evidence is recorded as a conflict rather than overwritten. A structured summary that records relationships between facts is another improvement the authors suggest. The authors state that their released artifacts are for research and not recommended for user-facing deployment without additional safety and quality assurance, and the paper is a preprint that has not yet been peer reviewed.

Source: Xingyu Wu, Yuchen Yan, Zhengxi Lu, Siqi Chen, Xin Zhang, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Jian Shao, Jun Xiao, Yongliang Shen et al., "IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis," arXiv 2609.29444v1 (September 24, 2026, Zhejiang University and Tencent). ASAP verified the abstract and body of the paper (method, main results table, ablations, and appendix controlled comparison, efficiency, and failure analysis) directly.