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, et al.
cs.CL · cs.AI
Quick verdict
Role-decoupled deep-search agent with rubric RL shows strong ablations, but headline margin rides on extra test-time compute and lacks variance reports.
Summary
The paper introduces IterSynth, a deep-search agent paradigm that decouples planning from synthesis within a single shared 8B policy: a Planner issues queries or terminates based on a compact global summary, while a Synthesizer integrates retrieved evidence into that summary each round, bounding context. Training proceeds via cold-start SFT on ~10K teacher-synthesized dual-role trajectories (Qwen3-8B backbone), followed by Role-Decoupled Policy Optimization (RDPO), which adds turn-level rubric rewards from an LLM judge to terminal outcome rewards and normalizes advantages separately per role. Across five benchmarks, IterSynth-8B averages 50.7 versus 46.5 for MiroThinker-v1.0-8B (+4.2 points), with the largest gain on BrowseComp-ZH (55.4, +15.2), while trailing some ≤8B baselines on GAIA (55.3 vs 66.4). Ablations attribute gains to role-decoupled advantage normalization (50.7 vs 47.2 with mixed normalization), architecture over trajectory quality (+11.5 over ReAct-SFT under matched data), and shared over split parameters (+3.5). As pure prompting, IterSynth beats ReAct and IterResearch on Claude-4.5-Opus and DeepSeek-V3.1. Key limitations: the Table 1 headline numbers include a not-attempted resampling strategy that baselines may not use; there are no variance estimates or significance tests despite small validation sets (GAIA n=103, Xbench n=100); one SFT data citation (DeepForge) appears bibliographically wrong; and full reproduction requires 96 H20 GPUs plus paid APIs.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 3
● low 0
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Decoupling planning and synthesis into prompt-separated roles over a bounded global summary lets a single 8B policy obtain multi-agent-like specialization while cutting ReAct's 59% context-exhaustion rate to under 5% (Section 3.1, Appendix F.3 Table 12).
- RDPO's per-role advantage normalization is load-bearing: 50.7 average versus 47.2 for identical rewards under mixed-role normalization and 48.9 for outcome-only GRPO (Table 3).
- IterSynth-8B averages 50.7 across five long-horizon benchmarks, +4.2 points over the strongest prior ≤8B agent (Table 1) — with headline numbers including an inference-time not-attempted resampling step (Appendix J).
- The workflow transfers without any training: as pure prompting it beats ReAct and IterResearch on Claude-4.5-Opus and DeepSeek-V3.1, up to +10.0 on BrowseComp-ZH (Table 2).
- Parameter sharing is a net win rather than a trade-off: one jointly trained 8B model beats two independently fine-tuned 8B role models by +3.5 average (Table 9).
Strengths
Controlled architecture-vs-data confound test
Appendix E (Table 8) rebuilds a ReAct baseline with the same teacher model, data scale and training pipeline: IterSynth-SFT 44.1 vs ReAct-SFT 32.6 (+11.5 avg), and the RL-stage gain is larger for IterSynth (+6.6 vs +5.5), isolating the workflow itself from teacher trajectory quality.
Ablation depth uncommon in agent papers
Role-swap (Table 4: -41.1 avg when replacing the Planner with base Qwen3-8B, -17.4 for the Synthesizer), shared-vs-independent parameters (Table 9: +3.5 for shared), per-role vs mixed-role advantage normalization (Table 3: 50.7 vs 47.2), alpha sweep (Table 13), judge-model swap (Table 14), plus training-dynamics curves (Fig. 5).
Honest cost and failure-mode accounting
Prompting generality on frontier models
Design targets a measured failure mode
Weaknesses
Test-time compute asymmetry in headline results
No variance or significance reporting
Uneven per-benchmark wins behind the average
Baseline comparability unstated
Citation integrity defect
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
IterSynth (Zhejiang University + Tencent; arXiv 2609.29444) is a method paper for LLM deep-search agents. It claims two contributions: (1) a workflow that alternates a Planner and a Synthesizer within one shared 8B policy, using an evolving global summary as bounded persistent state (Section 3.1); and (2) RDPO (Section 3.2.2), which adds turn-level rubric rewards from an LLM judge to terminal outcome rewards and normalizes group advantages separately per role. The headline claim is that IterSynth-8B averages 50.7 across five benchmarks (Table 1), +4.2 points over MiroThinker-v1.0-8B, and beats ReAct and IterResearch as a training-free prompt on two frontier models (Table 2). The extracted text includes Appendices A-K, so most claims are assessable in detail; figures were only partially recoverable.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is IterSynth?
IterSynth is a deep-search agent paradigm in which one shared LLM alternates between two prompt-defined roles: a Planner, which decides the next search query or the final answer from a compact global summary, and a Synthesizer, which folds newly retrieved evidence into that summary, keeping the working context bounded instead of ever-growing.
How much better is IterSynth-8B than previous small agents?
On five benchmarks it averages 50.7 points versus 46.5 for the strongest prior ≤8B agent (MiroThinker-v1.0-8B), a +4.2-point margin; the biggest gain is BrowseComp-ZH (+15.2). On GAIA it trails some small baselines (55.3 vs 66.4). Note the headline numbers include a resampling step at inference time.
Does IterSynth require training, or can I just prompt my model with it?
Both work. The paper trains IterSynth-8B via SFT plus a role-decoupled RL method (RDPO), but also shows that using the Planner-Synthesizer loop purely as a prompt improves Claude-4.5-Opus (+5.5 average over ReAct) and DeepSeek-V3.1 (+4.5) with no parameter updates.
Is the code or model checkpoint available?
The paper lists a code repository (github.com/Tencent/IterSynth) and states that the IterSynth-8B checkpoint, SFT corpus and RDPO pipeline will be released under a permissive research license. Full retraining, however, requires roughly 96 H20 GPUs plus paid search and judge APIs.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 were only partially recoverable from text extraction; numeric readings for Figure 4/5 rely on garbled caption text, though all headline numbers come from clean table text.
- Baseline provenance (reproduced vs quoted) cannot be verified from the text itself; the corresponding criticism rests on the absence of any such statement.
- The DeepForge citation mismatch is judged from internal evidence (title/journal vs stated use) without external lookup; the true dataset identity is unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.