IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

arXiv · Xingyu Wu, Yuchen Yan, Zhengxi Lu, Siqi Chen, Xin ZHANG, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Jian Shao, Jun Xiao, Yongliang Shen · 2026-09-27

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

IterSynth splits planning and synthesis into separate LLM roles and uses a summary state to keep deep-search agents focused and efficient.

What it does

The authors identify two bottlenecks in ReAct-style agents: role coupling (one policy does planning, evidence gathering, and synthesis) and context accumulation (growing histories add noise). IterSynth introduces a Planner that decides what information to fetch and a Synthesizer that merges evidence into a persistent summary, alternating between the two. Training leverages Role-Decoupled Policy Optimization (RDPO), which blends terminal rewards with turn-level rubric scores and computes role-specific advantages for sharper credit assignment.

Experiments on five long-horizon deep-search benchmarks (including BrowseComp and Xbench-DS) show IterSynth-8B scoring 50.7, a +4.2% gain over the strongest ≤8B baseline, and the prompting paradigm yields strong zero-shot improvements on proprietary models.

Why it matters

Decoupling roles reduces interference between planning and synthesis, leading to clearer credit signals and more stable learning. The summary state curtails context bloat, preserving relevant evidence while discarding noise, which is crucial for long-running searches. These design choices translate into measurable performance lifts without extra model parameters.

How it applies

When building an LLM-driven search assistant, you can adopt the IterSynth template: a lightweight Planner model (or prompt) that outputs a query list, and a separate Synthesizer that consumes retrieved snippets into a running summary. The RDPO framework can be plugged into existing RL pipelines to fine-tune the two roles jointly. Existing LLMs can be used in a model-agnostic way, so you get gains even with off-the-shelf APIs.

The short version

Imagine a search agent that never loses its train of thought, no matter how long the query.

My take

I think the role-decoupled approach will become a standard pattern for agent design, much like the separation of encoder and decoder in seq2seq models. The summary state acts as a lightweight working memory, which aligns with emerging research on external memory modules for LLMs. The fact that the method works zero-shot on proprietary models suggests it’s not tied to a particular model size or training regime, making it broadly adoptable.

How it connects

• Role-specific policies echo the modular RL trend where sub-tasks get dedicated networks.
• Summary-based state management parallels recent work on retrieval-augmented generation and long-context transformers.

Bottom line

Split planning from synthesis, keep a concise summary, and you’ll get a deeper-search agent that stays sharp at scale.

Takeaways

• Role coupling in ReAct agents hampers performance; separating planning and synthesis alleviates this.
• A persistent summary state replaces ever-growing search histories, reducing context noise.
• RDPO provides role-specific advantage estimation, improving RL credit assignment.
• IterSynth-8B reaches an average score of 50.7 on five deep-search benchmarks, +4.2% over the best prior ≤8B agent.
• The paradigm works zero-shot on proprietary models, showing model-agnostic prompting benefits.

Topics: deep search, LLM agents, role decoupling, reinforcement learning