Deep search agents improve planning and answer summarization
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
Computation and LanguageArtificial Intelligence
Summary
When AI tries to answer complicated questions, it needs to plan, look for evidence, and combine information into a clear answer. The authors found that having one AI policy do all these tasks leads to problems and confusion. They created IterSynth, which splits planning and answer building into two separate parts that work together with a summary to keep track of progress. This method helps the AI find better answers by avoiding noise from long search histories. Tests show it works better than previous similar AI agents, even on tough problems.
What this means in practice
- •For chatbot developers: Build chatbots that handle complex queries by separating planning from answer synthesis to improve accuracy and reduce confusion.
- •For digital assistant teams: Create assistants that summarize ongoing searches effectively, maintaining clear context and improving response quality for multi-step queries.
Authors
Xingyu Wu, Yuchen Yan, Zhengxi Lu, Siqi Chen, Xin ZHANG, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Jian Shao, Jun Xiao, Yongliang Shen
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 $\leq$8B 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.