IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
IterSynth decouples planning and synthesis for deep-search agents, and an 8B model scores 50.7.
IterSynth splits deep search into a Planner that identifies information needs and a Synthesizer that folds evidence into an evolving summary, avoiding ReAct-style role coupling and noisy accumulated context. Training uses Role-Decoupled Policy Optimization, combining terminal outcome rewards with turn-level rubric scores and role-specific advantages. On five long-horizon benchmarks including BrowseComp and Xbench-DS, IterSynth-8B averages 50.7, surpassing the strongest prior agent at or below 8B by 4.2%. The same design also works as a model-agnostic prompting paradigm with zero-shot gains on frontier proprietary models.
- A Planner finds information needs; a Synthesizer updates a persistent summary.
- RDPO mixes outcome rewards, turn-level rubrics, and role-specific advantages.
- IterSynth-8B averages 50.7, ahead of the best prior ≤8B agent by 4.2%.
- As a prompting pattern it also improves frontier models over ReAct zero-shot.
Full article186 words · extracted from huggingface.co · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.29444