IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
IterSynth splits deep-search agents into planning and synthesis roles, keeping only an evolving summary as search state.
The paper argues that ReAct-style agents are hurt by role coupling and noisy context buildup during long searches. Its proposed RDPO training method assigns credit separately to the Planner and Synthesizer using outcome rewards and turn-level rubric checks. Across five benchmarks, IterSynth-8B reports a 50.7 average score, 4.2 points above the strongest prior agent at 8B parameters or smaller. The authors also present it as a prompting pattern that improves zero-shot results over ReAct-style methods on frontier proprietary models. HF Daily Papers' note
The paper argues that ReAct-style agents are hurt by role coupling and noisy context buildup during long searches. Its proposed RDPO training method assigns credit separately to the Planner and Synthesizer using outcome rewards and turn-level rubric checks. Across five benchmarks, IterSynth-8B reports a 50.7 average score, 4.2 points above the strongest prior agent at 8B parameters or smaller. The authors also present it as a prompting pattern that improves zero-shot results over ReAct-style methods on frontier proprietary models. HF Daily Papers' note
score 5