IterSynth Splits the Deep Search Agent in Two, and an 8B Model Wins
ReAct-style deep search agents ask one policy to do three jobs, plan the next query, read evidence, and write the answer, while dragging an ever-longer search history behind it. IterSynth (arXiv 2609.29444, Tencent) fixes both problems with one structural change. Split the agent into a Planner that decides what information is still missing and a Synthesizer that folds new evidence into an evolving summary. The summary, not the raw history, becomes the persistent state of the search.
Training gets its own trick. Role-Decoupled Policy Optimization combines the final outcome reward with turn-level rubric scores and computes separate advantages for each role, so the Planner is credited for good questions and the Synthesizer for good integration instead of both sharing one blurry reward.
Results: across five long-horizon deep search benchmarks including BrowseComp and Xbench-DeepSearch, IterSynth-8B averages 50.7, 4.2 percent above the strongest prior agent at 8B or below. The part that should interest people who never train anything: used as a pure prompting pattern on frontier proprietary models, zero-shot, it still delivers substantial gains over ReAct.
This is the same lesson the harness papers keep teaching from different angles. Context is the scarce resource, and the fix is structure, not more tokens. A rolling summary as state is cheap to adopt tomorrow in any search agent, with or without the RL. Code at github.com/Tencent/IterSynth.
← Back to all articles
Training gets its own trick. Role-Decoupled Policy Optimization combines the final outcome reward with turn-level rubric scores and computes separate advantages for each role, so the Planner is credited for good questions and the Synthesizer for good integration instead of both sharing one blurry reward.
Results: across five long-horizon deep search benchmarks including BrowseComp and Xbench-DeepSearch, IterSynth-8B averages 50.7, 4.2 percent above the strongest prior agent at 8B or below. The part that should interest people who never train anything: used as a pure prompting pattern on frontier proprietary models, zero-shot, it still delivers substantial gains over ReAct.
This is the same lesson the harness papers keep teaching from different angles. Context is the scarce resource, and the fix is structure, not more tokens. A rolling summary as state is cheap to adopt tomorrow in any search agent, with or without the RL. Code at github.com/Tencent/IterSynth.
Comments