Splitting a search agent's planning from its write-up lifts an 8B model

single source· 1 articles · confidence: medium · first seen 2026-09-23 20:00 UTC

What this means for you

Nothing to act on yet: this is a preprint, and the abstract does not say whether the 8B model's weights are released. If you run a search agent, the transferable part is architectural — keep the plan and the running summary in separate prompts — and you can test that without retraining anything.

IterSynth splits a deep-search agent (one that gathers evidence over many steps before answering) into two roles: a planner that decides what to look for next, and a summariser that folds each find into a running summary. The authors say this reduces context noise. Trained with reinforcement learning that scores individual turns alongside the final answer, IterSynth-8B averages 50.7 across five search benchmarks, including BrowseComp; the paper reports that as 4.2 above the strongest previous agent at or below 8B parameters. The scheme also improves frontier proprietary models used as prompts, the authors say. No evaluation date or harness is given.

Key facts

  • ·IterSynth splits a deep-search agent into a planner and a synthesizer, using an evolving summary as the persistent state of search. source
  • ·IterSynth-8B averages 50.7 across five long-horizon deep-search benchmarks including BrowseComp and Xbench-DS. source
  • ·The paper reports that average as 4.2 above the strongest prior agent at or below 8B parameters. source
  • ·Training uses Role-Decoupled Policy Optimization, which combines terminal outcome rewards with turn-level rubric evaluations. source
  • ·The method is also presented as a model-agnostic prompting paradigm that improves frontier proprietary models zero-shot, according to the authors. source
  • ·No evaluation date or harness is given for the reported benchmark scores. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire