One policy is carrying everything
An agent asked to research something is holding several jobs at once. It has to break a complicated question apart, decide where to look, and assemble a grounded answer from what it finds. In the common pattern one policy does all of this: think, act, observe, think again. Straightforward enough, but the authors identify two difficulties in it.
The first is role coupling. Planning, using evidence and composing an answer demand different kinds of judgement. Push them into a single policy and one of them gives way. The second is context accumulation. The longer the search history grows, the more noise it carries, and useful passages get buried in it. The primary source for this article is a preprint that has not been peer reviewed.
Make the summary itself the state
The proposal narrows to one idea. Separate the role that identifies what still needs to be known from the role that weaves gathered evidence into a single summary, and alternate between them. The authors call this a role-decoupled, summary-based paradigm and name it IterSynth[1].
The pivot is treating the summary not as output but as state. Each step of search rewrites the summary, and what to look for next is decided by reading that summary. Instead of carrying the full history, the agent moves forward holding only its current understanding. Planning separates from composition, and the noise in the context falls at the same time. Two aims from one move.
There is machinery on the training side as well. Alongside a reward for the terminal outcome, the method combines evaluations of each turn against rubrics, and computes separate advantages per role. The design reads as an attempt to assign more precisely which of the two roles moved the result.
What the abstract states
The authors evaluate on five long-horizon deep-search benchmarks, citing BrowseComp and Xbench-DS as examples. IterSynth-8B reaches an average score of 50.7, which they report as surpassing the strongest prior agent at that scale by +4.2%.
They further state that IterSynth works as a model-agnostic prompting paradigm, obtaining gains without additional training over think-and-act style prompting and similar paradigms, including on frontier proprietary models. That is the extent of the abstract, and it is the authors' claim. It is not a proven conclusion. A public implementation is also released[4].
What this looks like in practice
Applied to the job of pulling together an internal review of some question, the difference in stance looks like this.
One person does all of it
Decide where to look, read the material, write the report straight through. Thinking gets dragged along by reading order, and material that surfaces late is hard to place. The more documents pile up, the vaguer it becomes which ones were actually used.
Split the roles and alternate
First decide only what is still missing. Weave each document into the body of the summary as it arrives. When deciding what is missing next, what you look at is the current summary, not the pile.
The point is not speed. It is that what was understood at which moment survives. If the summary is the state, its intermediate versions are the record of how the work proceeded. Read the other way: a document that never got woven in disappears quietly from every later judgement. A design that does not record what it left out is weak exactly here.
What is not new, and what cannot be read
Separating roles is not new. Splitting a planning role from an executing role has been tried repeatedly in agent research. Carrying a rolling summary to hold the context down is also an old device. What this paper contributes is joining the two into one paradigm and training it with advantages assigned per role.
Much cannot be read from the abstract. Who built the rubrics used to score each turn, and how. How many cases turned on a document that never entered the summary. Across the five benchmarks, where the margin appeared and where it did not. None of these conditions are given. The consequence of a summary being state deserves particular attention: an error in the summary is carried forward unchanged. How such an error is detected is not addressed in the abstract.
Why this subject is clustering now
A paper published in the same window applies reinforcement learning with verifiable rewards to small search agents[3]. The question overlaps. Rather than calling a large model, can an agent of limited scale be taught the research process itself? IterSynth approaches it through role separation and summary-as-state; the other through the design of the reward. Same direction, different handle.
Reader votes stand at 9, comments at 2, and stars indicating implementation interest at 6. For figures shortly after release, these are small. This site measures the spread of a subject through independent channels — votes, implementation, press, and a cluster of papers on the same question — and here only the cluster stands. Few votes are not evidence that a paper is wrong, and many votes would not be evidence that it is right. Both are indicators of attention, not of correctness or importance.
Where this connects to material review
Reviewing promotional material means repeatedly grounding a judgement in outside documents. Which passage of the package insert was checked, whether a cited dataset preserves the conditions of its source, what precedent exists for a given objection. As a process, that is research. And the deliverable is not only the conclusion but the account of where the reviewer looked to reach it.
Translated into a rejection
Reviewer: "This statement goes beyond the approved scope."
Author: "Against which passage of which document was that judged?"
Reviewer: "...It is a judgement based on reading all of it."
An objection that cannot name what it was checked against does not prevent the same objection next time. Carrying a summary as state can work against that gap. If the current understanding always exists as a single document, and what was woven into it can be traced, the account of the judgement moves closer to being the review record itself.
Carried into regulated practice, though, the order reverses. The summary is not made state in order to raise a score or move faster, but in order to preserve which documents were seen and set aside. A design where unwoven documents vanish silently is a mechanism for hiding missing grounds. The score is measurable; the price of a dropped document is measurable only after the fact.