Incremental Open-Ended Deep Research with Structured Harness

Incremental Open-Ended Deep Research with Structured Harness. By Meilin Chen and one other, from Xiaohongshu and Zhejiang University. Published on the eighth of October.

Open-ended deep research systems mostly write their reports from scratch. That works for a one-off question. It is inefficient when a report has to be maintained over time, as new information keeps arriving. Each update would mean redoing the research and rewriting the whole document, even the parts that were still correct. The cost grows with every revision, and nothing guarantees that two versions written from scratch will stay consistent with each other. For a report that is supposed to track a subject as it changes, that lack of continuity is a problem of its own.

The authors define a new research setting, Incremental Open-Ended Deep Research. Here a report is treated as an evolving research state. Updating it means three things at once: preserving the knowledge that is still valid, revising content that has become outdated or was incomplete, and incorporating information that has newly become available.

To support that setting they propose Structured Harness. It represents a report as a structured collection of outlines, sections and supporting evidence, instead of a single block of text. On top of that representation it provides structured retrieval, a persistent structured evidence pool, and structured generation. Together these allow selective updating of a report and reuse of evidence gathered earlier. Because sections and their evidence are held as separate pieces, a new finding can be routed to the part of the report it affects, and the rest can be left alone.

Evaluating this needed its own framework. The authors build a temporal evaluation spanning ten years, with two task types. The Single-Step Task measures individual transitions from one version to the next. The Long-Chain Task measures long chains of updates, where errors or drift could accumulate.

Experiments run on DeepResearch Bench and on DeepConsult, under both an open-source configuration and a proprietary configuration. The incremental approach keeps report quality competitive while substantially improving continuity between versions and lowering research costs.

On DeepResearch Bench, compared with research from scratch, it reaches up to zero point five one higher content-level ROUGE-L F1 and up to zero point six three higher outline-level EM F1. It also uses thirty-three percent fewer tokens and makes sixty-one percent fewer search calls.

The shift is in what a report is taken to be. Not a finished artifact produced once, but a state that gets carried forward. Once the report is held as structure, with its evidence kept alongside, an update can touch only what changed, and both the continuity gains and the cost savings come from that.