Incremental Open-Ended Deep Research with Structured Harness
The paper proposes updating long-running research reports in place instead of regenerating them from scratch.
Its Structured Harness breaks reports into outlines, sections, and supporting evidence so systems can reuse valid material and revise only what changed. The authors test the setup over a ten-year temporal framework with single-step and long-chain update tasks. In their reported DeepResearch Bench results, the method keeps report quality competitive while cutting token use by 33% and search calls by 61%. HF Daily Papers' note
Its Structured Harness breaks reports into outlines, sections, and supporting evidence so systems can reuse valid material and revise only what changed. The authors test the setup over a ten-year temporal framework with single-step and long-chain update tasks. In their reported DeepResearch Bench results, the method keeps report quality competitive while cutting token use by 33% and search calls by 61%. HF Daily Papers' note
score 4