Incremental deep research cuts tokens 33% and search calls 61%
Meilin Chen · hf · 2026-10-09
The paper proposes Incremental Open-Ended Deep Research (Incremental-OEDR): instead of generating reports from scratch, a report is treated as an evolving research state — preserving valid knowledge, revising outdated content, and incorporating new information.
- The Structured Harness framework represents reports as structured collections of outlines, sections, and supporting evidence, with structured retrieval, a persistent evidence pool, and structured generation for selective updates and evidence reuse.
- A temporal evaluation framework spans ten years with Single-Step and Long-Chain tasks, testing individual transitions and long update chains.
- On DeepResearch Bench and DeepConsult (open-source and proprietary configs): quality stays competitive while content-level ROUGE-L F1 rises up to 0.51, outline-level EM F1 up 0.63, with 33% lower token consumption and 61% fewer search calls.
More from coding & agent
- Developer ships first App Store app Save4Two built with Cursor and Grok — chrisfleck · 2026-10-10
- Dev clones Photoshop and the whole Adobe suite in Rust using Claude — iamfakhrealam · 2026-10-10
- When AI Writes the Code, What Replaces Design Patterns? — JnBrymn · 2026-10-10
- Google Cloud launches App Optimize remote MCP server to query cloud costs from any AI assistant — rseroter · 2026-10-10
- Developer jokes: with AI, no more shipping tests — yoobinray · 2026-10-10
- Grok leak hints at merging your Cursor account into X subscription — nima_owji · 2026-10-10