LongCat-DeepResearch Technical Report
Meituan LongCat Team, He Zhu, Yue Xu, Wanli Wu, Haolin Ren, Yuxin Bian, Jiarui Zhao, Rongzhi Zhang and 23 others
Meituan
Published September 28, 2026 · on the feed September 30, 2026 · ▲ 63 · arXiv 2609.36071
The abstract
We present LongCat-DeepResearch, a deep research system that combines an enhanced LongCat model with a multi-agent workflow for producing comprehensive, evidence-grounded reports. The workflow separates global planning from detailed investigation and coordinates revision at the section level. Multiple planning agents first explore external sources and refine an actionable research plan, termed ResearchSpec. Research agents then investigate and draft their assigned sections in parallel, gathering additional evidence in separate contexts as their analyses develop. Once the sections are assembled, global review guides targeted local revisions, reducing reliance on repeated full-report rewriting. This workflow also supports the construction of research tasks and trajectories for the mid-training and post-training of LongCat's general-purpose models. LongCat-DeepResearch achieves 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics. On an in-house benchmark, it scores 76.04, ranking second among four compared systems. Development-set analyses show benefits from combining planning perspectives, while further planning refinement has mixed effects. Additional editing improves average automatic readability preference across two benchmarks, with different trends on each.
The Brief
2:08A short podcast summary of the paper. Written and voiced by AI.
Transcript
LongCat-DeepResearch Technical Report. By Meituan LongCat Team and thirty others, from Meituan. Published on the twenty-eighth of September. Its code repository is public and real: about twenty-five source files.
A deep research system has to produce a comprehensive report grounded in evidence. That means planning the whole report, investigating many parts of it in detail, and fixing what is weak once the parts come together. One way to repair problems is to rewrite the full report again and again, and reducing reliance on that habit is part of what this system is built for.
LongCat-DeepResearch pairs an enhanced LongCat model with a multi-agent workflow. The workflow keeps global planning separate from detailed investigation, and it coordinates revision at the level of individual sections.
It runs in three movements. First, several planning agents explore external sources and refine an actionable research plan, which the authors call ResearchSpec. Second, research agents take their assigned sections and work on them in parallel, each gathering more evidence in its own context as its analysis develops, then drafting. Third, once the sections are assembled, a global review directs targeted local revisions, which reduces the reliance on repeated full-report rewrites. The same workflow is also used to build research tasks and trajectories for the mid-training and post-training of LongCat's general-purpose models.
The system scores fifty-five point two five on DeepResearchBench, fifty-one point three five on DeepResearchBench II, and seventy-nine point eight three on ResearchRubrics. On an in-house benchmark it scores seventy-six point zero four, ranking second of four compared systems.
Development-set analyses add texture. Combining planning perspectives helps, while refining the plan further has mixed effects. Additional editing raises average automatic readability preference across two benchmarks, with a different trend on each.
This is a report that keeps its own caveats in view: a second place on the in-house comparison, mixed returns from extra planning. The structural bet is clear all the same. Plan globally, research locally and in parallel, and revise sections rather than the whole, with the workflow doubling as a source of training data for the models underneath it.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A long report, built from evidence nobody can list in advance. Where do you even start?
You are building an agent that answers open-ended questions with long, evidence-grounded reports. What a report needs only becomes clear as the investigation goes on.
The road ahead
- 1 · The problem ~7 min A long research report needs evidence nobody can list in advance, and one ever-growing draft chokes on that.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Iterate early on a compact research plan, research sections in parallel separate contexts, then fix the assembled draft with targeted local revisions.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Several planners explore, a plan is refined, researchers write sections in parallel, sections are assembled, then editors revise locally under a global review.
- 6 · The experiment ~6 min You’ll guess: On the authors' in-house benchmark, against four compared research systems in total, where did LongCat-DeepResearch rank?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.