LongCat-DeepResearch Technical Report
Meituan LongCat Team, He Zhu, Yue Xu, Wanli Wu, Haolin Ren, Yuxin Bian, Jiarui Zhao, Rongzhi Zhang, Quanchi Weng, Jinghao Cui, Yu Fan, Yuhan Liu, Yunhu Ye, Jiyuan Ren, Fengcheng Yuan, Zhao Yang, Jiacheng Zhang, Yuchuan Dai, Ruixuan Xiao, Haozhe Sun, Xiangyuan Liu, Cheng Sun, Yao Du, Yiming Hao, Hongbo Guo, Shuo He, Lei Wang, Xunliang Cai, Yan Chen, Fan Yang, Lingchuan Liu
cs.AI
2026-09-29
Meituan pairs an upgraded LongCat with a ResearchSpec multi-agent pipeline, scoring 55.25/51.35/79.83 first on three public benches and 76.04 in-house, 0.55 behind ChatGPT.
Open-ended deep research asks an agent to investigate a question that is still underspecified, then write a long, cited report. The evidence the report will need cannot be listed in full before the work starts. As the analysis moves, gaps appear, claims need checking, and new questions show up.
Current systems either interleave reasoning, retrieval, and drafting, or let a growing draft steer the next search and rewrite. Evidence, intermediate reasoning, and prose then compete for one finite context, and compression can drop material whose use is not yet obvious. Early findings can lock the narrative and shrink later search. Rewriting the whole report regenerates passages that new evidence never touched.
The design question is how to share one research agenda across agents, keep the detailed investigation each section needs, and avoid stuffing the entire inquiry into an ever-growing draft.
LongCat-DeepResearch couples an upgraded LongCat model with a research harness. Early iteration happens on an executable specification called ResearchSpec, not on the full report.
Global coordination sits in a short spec. Detailed retrieval sits in independent long contexts. The spec is revised before dispatch. Research extends only inside the assigned section. Editing coordinates text and does not reopen the global agenda.
The same stage interfaces also build training data. Questions and task-level rubrics are grounded in independently licensed review articles or frozen multi-source briefs. Teacher runs of the harness produce trajectories. That data, mixed with other corpora, goes into mid-training and post-training of LongCat's general-purpose models. System-level scores do not isolate this pipeline, any one data source, or any one training stage.
Comparisons use three official Deep Research products: Gemini-DeepResearch (Gemini 3.7 Flash), ChatGPT-DeepResearch (GPT-5.6 Sol, xhigh), and Claude-DeepResearch (Claude Opus 5, xhigh). Judges follow each benchmark: GPT-5.5 medium for DeepResearchBench and DeepResearchBench II, Gemini 2.5 Pro for ResearchRubrics.
| System | DeepResearchBench | DeepResearchBench II | ResearchRubrics | In-house |
| Gemini-DeepResearch | 50.21 | 46.72 | 64.92 | 42.49 |
| ChatGPT-DeepResearch | 54.95 | 47.16 | 74.21 | 76.59 |
| Claude-DeepResearch | 53.43 | 48.18 | 72.91 | 61.42 |
| LongCat-DeepResearch | 55.25 | 51.35 | 79.83 | 76.04 |
On the three public benches LongCat is first, at +0.30, +3.17, and +5.62 versus the strongest listed product on each. In-house it scores 76.04, 0.55 behind ChatGPT. Analysis is 70.74 versus ChatGPT's 66.58; presentation is 52.99 versus 74.64. These are point estimates without significance tests.
The public lead comes from coverage and analysis, not from readability. DeepResearchBench readability is 49.96, below ChatGPT at 51.51 and Claude at 51.35. On DeepResearchBench II, analysis is 60.69 against ChatGPT's 52.26; presentation is 83.13 against ChatGPT's 86.98. On ResearchRubrics, explicit requirements, implicit requirements, and synthesis lead; instruction following, citation quality, and communication each trail at least one comparator.
The harness move is larger than the model upgrade. Previous LongCat (LongCat-2.0) with ReAct then a direct write averages 47.04 across the three public benches. The same previous model with the current harness reaches 58.05. The current model on that harness reaches 62.14. On development subsets, the full pipeline averages 63.91. A single-Writer plan that also drops the Judge and Critic falls to 59.56. One whole-report Researcher scores 62.07. Dropping the Editor leaves 63.30. Merging three Writers through the Judge lifts the ResearchRubrics total from 81.67 to 85.13. Extra Critic/Reviser rounds can raise planned coverage while the report total slips to 84.34 and 84.89. Three enhanced Editor rounds move automatic readability preference from 50.00 to 53.96 and the two-benchmark mean from 70.89 to 72.19. One enhanced round once scored below the unedited draft on native metrics.
In the DeepResearchBench II 012 case, four user themes become 19 subsections and the Critic inserts Deep Portfolio Theory. After independent drafting, Fama-French detail is assigned to S1.5, with only a cross-reference left in S1.2. The finished report still misses Treynor and CPPI and leaves an internal section ID in the prose. An executable agenda plus a traceable edit does not guarantee complete coverage.
This is a product technical report, not a new learning algorithm. For people building deep-research agents, the transferable piece is the interface: freeze a shared spec, keep evidence in section contexts, edit locally. Compared with rewriting a growing draft, passages that new evidence never touched are not regenerated.
Putting the current harness on the previous LongCat model already moves the three-bench mean from 47.04 to 58.05. Orchestration is doing real work, independent of a stronger base model.
Comparators are full products with their own tools and budgets, not matched model ablations. The DeepResearchBench margin is 0.30. The in-house bench still loses to ChatGPT. Readability and presentation are explicit weak spots. The open questions are whether a frozen ResearchSpec can be reopened after drafting, and whether extra Editor rounds spend retrieval budget on style.
ResearchSpec is frozen after planning; the core pipeline cannot reopen the global agenda after drafting. Reports can be long, dense, and repetitive. Both editor types read the full draft and still hit context limits. The enhanced Editor changes source access, instructions, and round count together, so their separate effects are unknown.
Native scores, planned coverage, and citation support measure different things. Unchanged text can receive different rubric judgments from the automatic evaluators. Component and planning studies use previously inspected development subsets, some selected using prior model scores. There is no matched training ablation, so system scores do not separate data sources or training stages. The paper does not report DeepResearchBench FACT citation-support numbers. Point estimates have no significance tests.
The case study scores only the final report, so it cannot isolate planning or editing gains. Reading the public-bench ranking as a clean win over ChatGPT does not survive the in-house, readability, or presentation numbers.