PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.36071v1cs.AI

TL;DR

Long-form deep research must coordinate an initial agenda with evidence gathering that evolves as sections develop. LongCat-DeepResearch combines an improved LongCat model with a shared ResearchSpec, independent section research, and coordinated editing, and reports benchmark scores across three public benchmarks and an in-house evaluation.

  • Problem

    Open-ended deep research must coordinate a useful initial agenda with evidence gathering that continues as explanations, unresolved claims, and new questions emerge.

    Source: §1 · p. 2

  • Method

    LongCat-DeepResearch uses planners to refine a shared ResearchSpec, independent researchers to investigate and draft assigned sections, and global-to-local editing after assembly.

    Source: §1 · p. 2

  • Results

    LongCat-DeepResearch scores 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics, and ranks second among four systems with 76.04 in-house.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The system’s stage interfaces also support constructing research questions, task-specific rubrics, and trajectories for general-purpose model training.

    Source: §1 · p. 2

    ResearchSpec is fixed during section research, so reopening the global agenda after drafting is outside the core pipeline.

    Source: Limitations · p. 13

Abstract

from arXiv · show

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.

1 Introduction

Long-form deep research must coordinate an initial agenda with evidence gathering that develops as sections progress. LongCat-DeepResearch uses a shared ResearchSpec and section-level research and editing, and reports strong benchmark results alongside support for research-task and trajectory construction.

1 Introduction

  • Open-ended research requires evidence gathering to continue as explanations, unresolved claims, and new questions emerge during report development.The section’s central question is how to coordinate a shared research agenda while preserving detailed investigation for each section.Source: §1 · p. 2
  • Growing research histories and repeated full-report updates can consume context, anchor later searches to early findings, and regenerate unaffected text.The challenge is to retain research-writing feedback without concentrating the investigation in an expanding draft.Source: §1 · p. 2
  • LongCat-DeepResearch combines an improved LongCat model with an executable ResearchSpec refined by multiple planners before section dispatch.Researchers investigate and draft assigned sections independently; after assembly, a Global Editor guides targeted Local Editor revisions without reopening the global agenda.Source: §1 · p. 2
  • 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics; the in-house score is 76.04, ranking second among four systems.The system trails ChatGPT-DeepResearch’s 76.59 by 0.55 points on the in-house benchmark.Source: §1 · p. 2
  • The same stage interfaces support construction of research questions, task-specific rubrics, and trajectories for LongCat’s general-purpose model training.The report describes these materials as used in mid-training and post-training.Source: §1 · p. 2

2 Related Work

Related work spans context management, structured research planning, long-form synthesis and revision, open research implementations, and synthetic research training data. LongCat-DeepResearch positions its design around shared planning and the preservation and revision of complete section drafts.

Context management and evidence preservation

  • Context-management approaches use external storage, folding, structured summaries, or evidence graphs, while LongCat-DeepResearch retains complete section artifacts through assembly and editing.The report distinguishes this synthesis path from separate methods for managing research history through external memory and compression.Source: §2 · p. 2

Research planning and structured state

  • Structured planning ranges from perspective development and executable graphs to mutable outlines, coverage objectives, and research DAGs.ResearchSpec instead fixes section-level execution responsibilities after planning; researchers retain findings in section artifacts for editorial reconciliation, rather than using a persistent knowledge graph or input-level query-refinement specification.Source: §2 · pp. 2–3

Long-form synthesis and revision

  • Long-form systems combine retrieval, drafting, feedback, verification, and report revision; LongCat-DeepResearch assembles independent sections before assigning global ownership and local edits.The report notes that local revision scope does not guarantee retention of every useful detail.Source: §2 · p. 3

Open research systems

  • Open implementations provide precedents for structured planning, concurrent research, evidence collection, compression, and report synthesis.LongCat-DeepResearch builds on these workflow patterns with shared planning and complete section-draft preservation and revision.Source: §2 · p. 3

Synthetic research tasks and trajectories

  • Research-training work includes rubric-based supervision, synthetic tasks and trajectories, evidence trees, and methods for aligning open-ended questions with requirements.The report distinguishes evaluation resources from synthetic training data and frames its contribution around provenance, question–rubric alignment, and trajectory interfaces.Source: §2 · p. 3

3 Method

The method coordinates detailed evidence gathering through a fixed, executable ResearchSpec and independent section work, then assembles and edits the resulting artifacts. The same stage interfaces guide evidence-grounded task and trajectory construction for model training.

3 Method

  • The harness separates global research responsibilities from detailed evidence gathering in independent section contexts, then assembles section artifacts for editing.Retaining artifacts does not mean every retrieved page or interaction detail is preserved.Source: §3 · p. 4, §3.1 · p. 4

3.1 Research Harness

  • Planners explore sources independently, then a Judge, Critic, and Reviser consolidate candidate plans and seek missing questions before assignments are dispatched.ResearchSpec records each section’s scope, questions, required entities or cases, and provisional source leads; it is validated and fixed during section research.Source: §3.1 · p. 4
  • Researchers receive the query, full ResearchSpec, and one assignment, investigate concurrently in separate contexts, and write citation-bearing sections from acquired evidence.They do not depend on previously generated sections, and carry local evidence into writing rather than first compressing it for a separate writer.Source: §3.1 · p. 5
  • After mechanical assembly, a Global Editor assigns ownership and identifies conflicts, while Local Editors revise assigned sections using the full draft and specification.Local output scopes reduce regenerated text, but both editor types read the full draft, remain subject to context limits, and may omit useful details.Source: §3.1 · p. 5
  • Stage interfaces map queries and explored sources to ResearchSpecs, assignments to grounded sections, and drafts plus directives to revised sections.These units support targeted data synthesis and rubric-based evaluation at each stage.Source: §3.1 · p. 5

3.2 Research Data Construction

  • Research tasks pair questions grounded in independent source material with rubrics whose factual criteria require evidence and analytical criteria specify warranted comparisons or inferences.Construction uses review articles or frozen multi-source briefs; the answering agent receives the query, not the separate construction evidence and rubrics.Source: §3.2 · p. 5, Figure 3 · p. 6
  • Task checks test bounded evidence availability, rubric quality, leakage, and coherence, while trajectories record stage-specific artifacts and retain complete, partial, and failed attempts.Trajectory validity is one part of dataset selection; dataset-level overlap checks, distribution-aware selection, and human review where specified are additional checks.Source: §3.2 · p. 5
  • Research-related data are used alongside other data in mid-training and post-training of LongCat’s general-purpose models.System-level scores do not isolate the contribution of this pipeline, an individual data source, or a training stage.Source: §3.2 · p. 6

4 Evaluation

Evaluation compares LongCat-DeepResearch with deployed deep-research systems on three public benchmarks and an in-house benchmark. It leads the public-benchmark overall scores, while ranking second in-house.

4 Evaluation

  • Evaluation covers DeepResearchBench, DeepResearchBench II, ResearchRubrics, and an in-house benchmark, using each benchmark’s evaluation setup and specified report evaluator.DeepResearchBench and DeepResearchBench II use GPT-5.5 (medium); ResearchRubrics uses Gemini 2.5 Pro.Source: §4 · p. 6

4.3 Evaluation Protocol

  • Generation uses the original task prompt, while benchmark rubrics and reference reports are reserved for scoring; Public-Web retrieval remains available.Scored task coverage varies across systems, which are compared as deployed with their native tools and budgets.Source: §4 · p. 6, §4.3 · p. 7

4.1 Benchmarks

  • DeepResearchBench covers 100 tasks, DeepResearchBench II covers 132, and ResearchRubrics covers 101; their scoring uses task-level aggregation and benchmark-specific criteria.DeepResearchBench RACE uses relative scores, DeepResearchBench II reports rubric pass rates, and ResearchRubrics applies weighted criteria.Source: §4.1 · pp. 6–7

4.2 Comparison Systems

  • The product comparisons use each provider’s complete Deep Research system through its official client, so system-level scores do not isolate a model or harness component.The selected client models are Gemini 3.7 Flash, GPT-5.6 Sol (xhigh), and Claude Opus 5 (xhigh).Source: §4.2 · p. 7

4.4 Main Results

  • LongCat-DeepResearch scores 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics, leading each public-benchmark comparison by 0.30, 3.17, and 5.62 points.Table 1 presents the benchmark-specific dimensions alongside these overall results.Source: §4.4 · p. 7, Table 1 · p. 8
  • Dimension results are uneven: readability trails two systems on DeepResearchBench, presentation trails ChatGPT on DeepResearchBench II, and several ResearchRubrics categories trail a comparator.LongCat leads in comprehensiveness, insight, and instruction following on DeepResearchBench; information recall and analysis on DeepResearchBench II; and explicit requirements, implicit requirements, and synthesis on ResearchRubrics.Source: §4.4 · p. 8

4.5 In-house Benchmark

  • 76.04 is LongCat-DeepResearch’s in-house overall score, ranking second behind ChatGPT-DeepResearch’s 76.59 by 0.55 points.In-house results use an automatic evaluator; Overall is the weighted total.Source: Table 1 · p. 8, §4.5 · p. 9

5 Design Analysis

The design analyses examine how planning, independent section research, and coordinated editing contribute to LongCat-DeepResearch. Development-set studies report benefits from planning aggregation and enhanced editing, while the case study traces agenda revision and cross-section coordination.

5.1 Component Ablations

  • Full outperforms simplified planning and a single whole-report Researcher, supporting compact global coordination alongside independent section investigation.The Writer → Reviser intervention also removes the Planning Judge and Critic, so it changes the planning procedure as well as Writer count.Source: §5.1 · p. 9, Table 2 · p. 9

5.2 ResearchSpec Refinement

  • Plan aggregation improves final-report quality, while further Critic/Reviser cycles improve planned coverage on the fixed ResearchRubrics development subset.The same downstream Researcher and Editor configuration is used at each stage, with planning compaction disabled.Source: §5.2 · p. 9, Table 3 · p. 10

5.3 Editor Scaling

  • Enhanced Editor rounds improve average readability preference, and the final round preserves or improves native benchmark scores relative to the unedited draft.Table 4 reports native scores and mean automatic preference against the same A0 draft; its caption cautions that the highest mean does not imply statistical significance.Source: §5.3 · p. 9, Table 4 · p. 10

5.4 Model and Harness Configurations

  • The current harness improves scores with the previous model, and the current model further improves scores under that harness.The three configurations are evaluated on full benchmarks; Table 5 reports native report quality and unweighted benchmark averages.Source: §5.4 · p. 10, Table 5 · p. 10

5.5.1 Turning the Question into Research Assignments

  • The merged ResearchSpec divides four requested themes into 19 subsections, each specifying coverage, questions, evidence targets, and source leads.A Critic flags the omission of Deep Portfolio Theory, and the Reviser adds a research question and named entity to S2.2 without creating another subsection.Source: §5.5.1 · p. 11, REVISER: MAKE THE NEW QUESTION… · p. 11

5.5.2 Coordinating Evidence Across Section Drafts

  • The CAPM section’s cited empirical evidence is assigned detailed treatment in S1.5, while S1.2 retains a brief mention and cross-reference.The Global Editor makes the ownership decision and the Local Editor applies it to the existing section text.Source: §5.5.2 · p. 12, Figure 5 · p. 12

LOCAL REVISION RETAINED IN THE FINAL REPORT

  • The revision preserves the finding that market beta alone had no explanatory power for average returns, while directing readers to S1.5 for detail.The case also records ownership decisions for error maximization and NSGA-II material, not only sentence-level edits.Source: LOCAL REVISION RETAINED IN THE FINAL… · p. 12
  • The case report retains an internal S1.5 reference without a matching displayed heading identifier and misses requirements including Treynor and CPPI.Only the final report was scored, so the case does not isolate planning or editing gains.Source: LOCAL REVISION RETAINED IN THE FINAL… · p. 12

6 Conclusion

LongCat-DeepResearch combines a model-and-harness workflow centered on ResearchSpec with independent section research and coordinated editing. It achieves strong benchmark scores, while the reported findings remain specific to evaluated configurations and do not isolate training-stage contributions.

6 Conclusion

  • The system combines parallel planning, independent section research, and coordinated editorial revision around ResearchSpec.The system shifts early iteration from full reports to the specification, which defines evidence requirements and section responsibilities.Source: §6 · p. 13
  • The system achieves 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics; on matched development subsets, the full LongCat pipeline has the highest component-study means.Source: §6 · p. 13
  • ResearchSpec coverage supports early inspection of planned requirements, but factual verification and independent report evaluation remain necessary.The reported observations are specific to evaluated configurations and do not isolate training-stage contributions.Source: §6 · p. 13

Limitations

The evaluation and planning pipeline have bounded interpretations, while development evidence and editorial practice leave reproducibility, attribution, and readability constraints.

Limitations

  • Automatic rubric judgments can differ for unchanged content, and question-bootstrap intervals capture question—not generation or judging—variation.Coverage diagnostics aid inspection of planned requirements, but factual verification and independent report evaluation remain necessary.Source: Limitations · p. 13
  • Development studies use previously inspected subsets, with variable retrieval windows and computation; broader task coverage and repeated runs would better characterize stability and resource needs.Changing web content and hosted products also limit exact replay.Source: Limitations · p. 13
  • Without matched training ablations, the contributions of individual data sources and training stages remain unidentified.Task and trajectory construction is part of a broader model-development process.Source: Limitations · p. 13
  • ResearchSpec is fixed during section research, and the workflow can omit relevant details; reports may remain lengthy, dense, or repetitive.Global and Local Editors read the full draft under input-context limits; additional refinement has varying effects and consumes resources that could support evidence gathering.Source: Limitations · p. 13

A Official Leaderboards and Evaluated Systems

The appendix places official leaderboard snapshots beside this work’s system comparisons, while distinguishing their provenance, settings, and status.

A Official Leaderboards and Evaluated Systems

B.1 Main Evaluation and Shared Conventions

The section documents evaluation conventions, component-comparison conditions and limitations, and model–harness settings. It specifies how reports are retrieved and scored, which comparisons are paired, and where compute or evaluation evidence is not equivalent.

B.1 Main Evaluation and Shared Conventions

  • For ResearchRubrics, LongCat uses thinking, three Planning Writers, up to 20 tool rounds, and research concurrency three; Claude’s overall score is an external aggregate without verifiable complete per-task grades.Scores average reports with completed grades, and scored coverage varies across systems.Source: §B.1 · p. 18
  • The web harness retrieves public information through managed search and page-reading services, while agents control when to search and which pages to read.Search returns titles, URLs, and snippets; page retrieval parses selected pages for further reading.Source: §B.1 · p. 18
  • ResearchRubrics uses Gemini 2.5 Pro with binary verdicts and signed criterion weights, while the other benchmarks use their specified native evaluators and scoring conventions.ResearchRubrics scores use the sum of positive weights as the denominator; DeepResearchBench II development studies use 50-rubric chunks and the complete rubric denominator.Source: §B.1 · p. 18

B.2 Component Comparison

  • Component comparisons use fixed development subsets and questions completed by all four arms, but the studies have one grade per report and lack established reference-result alignment.An incomplete single-Researcher case is excluded; dedicated fact-support and readability evaluations are also incomplete for these cohorts.Source: §B.2 · p. 18

B.3 Model–Harness Configuration Settings

  • The full-benchmark configuration comparison uses different response ceilings and retrieval windows, and its cumulative allowances do not imply equal realized compute.ReAct/direct reporting has a 65,536-token final-response ceiling versus 32,000 tokens per call for the current harness; same-model recovery retains validated plans and sections under unchanged scoring rules.Source: §B.3 · p. 18

B.4 ResearchSpec Coverage Evaluation

The ResearchSpec diagnostic scores rubric coverage before report generation, with explicit limits on interpreting that score as planning completeness.

B.4 ResearchSpec Coverage Evaluation

  • ResearchSpec coverage is the mean per-question percentage of positive-weight rubric criteria judged fully covered by the plan, with criteria counted equally.Partial, missing, and not-plan-evaluable criteria remain in the denominator.Source: §B.4 · p. 18
  • The diagnostic evaluates the plan before report generation; it measures neither final-report quality nor factual correctness.The evaluator maps rubric criteria to supporting plan excerpts and section identifiers.Source: §B.4 · p. 18
  • Including criteria that cannot be evaluated from a plan limits interpreting coverage as planning completeness; comparisons use the same criterion set.Rubrics and diagnostic judgments are withheld from generating agents.Source: §B.4 · p. 19

B.5 Readability Preference Evaluation

The appendix specifies automatic readability-preference scoring and a matched development-set protocol for ResearchRubrics planning refinement.

B.5 Readability Preference Evaluation

  • Readability preference compares each edited report with its own unedited draft, averages evaluator judgments across orders and guides, and maps margins to a scale where 50 is parity.Question scores are averaged within each benchmark, then the two benchmark means receive equal weight; these are automatic preferences, not absolute readability or human ratings.Source: §B.5 · p. 19

B.6 ResearchRubrics Refinement Protocol

  • Four planning snapshots—W0, J0, R1, and R2—are evaluated on the same development subset using the original parallel Researchers and Editor.W0 reuses the first Writer rather than a score-selected plan; R1 and R2 each add one Critic/Reviser cycle to the preceding plan.Source: §B.6 · p. 19
  • ResearchRubrics scores use per-question weighted criterion satisfaction; Total uses all rubric categories, not the average of the displayed explicit and implicit subscores.Coverage judgments are reused, and diagnostic judgments are not supplied to the generating agents.Source: §B.6 · p. 19

B.7 Editor Refinement: Settings and Paired Results

The appendix reports editing comparisons and traces selected changes from ResearchSpec through final reports, while noting exploratory settings and limits on attributing score changes to individual stages.

B.7 Editor Refinement: Settings and Paired Results

  • Development subsets were expanded after initial-result inspection, making the analysis exploratory; enhanced editing used source-bounded excerpts without new retrieval.Source access, fact-checking instructions, and iteration count form a joint intervention, and a higher score need not mean editing added credited information.Source: §B.7 · pp. 19–20
  • +0.158 overall, LongCat B3’s automatic readability-preference margin relative to A0, despite −0.183 on DRB-II and +0.500 on ResearchRubrics.The aggregate trend need not hold within either benchmark.Source: §B.7 · p. 19
  • +2.51 on DRB-II and +0.09 on ResearchRubrics are the paired native-score changes for LongCat B3–A0.Table 8 reports paired contrasts on fixed development subsets; positive changes favor the first condition.Source: §B.7 · p. 19, Table 8 · p. 20
  • One edit corrected an online-learning study table and raised its native score by 1.59 points, while a UBI edit lost a higher-weight citation criterion and scored −1.25.The examples show a factual correction in one report and weaker quantitative-claim attribution in another.Source: §B.7 · p. 20

FINAL EVALUATION

04 LOCAL EDIT →FINAL REPORT

  • The final report retained the Fama–French finding that beta lacked explanatory power while size and book-to-market equity were related to returns, with detail cross-referenced to S1.5.Global editing assigned ownership of the repeated finding to S1.5, and local editing shortened the S1.2 discussion to a cross-reference.Source: §04 · p. 21, §03 · p. 21, WHAT CHANGED · p. 21

REMAINING COVERAGE GAP

  • 11.43/100 was the case 02 final score; the report omitted two rubric-required studies, and information recall remained zero.The case illustrates that refining an existing agenda can leave required evidence uninvestigated; its final score does not isolate an individual stage’s contribution.Source: WHAT CHANGED · p. 22, REMAINING COVERAGE GAP · p. 22, TAKEAWAY · p. 22

TASK AND RESEARCH AGENDA

  • The case 02 report preserved a study result that the treatment group completed the task 55.8% faster than the control group: 71.17 versus 160.89 minutes on average.The reported result includes p = 0.0017 and a 95% CI of 21–89%.Source: TASK AND RESEARCH AGENDA · p. 22

Related papers

Loading 2609.36071v1…