Deep research, without the single-context bottleneck

A personal author post about LongCat-DeepResearch, not an official team announcement.

Ask a deep research system a real question and watch what happens inside it. The question splits into sub-questions. One source contradicts another. A promising lead turns out to be a press release repeating the same unsourced number. By the time the system starts writing, its context holds search results, abandoned branches, half-finished paragraphs, and the original request, all competing for the same window.

That is the problem LongCat-DeepResearch is built around. In a new technical report, my co-authors and I describe a system that separates global coordination from local investigation. The claim is deliberately modest: on three public benchmarks, treating each compared product as a complete system with its native tools and budgets, it scores highest among the four systems compared — but the comparison does not isolate model quality, and several development studies are not independent confirmation. I'm one of the authors; this is a personal explanation of the work.

The context problem is an architecture problem

Most agent designs keep one expanding history. Evidence, intermediate reasoning, and draft text accumulate together. Compressing that history early is risky: you may discard a detail before you know whether it matters. Keeping everything is expensive and eventually impossible.

LongCat-DeepResearch instead makes the shared state small and inspectable. Planners briefly search and read before committing to a plan. Several planning writers propose candidate specifications; a judge merges their perspectives; a critic looks for missing questions and cases; a reviser produces the final ResearchSpec. That document records scope, research questions, required entities or cases, provisional source leads, and who owns which section. The detailed evidence stays with the researcher that will use it.

Figure 1. Public-benchmark scores. These compare deployed systems with their native tools and budgets, not isolated model components. Source: LongCat-DeepResearch project page.

Each researcher receives the original query, the complete ResearchSpec, and exactly one assignment. It gathers evidence in its own context and writes a cited section from that history. The harness then assembles the completed sections mechanically, without asking a final writer to regenerate the report. A global editor identifies conflicts, repetition, and ownership problems; local editors revise only their assigned sections. The original section artifacts remain available for inspection.

View Figure 2: the LongCat-DeepResearch harness

Figure 2. The research harness. Planning produces an executable ResearchSpec; independent researchers write citation-bearing sections; assembly is followed by global-to-local editing. Source: project page, figure from the technical report.

What the evaluation does and does not show

On DeepResearchBench, DeepResearchBench II, and ResearchRubrics, the reported scores are 55.25, 51.35, and 79.83. The strongest of the three compared products scores 54.95, 48.18, and 74.21 on the same benchmarks, so the recorded gaps are +0.30, +3.17, and +5.62 points. On a separate in-house benchmark the system scores 76.04 overall — second among the four, 0.55 points behind ChatGPT-DeepResearch, while leading on that benchmark's Content and Analysis scores.

The more interesting evidence is a small ablation. Holding the previous LongCat model fixed and swapping in the current harness moves the three scores from 47.65 / 33.13 / 60.33 to 53.58 / 48.68 / 71.90. Holding the harness fixed and using the current model moves them to 55.25 / 51.35 / 79.83. Model capability and orchestration look complementary rather than substitutable.

Component studies point the same way, with caveats. On a fixed ResearchRubrics development subset, integrating three planning writers through the judge raises the final total from 81.67 to 85.13 versus a single-writer plan. Replacing parallel section researchers with one whole-report researcher lowers both displayed benchmark means. But more planning refinement is not automatically better: critic/reviser cycles raise planned coverage while report scores move differently. The ResearchSpec is inspectable state, not a calibrated selector for the best report.

Scope, limits, and one traceable case

The public numbers compare complete products, so they cannot attribute the result to one model component or training stage. The readability study is model-judged, not a human rating. Component and editor analyses use previously inspected development subsets, so their trends are not independent full-benchmark confirmation. And in the current pipeline the ResearchSpec is fixed after dispatch; reopening the global agenda mid-research remains future work.

A recorded portfolio-management case shows both the value and the limit of traceability. The merged ResearchSpec assigned 19 subsections across four themes. The critic noticed that Deep Portfolio Theory was missing, and the reviser added a concrete question to one section. Later, the global editor assigned detailed Fama-French coverage to a single section, and a local editor kept a concise finding plus a cross-reference. Every step is inspectable — yet the final report still retained an internal section identifier and missed some benchmark requirements. Traceability helps diagnose gaps; it does not guarantee coverage.

That is the honest summary: a practical architecture for keeping research branches independent without losing a shared agenda, evidence that the harness and model contribute separately, and clear boundaries on what the evaluation can support.

Links and disclosure

• Technical report: arXiv:2609.36071
• Code (harness only): github.com/meituan-longcat/LongCat-DeepResearch
• Project page: meituan-longcat.github.io/LongCat-DeepResearch

Disclosure: I am an author of the technical report. This post was prepared and checked using AI tools, based on the technical report; it is a personal post, not an official Meituan LongCat team announcement. Only the research harness is released — the model weights, training data, hidden rubrics, trajectories, and training code are not open-sourced.