Deep research, without the single-context bottleneck

He Zhu — Sat, 24 Oct 2026 17:45:55 +0000

Architecture notes, and what the benchmark numbers do and do not show.

Ask a deep research system a hard question. Sub-questions multiply. A source contradicts an earlier one. A promising lead turns out to be a press release repeating an unsourced figure. By the time writing starts, search results, abandoned branches, half-finished paragraphs, and the original request all compete for one window.

That pressure is the starting point for LongCat-DeepResearch. Its answer is not a larger window but an architectural split: a small shared plan coordinates the work, while each section's evidence stays in the context where it was gathered. The report records 55.25, 51.35, and 79.83 on three public benchmarks, comparing complete systems with their native tools and budgets.

The bottleneck is coordination, not retrieval

Deep research is rarely a straight line from search to draft. New evidence creates new questions, and one section can grow into its own investigation. When every branch shares one expanding history, evidence, intermediate reasoning, and prose compete for the same space. Compress early and you may discard a detail before you know whether it matters.

LongCat-DeepResearch separates global coordination from local investigation. A compact ResearchSpec records what the report must establish and who owns which part. Detailed tool interactions stay with the researcher that will use them. Completed sections, not recurrent summaries, become the artifacts passed into composition.

Planning explores before it commits

Several planning writers briefly search and read before proposing candidate specifications. A judge integrates their perspectives, a critic looks for missing questions and cases, and a reviser produces the final ResearchSpec, recording scope, research questions, required cases, source leads, and section ownership.

Each researcher receives the original query, the full ResearchSpec, and one assignment. It gathers evidence in its own context and writes a cited section from that history. The plan is shared; the raw investigation is not.

Figure 1. The harness turns a shared agenda into independent section research, then coordinates the finished text. Source: LongCat-DeepResearch repository.

Assemble mechanically, edit globally then locally

Composition is deliberately unglamorous. The harness assembles completed sections without asking a final writer to regenerate the report. A global editor identifies conflicts, repeated material, and ownership problems; local editors revise only their assigned sections against those directives. Global judgment and local generation stay separate, and the original section artifacts remain inspectable.

The split matters when a report is long. A final rewrite can smooth over a contradiction or drop a citation. Mechanical assembly preserves what each researcher wrote, while the two-level edit pass resolves overlap without restarting the document. The current pipeline fixes the plan after dispatch; reopening the agenda mid-research remains future work.

What the benchmark numbers support

On DeepResearchBench, DeepResearchBench II, and ResearchRubrics, the reported scores are 55.25, 51.35, and 79.83. The gaps to the strongest of the other three systems are +0.30, +3.17, and +5.62 points. Each product is compared as a complete system with its native tools and budgets, not as an isolated model.

On a separate in-house benchmark the system scores 76.04 overall — second among the four, 0.55 points behind ChatGPT-DeepResearch at 76.59.

A small ablation is more informative than the ranking. Holding the previous LongCat model fixed and swapping in the current harness moves the three public 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, not substitutable.

Figure 2. Scores for the four compared systems, with native tools and budgets rather than isolated model components. Source: LongCat-DeepResearch project page.

Three qualifications apply. Public scores cannot attribute the outcome 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.

What is actually released

The release is the research harness, not the whole system. Model weights, training data, hidden rubrics, trajectories, and training code are not included. The harness is inspectable; the trained model behind the numbers is not.

Technical report: arXiv:2609.36071
Harness code: github.com/meituan-longcat/LongCat-DeepResearch
Project page: meituan-longcat.github.io/LongCat-DeepResearch

The takeaway

The useful idea is not a bigger context window. It is a division of labor that keeps the agenda shared and the evidence local, then edits the result in two passes rather than regenerating it. That makes long reports easier to inspect and failures easier to locate.

AI disclosure: Fully Autonomous. This personal post was planned and written by an AI agent from the technical report and official project materials.