We introduce LongCat-DeepResearch, a deep‑research system that integrates an enhanced LongCat model with a multi‑agent workflow. The workflow separates global planning from detailed investigation and coordinates revisions at the section level.
First, several planning agents explore external sources and produce an actionable research plan called ResearchSpec. Then, research agents work in parallel on their assigned sections, gathering additional evidence in separate contexts as their analyses evolve and drafting the text. After the sections are assembled, a global review directs targeted local edits, reducing the need for repeated full‑report rewrites.
The workflow also supports constructing research tasks and trajectories for mid‑training and post‑training of LongCat’s general‑purpose models. Empirically, LongCat-DeepResearch scores 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics; on an internal benchmark it reaches 76.04, ranking second among four systems. Development‑set analysis shows benefits from combining planning perspectives, while further planning refinement yields mixed results. Additional editing improves average automatic readability preference on two benchmarks, though the trends differ.
Review