LongCat-DeepResearch Technical Report: A Multi-Agent System for Evidence-Grounded Research

Summary

The report presents LongCat-DeepResearch, a deep research system that combines an enhanced LongCat model with a multi-agent workflow for producing comprehensive, evidence-grounded reports. Its process separates global planning from detailed investigation and coordinates revision at the section level. Planning agents explore external sources and refine an actionable ResearchSpec, after which research agents investigate and draft assigned sections in parallel, using separate contexts to gather evidence. A global review then directs targeted local revisions instead of repeatedly rewriting the full report. The workflow also supports building research tasks and trajectories for the mid-training and post-training of LongCat’s general-purpose models. LongCat-DeepResearch scores 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics. On an in-house benchmark, it scores 76.04 and ranks second among four compared systems. Development analyses find benefits from combining planning perspectives, while additional planning refinement has mixed effects. Further editing improves average automatic readability preference across two benchmarks, although the trends differ between them.