RESEARCH · RESEARCH · #1246
LongCat-DeepResearch technical report: enhanced LongCat and multi-agent workflow for evidence-grounded reports
The arXiv technical report introduces LongCat-DeepResearch, a system that pairs an enhanced LongCat model with a multi-agent workflow that separates global planning (producing a ResearchSpec) from parallel section-level investigation and drafting, followed by targeted global review to guide local revisions. The authors report benchmark results (DeepResearchBench 55.25, DeepResearchBench II 51.35, ResearchRubrics 79.83) and an in-house score of 76.04 (second of four systems), and note that combining planning perspectives helps while additional planning refinement has mixed effects; post-drafting edits improved automatic readability preferences on some benchmarks.
KEY POINTS
- The arXiv technical report introduces LongCat-DeepResearch, a system that pairs an enhanced LongCat model with a multi-agent workflow that separates global planning (producing a ResearchSpec) from parallel section-level investigation and drafting, followed by targeted global review to guide local revisions.
- The authors report benchmark results (DeepResearchBench 55.25, DeepResearchBench II 51.35, ResearchRubrics 79.83) and an in-house score of 76.04 (second of four systems), and note that combining planning perspectives helps while additional planning refinement has mixed effects; post-drafting edits improved automatic readability preferences on some benchmarks.
- The paper proposes a structured multi-agent workflow and model enhancements aimed at more scalable, evidence-grounded research report generation and provides benchmarked performance to evaluate trade-offs in planning and editing stages.
WHY IT MATTERS
The paper proposes a structured multi-agent workflow and model enhancements aimed at more scalable, evidence-grounded research report generation and provides benchmarked performance to evaluate trade-offs in planning and editing stages.