LongCat-DeepResearch Technical Report
Meituan’s LongCat team is describing a research-report system built around separate planning, parallel section work, and targeted revision.
LongCat-DeepResearch pairs an enhanced LongCat model with a multi-agent workflow meant to produce evidence-grounded reports. Planning agents explore sources and create a ResearchSpec, while research agents draft assigned sections in parallel with their own evidence-gathering contexts. The assembled report is then reviewed globally, with revisions handled locally by section instead of through repeated full-report rewrites. The paper reports scores of 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics. HF Daily Papers' note
LongCat-DeepResearch pairs an enhanced LongCat model with a multi-agent workflow meant to produce evidence-grounded reports. Planning agents explore sources and create a ResearchSpec, while research agents draft assigned sections in parallel with their own evidence-gathering contexts. The assembled report is then reviewed globally, with revisions handled locally by section instead of through repeated full-report rewrites. The paper reports scores of 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics. HF Daily Papers' note
score 5