Meituan's LongCat-DeepResearch: plan-parallel-revise workflow scores 55.25 on DeepResearchBench
_reachsumit · x · 2026-09-30
Meituan's LongCat team released the LongCat-DeepResearch technical report (arXiv:2609.36071), a deep research system pairing an enhanced LongCat model with a multi-agent workflow: planning agents refine a ResearchSpec from external sources, research agents draft sections in parallel in separate contexts, and global review drives targeted local revisions instead of full-report rewrites. The workflow also generates training trajectories for LongCat's mid/post-training. Scores: 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, 79.83 on ResearchRubrics, 76.04 on an in-house benchmark (2nd of 4 systems). Ablations show combining planning perspectives helps; further planning refinement has mixed effects.
Related event: Meituan LongCat Unveils DeepResearch Multi-Agent Report System(2 posts)→
More from coding & agent
- New Yorker-style illustration Skill hits 400 stars, monetizes via Baidu agent ecosystem — oran_ge · 2026-09-30
- Dev builds a Grok-powered 'X newsroom' that saves 3 hours a day on posting — jamestagg · 2026-09-30
- Dimillian hails new Codex cloud environments as fantastic, teases deep dive — Dimillian · 2026-09-30
- Agent kept claiming 'CRM updated' when it wasn't: a pragmatic external-verification fix — Kindly_Ganache9027 · 2026-09-30
- Omnigent v0.16.0 ships copy-on-write sandbox edits, unified workspace browser for AI agents — matei_zaharia · 2026-09-30
- Midas Touch code dataset questioned: no baselines, single seed, possible repo overlap — maier_ak · 2026-09-30