Mind DeepResearch: 30B-Parameter Multi-Agent Framework Achieves Leading Performance
Li Auto has unveiled its innovative Mind DeepResearch (MindDR) framework, designed for advanced deep research and featuring approximately 30 billion parameters. The framework's notable element is its three-agent configuration, which consists of Planning, DeepSearch, and Report. To enhance its functionality, MindDR employs a four-phase training regimen, including SFT cold-start, Search-RL, Report-RL, and preference alignment. The model has recorded impressive evaluation scores, such as 45.7% on BrowseComp-ZH and 75.0% on xbench-DS. Additionally, MindDR has been launched as an online service, alongside the introduction of MindDR Bench, a benchmark with 50 distinct tasks for assessment.
Key facts
- MindDR uses ~30B-parameter models.
- Architecture includes Planning, DeepSearch, and Report agents.
- Training pipeline has four stages: SFT cold-start, Search-RL, Report-RL, preference alignment.
- Scores: 45.7% on BrowseComp-ZH, 42.8% on BrowseComp, 46.5% on WideSearch, 75.0% on xbench-DS, 52.5 on DeepResearch Bench.
- Deployed as an online product in Li Auto.
- Introduces MindDR Bench, a benchmark of 50 tasks.
- Outperforms comparable-scale open-source agent systems.
- Rivals larger-scale models.
Entities
Institutions
- Li Auto