MindDR 1.5 - A multi-agent deep research framework launched by Li Auto

What is MindDR 1.5?

MindDR 1.5 is a multi-agent deep research framework developed by the Li Auto Information Agent Team. With only about 30 bytes of parameters, it achieved a score of 52.54 on the DeepResearch Bench, reaching an industry-leading level. The framework adopts a three-agent collaborative architecture of Planning, DeepSearch, and Report, coupled with a four-stage training pipeline of SFT cold start, Search-RL, Report-RL, and preference alignment. It skips expensive mid-training, reducing training tokens by 71.4% and time consumption by 60%, achieving low-cost, high-performance deep research capabilities. It is now deployed in Li Auto's online product.

Main features of MindDR 1.5

• Intelligent task planningThe Planning Agent automatically breaks down user queries into independent subtasks.
• Deep Parallel SearchThe DeepSearch Agent performs multiple rounds of search, verification, and long-range inference.
• High-quality report generationReport Agent integrates evidence from multiple sources and outputs structured long reports.
• Memory sharing and tracingIt enables cross-agent information flow and traceability through Extended Chain-of-Thought (XoT) and Tool Memory.
• Multi-tool environment callSupports unified tool interfaces for Web, Database, Browser, Python, etc.

Technical Principles of MindDR 1.5

• Multi-agent collaborative architectureThe entire deep research process is broken down into three clearly defined intelligent agents: Planning Agent, DeepSearch Agent, and Report Agent. The XoT inference trajectory and tool call records are shared through the Memory module, avoiding long context expansion of a single model and interference from capability coupling.
• Four-stage training pipeline:
  • SFT Cold StartEstablish the basic behaviors for tool calls, format adherence, and multi-round reasoning.
  • Search-RLBased on real-world tool environments, online reinforcement learning is implemented, employing dynamic reward scheduling (tool invocation → format → PRM → ORM progression) to optimize long-link search and decision-making efficiency.
  • Report-RLThe core rewards are RACE Rubrics (comprehensiveness, insight, readability, and instructional compliance), combined with citation and formatting rewards, to optimize the quality of long report generation.
• Preference AlignmentBy using DPO and Self-SFT, we can solve fine-grained user experience issues such as temporal consistency and table formatting, and align with human preferences.
• Data SynthesisBased on Baidu Encyclopedia and English Wikipedia, a knowledge graph is constructed. Through subgraph sampling, multi-hop QA generation, condition obfuscation and quality filtering, high-quality complex reasoning training data is synthesized. This data is then mixed with real user queries to bridge the distribution gap.

Key information and usage requirements of MindDR 1.5

• Development TeamThe Li Auto Information Intelligence Agent is independently developed and maintained by the Li Auto team.
• Open source statusThe technical report and paper have been published (arXiv 2604.14518), but the model weights and product interface have not yet been open-sourced.
• Parameter sizeIt employs a dual-path approach using a dense model with approximately 30B parameters and a MoE model, achieving leading performance in open-source systems of similar scale.
• Training costsCompared to MindDR 1.0, the number of training tokens is reduced by 71.4%, and the training time is reduced by 60%, achieving a significant efficiency improvement.
• Product FormIt has been deployed as a core capability in the Livis online product, providing in-depth research services for real user scenarios.
• Evaluation CriteriaIt covers multiple deep search and research report evaluation systems, including DeepResearch Bench, MindDR Bench, BrowseComp, xbench-DS, and WideSearch.

The core advantages of MindDR 1.5

• Low cost and high performanceSkip mid-training with a 30B-level small model and achieve industry-leading performance through multi-stage training, significantly reducing training and inference costs.
• Intelligent agents have a clear division of laborThe three-agent decouples search and writing, avoiding the problems of reward sparsity and capability interference in end-to-end training, and supports parallel subtask execution.
• Dynamic progressive rewardsSearch-RL employs threshold-triggered dynamic reward scheduling, gradually transitioning from basic tool calls to deep reasoning optimization, simulating an "insightful" leap in capabilities.
• Real-world alignmentMindDR Bench is built based on 500 real user queries, with evaluation dimensions covering content quality and presentation format, directly aligning with user experience.
• Efficient search decisionAt the same level of accuracy, the number of tool calls and context token consumption are significantly lower than those of competing products of the same scale, achieving both accuracy and efficiency.

MindDR 1.5 project address

• HuggingFace model libraryhttps://huggingface.co/papers/2604.14518
• arXiv technical paper: https://arxiv.org/pdf/2604.14518

MindDR 1.5 Comparison with Similar Products

Application Scenarios of MindDR 1.5

• In-depth research on the automotive industryAnalyze complex business issues such as market competition patterns, price war strategies, and technological routes.
• Academic research supportIt automatically retrieves literature, integrates multi-source evidence, and generates citation-standardized research reviews.
• Financial Investment Research: Conduct multiple rounds of information verification and structured report output for listed companies and industry trends.
• Product Decision SupportBased on massive amounts of publicly available data and user behavior logs, generate actionable business analysis reports.
• Smart Cockpit Q&AAs the core competency of Ideal Classmate, it provides car owners with highly credible and in-depth knowledge Q&A.