MindDR 1.5 - Li Auto's Multi-Agent Deep Research Framework

What is MindDR 1.5

MindDR 1.5 is a multi-agent deep research framework launched by Li Auto's Information Agent team. With only about 30B parameters, it achieved a score of 52.54 on DeepResearch Bench, reaching industry-leading performance. The framework adopts a collaborative architecture of three agents: Planning, DeepSearch, and Report, combined with a four-stage training pipeline of SFT cold start, Search-RL, Report-RL, and preference alignment. It skips costly mid-training, reduces training tokens by 71.4% and GPU hours by 60%, and has achieved low-cost, high-performance deep research capabilities. It is now deployed in the Li Xiang online product.

Main Features of MindDR 1.5

• Intelligent Task Planning: The Planning Agent automatically decomposes user queries into independent subtasks.
• Deep Parallel Retrieval: The DeepSearch Agent performs multi-round search, verification, and long-horizon reasoning.
• High-Quality Report Generation: The Report Agent integrates multi-source evidence and outputs structured long reports.
• Memory Sharing and Traceability: Enables cross-agent information flow and traceability through Extended Chain-of-Thought (XoT) and Tool Memory.
• Multi-Tool Environment Invocation: Supports unified tool interfaces such as Web, Database, Browser, and Python.

Technical Principles of MindDR 1.5

• Multi-Agent Collaborative Architecture: Splits the entire deep research process into three clearly divided agents: Planning Agent (task planning), DeepSearch Agent (deep search), and Report Agent (report writing). Through the Memory module, they share XoT reasoning traces and tool invocation records, avoiding long-context inflation and capability coupling interference in a single model.
• Four-Stage Training Pipeline：
  • SFT Cold Start: Establishes foundational behaviors for tool invocation, format adherence, and multi-round reasoning.
  • Search-RL: Conducts online reinforcement learning based on a real tool environment, using dynamic scheduling rewards (tool invocation → format → PRM → ORM progressive), optimizing long-chain search and decision-making efficiency.
  • Report-RL: Uses RACE Rubrics (comprehensiveness, insight, readability, instruction following) as core rewards, combined with citation and format rewards, to optimize long report generation quality.
• Preference Alignment: Addresses fine-grained user experience issues such as temporal consistency and table formatting through DPO and Self-SFT, aligning with human preferences.
• Data Synthesis: Builds a knowledge graph based on Baidu Baike and English Wikipedia, and synthesizes high-quality complex reasoning training data through subgraph sampling, multi-hop QA generation, conditional confusion, and quality filtering; it is also mixed with real user queries to bridge the distribution gap.

Key Information and Usage Requirements of MindDR 1.5

• Development Team: Independently developed and maintained by Li Auto's Information Agent team.
• Open-Source Status: The technical report and paper have been made public (arXiv 2604.14518), while the model weights and product interfaces are not yet open source.
• Parameter Scale: Adopts dual routes of a dense model and an MoE model with about 30B parameters, leading in performance among open-source systems of the same scale.
• Training Cost: Compared with MindDR 1.0, the number of training tokens is reduced by 71.4% and training GPU hours by 60%, achieving significant efficiency improvements.
• Product Form: Deployed as a core capability in the online product Li Xiang Tong Xue (Livis), providing deep research services for real user scenarios.
• Evaluation Benchmarks: Covers multiple deep search and research report evaluation systems, including DeepResearch Bench, MindDR Bench, BrowseComp, xbench-DS, WideSearch, and others.

Core Advantages of MindDR 1.5

• Low Cost, High Performance: Using a 30B-class small model to skip mid-training, achieving industry-leading performance through multi-stage training, significantly reducing training and inference costs.
• Clear Agent Division of Labor: Three agents decouple search and writing, avoiding the sparse reward and capability interference problems of end-to-end training, and supporting parallel subtask execution.
• Dynamic Progressive Rewards: Search-RL adopts threshold-triggered dynamic reward scheduling, gradually transitioning from basic tool invocation to deep reasoning optimization, simulating "epiphany-style" capability leaps.
• Real-Scenario Alignment: MindDR 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 Decisions: At the same accuracy, the number of tool calls and context token consumption are significantly lower than competitors of the same scale, achieving accuracy and efficiency.

MindDR 1.5 Project Addresses

• HuggingFace Model Repository：https://huggingface.co/papers/2604.14518
• arXiv Technical Paper：https://arxiv.org/pdf/2604.14518

Comparison of MindDR 1.5 with Similar Competitors

Application Scenarios of MindDR 1.5

• In-Depth Research in the Automotive Industry: Analyzing complex business issues such as market competition landscape, price war strategies, and technology roadmaps.
• Academic Research Assistance: Automatically retrieving literature, integrating multi-source evidence, and generating research reviews with standardized citations.
• Financial Investment Research: Conducting multi-round information verification and structured report output for listed companies and industry trends.
• Product Decision Support: Generating actionable business analysis reports based on massive public data and user behavior logs.
• Intelligent Cockpit Q&A: As a core capability of Li Xiang Tong Xue, providing car owners with highly credible in-depth knowledge Q&A.