Mind DeepResearch (Li Auto MindDR): Three-Agent Collaboration + Four-Stage Training for Efficient Deep Research
In one sentence: Mind DeepResearch (MindDR, Li Auto) takes the "small but refined" route — using a model of roughly 30B scale, paired with an architecture of three agents collaborating sequentially (Planning / DeepSearch / Report), plus a four-stage training pipeline of SFT cold-start → Search-RL → Report-RL → preference alignment, to push open-source deep research to "better than open-source systems of the same scale, and able to go toe-to-toe with larger models," and it has already been deployed in Li Auto's own products; it also proposes MindDR Bench, scored with a multi-dimensional rubric over 500 real-world Chinese queries. Year proposed: 2026 (arXiv:2604.14518, 2026-04) · Institution: MindDR Team, Li Auto Inc · Scale: ~30B Prerequisite reading: Deep Research Overview · Tongyi DeepResearch · REDSearcher · Long-Horizon Web Navigation RL
Positioning: Doing Efficient Deep Research with a Small Model
MindDR's core thesis lies on the same battle line as Alibaba's Tongyi DeepResearch and Xiaohongshu's REDSearcher — rather than piling on parameters, it squeezes a ~30B model to deep-research SOTA level through "training + architecture." Its differentiator is to explicitly decompose "deep research" into three agents with clearly divided roles, and to design a dedicated RL stage to polish each part separately.
Three Agents: Planning → DeepSearch → Report
MindDR realizes the deep-research loop as three sequentially collaborating agents, with information passed down stage by stage:
- Planning Agent: decomposes the user question into subtasks and generates the corresponding search strategy — deciding "which aspects to research, and in what order to search."
- DeepSearch Agent: executes web search and information retrieval per the plan, with multi-round crawling, reading, and supplemental search — this is the part that truly "digs up information."
- Report Agent: synthesizes the collected materials into a coherent, structured report, responsible for clarity, accuracy, and comprehensiveness.
The benefit of this explicit division of labor is that each agent's capability can be trained and optimized separately — which is precisely the design premise behind the four-stage training pipeline below.
Four-Stage Training Pipeline
MindDR does not RL all the way through in one shot; instead it builds up different capabilities layer by layer in four steps:
- SFT cold-start: supervised fine-tuning on annotated trajectories for planning, search, and report-writing tasks, first establishing the base capability to "follow this workflow."
- Search-RL: reinforcement learning targeting the search stage, using search-specific rewards (coverage / relevance / efficiency type) to optimize "how to search better."
- Report-RL: reinforcement learning targeting the report stage, using report-quality rewards (clarity, accuracy, comprehensiveness) to optimize "how to write better."
- Preference alignment: finally, DPO-style preference learning to align with human preferences on "what makes good research."
This design of doing RL separately for "search" and "report" echoes its architecture of splitting the agents into a deep-search part and a report part — each part has its own reward signal, avoiding mashing two objectives into a single fuzzy aggregate score.
MindDR Bench: 500 Real-World Chinese Queries + Multi-Dimensional Rubric
To make evaluation closer to real usage, MindDR proposes its own benchmark, MindDR Bench: based on 500 real-world Chinese user queries, scored not with a single metric but with a multi-dimensional rubric (research depth, factual accuracy, report coherence, etc.). This aligns with the approach in this chapter's DR-Rubric and AgentDisCo of evaluating deep research with rubrics / report-quality benchmarks — open-ended deep research increasingly favors a "multi-dimensional scale" over "a single score."
Benchmark Performance (defer to the original paper)
The technical report reports multi-leaderboard results at ~30B scale (numbers per the arXiv original):
| Benchmark | MindDR ~30B |
|---|---|
| BrowseComp-ZH | 45.7 |
| BrowseComp | 42.8 |
| WideSearch | 46.5 |
| xbench-DeepSearch | 75.0 |
| DeepResearch Bench | 52.5 |
| MindDR Bench | 51.8 (self-reported SOTA) |
The paper claims it is better than open-source agent systems of the same scale, and able to compete with larger-scale models. As is customary in this chapter: these scores differ in measurement protocol, timing, and tool configuration, so directly comparing absolute values across systems has limited meaning — treat them as an order-of-magnitude reference for "the deep-research capability of contemporaneous ~30B-class open source."
Position in the Deep Research Lineage
- vs Tongyi DeepResearch / REDSearcher: all three work in the 30B class along the route of "specifically training long-horizon search / deep-research agents." MindDR's distinguishing feature is the four-stage pipeline of explicit three-agent division of labor + independent RL for search and report (Search-RL / Report-RL), and it has already been deployed in Li Auto's own products; REDSearcher emphasizes more "task-difficulty formalization + low-cost RL over a local closed corpus," while Tongyi emphasizes "agentic mid/post-training + fully automated data synthesis." See Tongyi and REDSearcher for comparison.
- vs AgentDisCo / open-source frameworks: AgentDisCo does not train a base model and instead does agent orchestration on a strong general model; MindDR builds the model from the training side, making the two complementary.
- For overall positioning and the backdrop of the "domestic/open-source leaderboard race," see the Deep Research Overview.
References
- MindDR Team, Li Auto Inc. Mind DeepResearch Technical Report. arXiv:2604.14518, 2026-04. https://arxiv.org/abs/2604.14518
- Evaluation benchmarks: BrowseComp / BrowseComp-ZH, WideSearch, xbench-DeepSearch, DeepResearch Bench, and the in-house MindDR Bench (500 Chinese queries, multi-dimensional rubric)