Mind DeepResearch Technical Report

• Deep research systems face high training and inference costs, while inefficient exploration increases token use and latency and can dilute findings in accumulated context.The paper frames achieving leading performance and excellent user experience with a small model and low-cost training and inference as the core challenge.Source: §1 · pp. 1–2
• MindDR achieves state-of-the-art deep research capabilities with ~30B-parameter models by decomposing research into specialized inference subtasks and using targeted multi-stage training.Source: §1 · p. 2
• MindDR’s Planning, DeepSearch, and Report Agents enable parallel search and context isolation, with specialized agents handling retrieval, conflict resolution, and report generation.Source: §1 · p. 2
• MindDR’s four-stage pipeline uses SFT, Search-RL, Report-RL, and preference alignment; Search-RL improves BrowseComp-ZH accuracy while reducing context and tool-call consumption versus SFT, and Report-RL improves report RACE scores.The stages target behavioral cold-start, search reasoning and efficiency, report quality and conflict resolution, and alignment with human expectations.Source: §1 · pp. 2–3
• MindDR Bench comprises 500 Chinese deep research queries from real-world Livis interactions and evaluates search process, content quality, and report format rather than relying only on RACE.Source: §1 · p. 3

2.1 Deep Research Agents

• Open agents explore end-to-end training (Tongyi [34]), interactive scaling (MiroThinker [33] [32]), uncertainty-aware navigation (WebSailor [16] [15]), dual-agent reporting (WebWeaver [18]), and specialized small models (Nanbeige4.1-3B [45]).Proprietary systems Gemini Deep Research and OpenAI Deep Research demonstrated near-human performance, but closed-source limits reproducibility; existing systems largely optimize retrieval accuracy with complex monolithic end-to-end RL objectives.Source: §2.1 · p. 3

2.2 Search Reinforcement Learning

• Knowledge-graph data methods (WebSailor [16], DeepDive [20]) face coverage limits, while agent simulation (MiroThinker [33], Cognitive Kernel-Pro [9]) is costly and lacks fine-grained per-step retrieval contributions.These limitations make it difficult to distinguish critical from peripheral search steps in training data.Source: §2.2 · pp. 3–4
• MindDR addresses search-RL credit assignment with lightweight step-level advantage estimation, avoiding critic overhead and exponential sampling while optimizing retrieval accuracy and search efficiency.Trajectory-level RL spreads uniform rewards across steps; PPO-based step rewards depend on costly critics, while branch-sampling approaches remain infeasible for full-coverage, long-horizon tasks.Source: §2.2 · p. 4

2.3 Report Reinforcement Learning

• MindDR targets report-RL gaps with RACE Rubrics-based reward shaping and preference alignment, addressing conflicting evidence and real-world user expectations.Prior work includes checklist-based rewards and end-to-end RL [13] [34], reflection-guided writing [41], recursive revision [46], and backward-inference trajectory construction [35]; evaluation frameworks include RACE [7], WritingBench [42], ResearchRubrics [27], and DEER [11].Source: §2.3 · p. 4

3 MindDR Framework

• MindDR pairs a three-agent inference pipeline with a four-phase training pipeline to generate research reports from natural-language queries.The inference pipeline decomposes queries, retrieves and synthesizes evidence from heterogeneous sources, and assembles structured reports; training builds the agents’ required capabilities.Source: §3 · p. 4

3.1 Inference Pipeline

• The Planning Agent decomposes queries, parallel DeepSearch Agents retrieve and integrate evidence, and the Report Agent synthesizes citation-grounded reports.The agents coordinate through shared memory and Extended Chain-of-Thought (XoT); DeepSearch uses a ReAct-style loop, while the Report Agent creates an outline before global synthesis.Source: §3.1 · pp. 4–5, Figure 2 · p. 5
• XoT and tool memory preserve reasoning across agent interactions and tool calls, giving downstream agents information provenance for grounded reports.DeepSearch and Report append to a shared reasoning context that captures inter-agent connections as well as each agent’s reasoning.Source: §3.1 · p. 5

3.2 Training Pipeline Overview

• The four-phase curriculum uses capability-specific data and dense phase-level signals to avoid sparse, noisy end-to-end rewards and follow dependencies between capabilities.SFT stabilizes format adherence before RL; report quality depends on retrieval completeness, and preference alignment presupposes functionally correct outputs.Source: §3.2 · pp. 5–6
• The phases comprise SFT cold-start, Search-RL, Report-RL, and preference alignment using DPO and Self-SFT.Search-RL uses real tool execution and scheduled rewards; Report-RL uses RACE Rubrics [7] judged by an LLM, plus rule-based citation and format rewards. Preference alignment targets report UX issues such as temporal correctness and table formatting while avoiding catastrophic forgetting.Source: §3.2 · pp. 6–7

4.1 Query Synthesis

• The four-stage query pipeline samples verifiable, necessary, structurally independent graph paths, generates implicit multi-hop QA, obfuscates clues, and filters for solvability and quality.Sampling prevents shortcuts and cross-branch associations; obfuscation preserves solvability and answer uniqueness, while filtering checks direct answerability, plausibility, retrieval support, and semantic coherence.Source: §4.1 · pp. 7–8

4.2 Query Source Diversification and Mixing

• MindDR mixes structured knowledge-graph queries with real user queries to balance controllable reasoning chains with coverage of genuine user intent, and curates real queries into MindDR Bench.Knowledge-graph queries provide complete reasoning chains and answers, while real-query mixing addresses their corpus-coverage and distribution-shift limitations.Source: §4.2 · p. 8

4.3 MindDR Bench

• MindDR Bench contains 500 expert-reviewed queries from Li Auto user logs across 16 domains and evaluates four pipeline stages alongside content-focused RACE metrics.It addresses the scarcity of authentic Chinese queries and the limited actionable feedback of aggregate-only scoring; the module-level system builds on DeepResearch Bench [7].Source: §4.3 · p. 8

4.4 SFT Data Synthesis

• SFT data comprises approximately 12K ReAct trajectories from knowledge-graph, real-world, and human-annotated sources, with controlled complexity, trajectory length, tool patterns, and context length.Source proportions are 60% knowledge-graph, 35% real-world, and 5% human-annotated; 128K-context format correctness rises from 72% to 94%. Three-tier filtering and dynamic configuration prioritize complex multi-turn samples at 60–70%.Source: §4.4 · pp. 8–9

4.5 Search-RL Data Synthesis

• Search-RL uses approximately 35K synthesized queries that preserve complete reasoning chains and provide entity-based process-reward and answer-based outcome-reward supervision.Each instance includes intermediate entities, relational paths, ground-truth answers, and a three-level difficulty label used for adaptive sampling; entity matching avoids additional LLM inference for process-reward verification.Source: §4.5 · pp. 9–10

4.6 Report-RL Data Synthesis

• Report-RL reuses SFT trajectories to create rubric-scored long-form query–report data and cheaper short-form data without retrieval context.Long-form examples retain retrieval context and reasoning chains, with query-specific RACE Rubrics generated using the DeepResearch Bench [7] template; short-form examples reuse those rubrics and synthesize reports against their criteria.Source: §4.6 · p. 10

5.1 Supervised Fine-Tuning

• SFT establishes tool-use and reasoning foundations, while a staged long-context curriculum raises 128K-context format correctness from 72% to 94%.Under-training leaves long-context format adherence too weak for RL, while excessive SFT can collapse exploration; stopping criteria balance format correctness with retained policy entropy. In practice, the criteria are typically met at 100K–150K high-quality samples.Source: §5.1 · p. 11

5.2 Search Reinforcement Learning (Search-RL)

• Search-RL uses one continuous GRPO-based process with threshold-triggered reward-weight scheduling and dynamic task difficulty to build capabilities from tool use toward long-horizon reasoning.Its rewards cover tool invocation, format compliance, intermediate-entity coverage, and final-answer correctness; stronger negative penalties for error-prone behavior accelerate error avoidance.Source: §5.2 · pp. 11, 12, 14
• For MoE Search-RL, MindDR adopts GSPO [52] because sequence-level importance ratios stabilize clipping under expert-routing shifts and help prevent search-capability degradation.GRPO and attempted Routing Replay remained unstable when sparse activation changed expert paths between sampled responses and policy updates.Source: §5.2 · p. 13

5.3 Report Reinforcement Learning

• Report-RL uses RACE Rubrics to reward reports for comprehensiveness, readability, insight, and instruction following, supplemented by citation and format rewards.The combined reward is RReport = RRACE + λcRcite + λfRformat, with auxiliary reward weights calibrated to observed citation and formatting errors.Source: §5.3 · pp. 14–16
• For dense-model Report-RL, MindDR uses DAPO [49] to address long-response gradient imbalance and limited exploration; for MoE, it uses GSPO [52] to stabilize routing-sensitive updates.DAPO combines token-level gradients, asymmetric clipping, removal of the KL constraint, and dynamic sampling filters; GSPO’s sequence-level ratios help prevent report-writing degradation under sparse activation.Source: §5.3 · p. 16

5.4 Preference Alignment

• Preference alignment targets holistic quality issues beyond rule-based RL rewards, using self-sampled outputs scored by LLM judges and rules to create high- and low-quality sets for DPO [24] and Self-SFT [1].The targeted issues include tense inconsistency, unnatural tone, abrupt transitions, and conclusions inconsistent with retrieved content; Self-SFT reinforces high-quality samples through static iteration.Source: §5.4 · pp. 16–17
• Preference-alignment training uses the model’s own samples to keep policy updates within its existing output space, preserving SFT- and RL-acquired capabilities while shifting outputs toward human-aligned quality.This avoids pulling the policy toward unexplored regions.Source: §5.4 · p. 17

6.1 Experimental Setup

• Evaluation covers deep-search benchmarks and long-form deep-research reports, with RACE dimensions and user-experience metrics including citation accuracy, table format, and temporal error.The compared MindDR variants range from RFT-only MindDR-v1.0 to fully upgraded MindDR-v1.5, including 32B and 30B-A3B base-model versions.Source: §6.1 · p. 18

6.2 Overall Performance

• MindDR-v1.5-30B-A3B leads on BrowseComp-ZH, BrowseComp, xbench-DS, and GAIA-DS, while MindDR-v1.5-32B leads on WideSearch; MindDR-v1.5 also leads on overall RACE and major report-quality dimensions.MindDR-v1.0 remains slightly stronger on citation accuracy, showing that report-quality and citation-faithfulness optimization are not identical objectives.Source: Table 3 · p. 18, Figure 8 · p. 19, §6.2 · p. 19

6.3 Detailed Analysis

• Sequence-level Report-RL methods GSPO and DAPO improve report quality and formatting while preserving more search capability than GRPO, which causes larger search degradation.The central trade-off is improving deep-research quality while retaining prior deep-search capability.Source: Table 4 · p. 19, §6.3 · p. 20
• RACE rises from 48.82 to 50.60 when short-form data is added to long-form Report-RL training, with Comprehensiveness increasing from 49.10 to 51.08 and Insight from 47.68 to 50.96.The comparison starts from an SFT checkpoint and uses DAPO with RACE rubrics as the reward.Source: §6.3 · p. 20, Table 5 · p. 20
• DPO reduces structured table and temporal errors, while Self-SFT improves expression and coherence and adds benchmark gains; together they refine reports without sacrificing Report-RL performance.DPO targets table and temporal correctness; Self-SFT uses high-quality self-sampled reports to improve consistency.Source: Table 6 · p. 20, §6.3 · p. 20
• 2.0%: MindDR-v1.5 has the lowest temporal error rate among compared systems, outperforming MindDR-v1.0 and external baselines.Temporal error is defined as a predictive statement about a past time point without explicit attribution to a source.Source: Table 8 · p. 21, §6.3 · p. 21
• MindDR-v1.5-30B-A3B matches or surpasses compared models across context limits of 16k–128k and tool-call limits of 8–64, while using about 1.03B training-related tokens and 6K GPU card-hours.MindDR 1.5 uses about 0.18B SFT tokens and 0.85B RL sampling tokens; MindDR 1.0 used about 3.6B tokens and 15K GPU card-hours in a single training run.Source: Figure 9 · p. 21, §6.3 · p. 22

7.1 Limitations

• Scaling beyond 128K context for extremely complex research tasks remains an open challenge despite 94% format correctness at 128K.The discussion proposes hierarchical memory, selective context compression, or adaptive attention to keep agents focused on relevant evidence over long horizons.Source: §7.1 · p. 22
• Automated RACE and factual-accuracy metrics do not readily capture methodological soundness, argument novelty, or appropriate hedging of uncertain claims.The discussion identifies more nuanced human evaluation protocols as a needed direction.Source: §7.1 · p. 22

7.2 Conclusion

• MindDR is presented as a cost-effective, open framework that avoids computationally expensive continual pre-training and enables ~30B-parameter models to match or surpass the deep-research capabilities of much larger foundation models.Its targeted optimization is intended to deliver top-tier performance and user experience without prohibitive training and inference costs.Source: §7.2 · p. 22
• MindDR coordinates Planning, DeepSearch, and Report Agents via Extended Chain-of-Thought, then trains through SFT, Search-RL, Report-RL, and preference alignment.Search-RL targets search efficiency and redundant token use; Report-RL resolves information conflicts for long-form generation; preference alignment addresses residual formatting and temporal defects.Source: §7.2 · p. 22
• MindDR-v1.5-30B-A3B achieves the best results among open-source agent-style systems on BrowseComp-ZH, BrowseComp, xbench-DS, and GAIA-DS, maintaining its advantage under varied context-length and tool-call constraints.The supplied passage ends mid-sentence after “Furthermore, we”; no additional result is included.Source: §7.2 · p. 22
• 51.8 RACE is MindDR’s state-of-the-art score on MindDR Bench, where it leads in comprehensiveness, insight, instruction following, and readability.MindDR Bench uses a multidimensional evaluation system to assess content quality and presentation format beyond a single metric.Source: §7.2 · p. 23

8.1 RACE Rubrics Example

• RACE rubrics define four weighted dimensions—comprehensiveness, insight, instruction following, and readability—with fine-grained criteria that provide per-dimension reward signals for Report-RL.The example includes criteria on coverage, reform analysis, risks, subject focus, and readability.Source: §8.1 · p. 23

8.2 Short-form Data Synthesis Prompts

• Report-generation prompts require comprehensive reports with sectional conclusions, a final summary, and tables for factual or statistical content; reference-report prompts additionally supply evaluation criteria.The reference prompt conditions report generation on the query and rubric criteria.Source: §8.2 · p. 24

8.3 Scoring Model Prompt

• The scoring prompt compares generated and reference articles criterion by criterion across four dimensions, assigning each a 0–10 score with analysis and justifications.Score bands range from very poor (0–2) to excellent (8–10).Source: §8.3 · p. 24

8.4 Temporal Tense Error Detection

• Temporal-tense detection flags future-tense claims tied to dates before the current date, then checks nearby context for an explicit forecasting institution; both a past prediction date and no identified institution are required for an error.Regex extraction finds candidate sentences, and an LLM verifies the surrounding one or two sentences for an institution.Source: §8.4 · p. 25