Source-linked AI summary
MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification
MiroMind Team, S. Bai, L. Bing, L. Lei, R. Li, X. Li, X. Lin, E. Min, L. Su, B. Wang, L. Wang, L. Wang, S. Wang, X. Wang, Y. Zhang, Z. Zhang, G. Chen, L. Chen, Z. Cheng, Y. Deng, Z. Huang, D. Ng, J. Ni, Q. Ren, X. Tang, B. L. Wang, H. Wang, N. Wang, C. Wei, Q. Wu, J. Xia, Y. Xiao, H. Xu, X. Xu, C. Xue, Z. Yang, Z. Yang, F. Ye, H. Ye, J. Yu, C. Zhang, W. Zhang, H. Zhao, P. Zhu
TL;DR
Deep research requires reliable multi-step evidence acquisition, reasoning, and verification beyond single-pass language-model responses. The paper introduces MiroThinker-1.7 with agentic mid-training and MiroThinker-H1 with local and global verification; H1 reports state-of-the-art deep-research performance while the released models improve efficiency. The evaluation also spans scientific and financial reasoning, and the system blocks identified benchmark-leakage sources during evaluation.
Problem
Complex research tasks require long reasoning chains, iterative information gathering, and intermediate verification, while longer trajectories alone can accumulate noise and errors.
Method
MiroThinker-1.7 strengthens step-level planning, reasoning, tool use, and summarization through agentic mid-training, while MiroThinker-H1 adds local and global trajectory verification.
Results
MiroThinker-H1 achieves state-of-the-art results across deep-research benchmarks, including 88.2 on BrowseComp, 84.4 on BrowseComp-ZH, and 88.5 on GAIA.
Takeaways & Limitations
The results support effective interaction scaling through stronger step quality and verification rather than simply lengthening reasoning trajectories.
Takeaways & Limitations
Evaluation blocks identified benchmark-leakage sources, including publicly hosted benchmark questions and answers, through an infrastructure-level blocklist.
Abstract
from arXiv · showhide
We present MiroThinker-1.7, a new research agent designed for complex long-horizon reasoning tasks. Building on this foundation, we further introduce MiroThinker-H1, which extends the agent with heavy-duty reasoning capabilities for more reliable multi-step problem solving. In particular, MiroThinker-1.7 improves the reliability of each interaction step through an agentic mid-training stage that emphasizes structured planning, contextual reasoning, and tool interaction. This enables more effective multi-step interaction and sustained reasoning across complex tasks. MiroThinker-H1 further incorporates verification directly into the reasoning process at both local and global levels. Intermediate reasoning decisions can be evaluated and refined during inference, while the overall reasoning trajectory is audited to ensure that final answers are supported by coherent chains of evidence. Across benchmarks covering open-web research, scientific reasoning, and financial analysis, MiroThinker-H1 achieves state-of-the-art performance on deep research tasks while maintaining strong results on specialized domains. We also release MiroThinker-1.7 and MiroThinker-1.7-mini as open-source models, providing competitive research-agent capabilities with significantly improved efficiency.
1. Introduction
MiroThinker-1.7 targets long-horizon tasks by improving step-level agentic capabilities, while MiroThinker-H1 adds local and global verification for more reliable reasoning. Across open-web, scientific, and financial benchmarks, H1 reports strong or state-of-the-art performance.
- Long-horizon research tasks require iterative information gathering, multi-step reasoning, and verification of intermediate conclusions.
- MiroThinker-1.7 improves effective interaction scaling through stronger planning, reasoning, and tool-use capabilities at each step.Its agentic mid-training emphasizes task decomposition, structured reasoning, contextual reasoning, and tool interaction.
- MiroThinker-H1 evaluates and refines intermediate decisions locally while auditing complete reasoning trajectories globally.The system can reconsider actions, correct potential errors early, and compare candidate solution paths for coherent evidence.
- 88.2 on BrowseComp and 84.4 on BrowseComp-ZH are reported for MiroThinker-H1, outperforming leading open-source and commercial research agents.The evaluation spans open-web research, scientific reasoning, and financial analysis, with additional strong results on specialized benchmarks.
2. Related Works
Agentic LLMs extend language models from single responses toward persistent multi-step workflows with tool use and environmental feedback. Deep research agents apply these capabilities to open-ended knowledge synthesis by retrieving, refining, and combining external evidence.
- Agentic LLMs decompose complex goals, invoke external tools, and iteratively refine decisions using environmental feedback.They maintain reasoning traces across multiple steps and dynamically coordinate tool execution.
- Recent frontier models integrate agentic capabilities through training or tightly coupled inference frameworks while supporting long-context processing and tool execution.Examples include GPT-5.4, Claude-4.6, Gemini-3.1 Pro, DeepSeek-V3.2, Qwen3.5-397B, GLM-5.0, Minimax-M2.5, Seed-2.0-Pro, and Kimi-K2.5.
- Collectively, these developments position foundation models as autonomous agents capable of executing complex workflows in real-world environments.
- Deep research agents acquire external information, refine hypotheses, and synthesize evidence from multiple sources into structured research outputs.Their tasks require long-horizon reasoning and intensive information retrieval beyond pretrained knowledge alone.
- Industrial systems deploy integrated web browsing and multi-step planning for autonomous, end-to-end research workflows.Representative systems include OpenAI Deep Research, Claude Research, Kimi-Researcher, and Grok DeepSearch.
3. Agentic Workflow
MiroThinker-1.7 uses an iterative agent–environment loop with managed context, tool execution, and robustness mechanisms for deep research. Its dual-loop design combines step-level reasoning with episode restarts, while multiple tool categories support retrieval, computation, and file transfer.
- 3. Agentic Workflow: The agent alternates reasoning, tool invocation, and observation until it gathers sufficient evidence for a final answer.The workflow contains a formal interaction loop, a modular tool interface, and implementation mechanisms for operating in external environments.
- 3.1. Formulation: MiroThinker-1.7 uses outer episode loops for trajectory-level restarts and inner step loops for reasoning, actions, and observations.Failed episodes restart from the original query, discarding degraded prior context while preserving a bounded interaction process.
- 3.1. Formulation: The context operator preserves thought-and-action history while retaining and truncating only recent observations within a sliding window.The window selects the K most recent steps, and observations outside it are masked; early steps retain all observations subject to truncation.
- 3.2. Tools: The framework organizes tools into information retrieval, code execution, and file-and-data transfer categories.Retrieval combines search with targeted scraping and evidence distillation; code runs in an isolated E2B sandbox, while transfer utilities move local and remote assets.
- 3.3. Implementation Details: Robustness mechanisms include sliding-window filtering, bounded tool-output truncation, clean episode restarts, and automatic correction of malformed tool calls.These mechanisms manage context budgets, preserve actionable evidence, and reduce failures during long-horizon trajectories.
4. High-Quality QA Construction
The paper constructs QA data through complementary corpus-based and WebHop pipelines, combining broad coverage with verified, difficulty-controlled multi-hop reasoning. WebHop expands graphs with web knowledge while enforcing solvability and limiting shortcut solutions.
- Corpus-based and WebHop Pipelines: The Corpus-based Pipeline generates high-volume QA pairs from structured knowledge graphs, while WebHop produces fewer questions with verified multi-hop structure and open-web grounding.Corpus-based data dominates early training, whereas WebHop provides calibrated depth and difficulty.
- WebHop Pipeline: WebHop uses structured reasoning graphs, web-based knowledge expansion, and hierarchical difficulty control to address implicit difficulty and leakage in corpus-based generation.These mechanisms jointly constrain reasoning depth and align generated questions with open-web conditions.
- WebHop Pipeline: Directed reasoning trees encode verifiable parent-child relationships, with tree depth controlling reasoning hops and edge-restricted facts preventing shortcut solutions.The answer entity roots each tree, and question construction permits only facts along graph edges.
- WebHop Pipeline: Hierarchical solvability verification checks that child entities narrow each parent to bounded candidates and that the root is uniquely identifiable from its first-level evidence.Verification is applied at every level of the reasoning graph to keep questions solvable but non-trivial.
- WebHop Pipeline: Adaptive leaf obfuscation replaces answer-leaking surface associations with functional descriptions and rejects descriptions that directly reveal the original entity.This expands plausible referents while reducing direct identification shortcuts.
- Difficulty-Adaptive Filtering: Post-hoc filtering assigns questions solvable by weaker agents to earlier training stages and harder questions to later reinforcement learning stages.The resulting corpus is difficulty-graded for curriculum-style training.
5. Training Pipeline
MiroThinker-1.7 uses a four-stage training pipeline that progressively builds atomic agentic skills, structured interaction behavior, preference alignment, and reinforcement-learning generalization. Its mid-training stage supervises planning, contextual reasoning, tool use, and summarization at individual interaction steps.
- Training Pipeline: MiroThinker-1.7 follows four stages: mid-training, supervised fine-tuning, preference optimization, and reinforcement learning.The stages target atomic capabilities, structured behaviors, task-aligned decisions, and exploration-based generalization, respectively.
- Agentic Mid-training: Mid-training strengthens planning, reasoning, tool use, and answer summarization through heterogeneous agentic supervision over single-turn and intermediate-step data.The data covers cold-start planning, context-conditioned reasoning, and answer aggregation under partial observations.
- Agentic Mid-training: Planning data teaches a structured plan and first tool call from the user query, while category-specific rejection criteria remove common planning failures.Failed generations are resampled up to K times, and persistent failures are discarded.
- Agentic Mid-training: Interleaved reasoning and summarization data rewrites one trajectory turn using preceding dialogue, tool calls, and outputs instead of supervising complete trajectories end-to-end.Randomized context summarization exposes the model to partially observed, dynamically evolving agent states.
- Agentic Mid-training: The unified mid-training objective predicts a single target assistant turn y_k conditioned on the preceding context C<k, with k=1 for planning and k>1 for later reasoning or summarization.General instruction-following and knowledge-intensive data are mixed in to preserve broad capabilities and mitigate catastrophic forgetting.
- Supervised Fine-tuning: Supervised fine-tuning imitates expert thought–action–observation trajectories after filtering repetition, malformed tool calls, and undesirable behaviors.Tool observations are pre-collected and supplied as context during dialogue-style training.
- Preference Optimization: Direct Preference Optimization ranks complete trajectories primarily by final-answer correctness while requiring coherent reasoning, explicit planning, and valid, well-formed outputs.DPO is combined with an auxiliary SFT loss on preferred trajectories for stability and behavioral consistency.
6. Heavy-duty Reasoning Mode
MiroThinker-H1 adds verification-centric heavy-duty reasoning to MiroThinker-1.7 through separate local and global verifiers. Local verification refines intermediate decisions, while global verification audits evidence and delays unsupported answers.
- Verification-Centric Reasoning: MiroThinker-H1 introduces Local Verifier and Global Verifier components that independently audit step-level decisions and the complete reasoning process.The verification-centric scheme is instantiated in MiroThinker-1.7 to produce the heavy-duty reasoning mode.
- Local Verification: Local verification prompts broader exploration and selective environmental feedback to counter habitual highest-probability reasoning paths on hard problems.It is intended to prevent exploration from degenerating into repeated confirmation.
- Global Verification: Global verification organizes the full evidence chain and requests resampling or completion when evidence is insufficient before selecting an answer.Under a controllable compute budget, selection favors the path backed by the most complete and reliable evidence.
7. Experiments
Experiments evaluate MiroThinker across agentic, professional-domain, and long-form research benchmarks, with additional analysis of interaction efficiency and verification. MiroThinker-H1 achieves leading results across several general and specialized tasks.
- Experimental Setup: Evaluation covers multi-step web browsing, information retrieval, and reasoning benchmarks alongside scientific, chemical, financial, and medical benchmarks.The study also evaluates long-form reports on 50 automatically generated deep research queries.
- Overall Performance: 88.2 on BrowseComp and 84.4 on BrowseComp-ZH outperform reported commercial-agent scores on the corresponding benchmarks.MiroThinker-H1 also scores 88.5 on GAIA and 61.3 on SEAL-0, setting new best results there.
- Professional-domain Performance: 79.0 on FrontierSci-Olympiad, 73.9 on FinSearchComp, and 56.5 on MedBrowseComp are the best results on three of four professional-domain benchmarks.On FrontierSci-Olympiad, MiroThinker-H1 surpasses GPT-5.2-high and Gemini-3-Pro.
- Long Report Evaluation: MiroThinker-H1 achieves the highest report quality among evaluated deep research agents while the series provides strong factual grounding.The comparison uses 50 deep research queries and evaluates report quality and factuality.
- Effective Interaction Scaling: 16.7% better performance with 43.0% fewer interaction rounds demonstrates more effective interaction scaling for MiroThinker-1.7-mini under identical 30B parameter budgets.On HLE, the model achieves 17.4% better performance with 61.6% fewer rounds.
- Verification-Centric Heavy-Duty Reasoning: +14.2 points on BrowseComp and +8.3 points on SEAL-0 are gains associated with global verification on search-intensive tasks.FrontierSci-Olympiad and HLE also improve by 7.5 and 4.8 points, respectively, indicating gains beyond search-intensive settings.
8. Conclusions
The paper introduces MiroThinker-1.7 and MiroThinker-H1 for long-horizon reasoning, emphasizing effective interaction scaling and local and global verification. Evaluations across complex benchmarks report state-of-the-art performance for MiroThinker-H1.
- Contributions: MiroThinker-1.7 and MiroThinker-H1 address long-horizon reasoning by improving planning, reasoning, and tool use rather than merely lengthening trajectories.H1 adds a verification-centric reasoning mode that audits and refines intermediate steps before a final solution.
- Contributions: MiroThinker-H1 establishes a new state-of-the-art across diverse benchmarks, outperforming leading open-source and commercial research agents.The reported evaluations include BrowseComp, FrontierScience-Olympiad, and FinSearchComp.
Contributions
The contributions section lists the paper’s authors and contributors.
- The paper lists S. Bai and additional authors among its contributors.
- The contributor list includes G. Chen and additional collaborators.