Source-linked AI summary
MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling
MiroMind Team, Song Bai, Lidong Bing, Carson Chen, Guanzheng Chen, Yuntao Chen, Zhe Chen, Ziyi Chen, Jifeng Dai, Xuan Dong, Wenhan Dou, Yue Deng, Yunjie Fu, Junqi Ge, Chenxia Han, Tammy Huang, Zhenhang Huang, Jerry Jiao, Shilei Jiang, Tianyu Jiao, Xiaoqi Jian, Lei Lei, Ruilin Li, Gen Luo, Tiantong Li, Xiang Lin, Ziyuan Liu, Zhiqi Li, Jie Ni, Qiang Ren, Pax Sun, Shiqian Su, Chenxin Tao, Bin Wang, Wenhai Wang, Haonan Wang, James Wang, Jin Wang, Jojo Wang, Letian Wang, Shizun Wang, Weizhi Wang, Zixuan Wang, Jinfan Xu, Sen Xing, Chenyu Yang, Hai Ye, Jiaheng Yu, Yue Yu, Muyan Zhong, Tianchen Zhao, Xizhou Zhu, Yanpeng Zhou, Yifan Zhang, Zhi Zhu
TL;DR
Open-source research agents have lagged proprietary systems because they are limited in scale, context, and interaction depth. MiroThinker v1.0 addresses this gap through interactive scaling, achieving strong cross-benchmark performance and predictable gains as agent–environment interactions deepen.
Problem
Open-source research agents are often limited in model scale, context length, and interaction depth compared with leading commercial research agents.
Method
MiroThinker combines model, context, and interactive scaling with ReAct tool use, a 256K context window, and reinforcement learning for deeper agent–environment interaction.
Results
Across diverse benchmarks, MiroThinker v1.0 shows consistent performance advantages over existing open-source research agents and approaches leading commercial systems.
Takeaways & Limitations
Interaction depth is established as a third scaling axis for research agents, complementing model capacity and context windows.
Takeaways & Limitations
Interactive scaling increases tool-use frequency but can produce marginal or redundant tool calls, while reinforcement learning can yield overly long, repetitive reasoning chains.
Abstract
from arXiv · showhide
We present MiroThinker v1.0, an open-source research agent designed to advance tool-augmented reasoning and information-seeking capabilities. Unlike previous agents that only scale up model size or context length, MiroThinker explores interaction scaling at the model level, systematically training the model to handle deeper and more frequent agent-environment interactions as a third dimension of performance improvement. Unlike LLM test-time scaling, which operates in isolation and risks degradation with longer reasoning chains, interactive scaling leverages environment feedback and external information acquisition to correct errors and refine trajectories. Through reinforcement learning, the model achieves efficient interaction scaling: with a 256K context window, it can perform up to 600 tool calls per task, enabling sustained multi-turn reasoning and complex real-world research workflows. Across four representative benchmarks-GAIA, HLE, BrowseComp, and BrowseComp-ZH-the 72B variant achieves up to 81.9%, 37.7%, 47.1%, and 55.6% accuracy respectively, surpassing previous open-source agents and approaching commercial counterparts such as GPT-5-high. Our analysis reveals that MiroThinker benefits from interactive scaling consistently: research performance improves predictably as the model engages in deeper and more frequent agent-environment interactions, demonstrating that interaction depth exhibits scaling behaviors analogous to model size and context length. These findings establish interaction scaling as a third critical dimension for building next-generation open research agents, complementing model capacity and context windows.
1. Introduction
MiroThinker addresses the performance gap between open-source and proprietary research agents by scaling model size, context length, and interaction depth. Its 72B variant achieves strong results across diverse research benchmarks while approaching commercial systems.
- Open-source research agents remain constrained by model scale, context length, and interaction depth compared with leading commercial systems.
- MiroThinker scales open-source research agents along model size, context length, and interaction depth through iterative reasoning and tool use.The system uses a 256K context window and supports up to 600 tool calls per task, with 8B, 30B, and 72B variants.
- 47.1% BrowseComp accuracy enables MiroThinker-v1.0-72B to surpass MiniMax-M2 by 2.0 points.
- 6.1 points higher on BrowseComp-ZH than GLM-4.6 demonstrates the model’s robust multilingual reasoning ability.
- 37.7% HLE accuracy places MiroThinker-v1.0-72B 4.8 points above Tongyi-DeepResearch.
- 81.9% GAIA-Text-Only accuracy surpasses MiniMax-M2’s 75.7% by 6.2 points.
2. Related Works
Related work has expanded LLMs with agentic capabilities and produced specialized deep research models for multi-hop, retrieval-intensive tasks. These systems combine planning, tool use, and external-environment interaction to acquire and synthesize knowledge.
- Agent Foundation Models incorporate decision-making, tool use, and external-environment interaction into general language-model training.
- Deep research models specialize in complex multi-hop reasoning and long-context, retrieval-intensive tasks through dynamic information seeking and iterative planning.
- Proprietary systems including OpenAI Deep Research, Claude Research, Kimi-Researcher, and Grok DeepSearch represent major developments in this research-agent direction.
3. Agentic Workflow
MiroThinker uses a ReAct loop that alternates reasoning, structured tool invocation, and observation, then summarizes the trajectory into a final answer. Modular tools and recency-aware context management support extended interaction within a 256K window.
- 3.1. Formulation: At each step, MiroThinker maintains thought–action–observation history and generates a thought followed by a structured tool invocation.
- 3.1. Formulation: The environment executes each action, returns an observation, and appends it to the trajectory for subsequent inference.
- 3.1. Formulation: When no further action is produced, a summary phase maps the completed trajectory to the final answer.
- 3.2. Tool Interface: The structured tool interface combines code execution, file management, and web retrieval to extend the agent beyond text generation.
- 3.3. Context Management: 600 tool calls fit within a 256K context window through result truncation and recency-based retention of recent tool responses.
- 3.3. Context Management: Recency-based retention preserves all thoughts and actions while masking older tool outputs, freeing context for extended reasoning without performance degradation.
4. Data Construction
MiroThinker’s training data combines multi-document QA synthesis with agentic trajectory synthesis to provide factual grounding and reasoning capabilities. The pipeline constructs linked-document reasoning tasks and diversifies trajectories across agent paradigms, tools, and models.
- 4.1. Dataset Overview: The synthetic dataset has two components: multi-document QA synthesis and agentic trajectory synthesis.Together, these stages provide factual grounding and reasoning capabilities.
- 4.2. Multi-document QA Synthesis: Multi-document QA construction links diverse sources into knowledge graphs and consolidated articles for cross-document reasoning.
- 4.2. Multi-document QA Synthesis: Fact extraction and constraint obfuscation produce questions requiring multi-hop integration rather than single-document retrieval or pattern matching.
- 4.3. Agentic Trajectory Synthesis: Agentic trajectory synthesis combines ReAct single-agent and MiroFlow multi-agent paradigms to generate iterative and collaborative reasoning workflows.
- 4.3. Agentic Trajectory Synthesis: Function calling and Model Context Protocol provide complementary structured and flexible mechanisms for tool invocation.
- 4.3. Agentic Trajectory Synthesis: Multiple leading LLMs and open-source QA datasets broaden trajectory styles, coverage, and reasoning diversity.
5. Training Pipeline
MiroThinker uses a three-stage training pipeline that progresses from supervised imitation to preference optimization and online reinforcement learning. The final stage trains the agent through direct environment interaction, supported by scalable environments, reward design, and trajectory curation.
- Pipeline overview: MiroThinker’s three-stage pipeline combines supervised fine-tuning, preference optimization, and reinforcement learning.The stages establish agentic behaviors, align decisions with task objectives, and promote exploration and generalization.
- 5.1. Agentic Supervised Fine-tuning: Supervised fine-tuning teaches multi-hop reasoning and tool use by imitating filtered expert thought–action–observation trajectories.The dataset pairs task instructions with expert trajectories, while filtering and repair remove repetition, duplication, and invalid tool calls.
- 5.2. Agentic Preference Optimization: Direct Preference Optimization ranks full interaction trajectories primarily by final-answer correctness without imposing fixed planning lengths or reasoning structures.An auxiliary SFT loss on preferred trajectories is combined with DPO to improve stability and preserve behavioral consistency.
- 5.3. Agentic Reinforcement Learning: Group Relative Policy Optimization trains the policy online from rollout trajectories, enabling creative solutions and adaptation through direct interaction with diverse environments.The GRPO objective uses group-relative advantages while maintaining proximity to a reference policy.
- 5.3. Agentic Reinforcement Learning: The RL system supports concurrent search, scraping, summarization, code execution, and Linux VM tasks, with rewards balancing correctness against format compliance.Streaming rollouts address long-tailed completion times, while trajectory curation removes environmental failures, redundant retries, formatting-only failures, and degenerate behavior.
6. Experiments
MiroThinker is evaluated across diverse agentic benchmarks using fixed ReAct-style settings, achieving strong open-source performance and competitive results against proprietary systems. Reinforcement learning increases interaction depth and frequency, improving accuracy while exposing limitations in tool-use quality, reasoning length, multilingual consistency, and sandbox proficiency.
- Overall Performance: 47.1% on BrowseComp and 51.0% on SEAL-0 place MiroThinker alongside advanced proprietary systems.The cited results compare MiroThinker with OpenAI DeepResearch, OpenAI o3, and Anthropic Claude 4.5.
- Overall Performance: 55.6% on BrowseComp-ZH and 77.8% on xbench-DeepSearch establish new open-source records on Chinese benchmarks.The results are presented as evidence of strong multilingual research performance.
- Overall Performance: The 8B and 30B variants achieve state-of-the-art performance within their respective size classes.This extends strong performance beyond the 72B model.
- Interactive Scaling: RL-trained MiroThinker-30B produces longer and deeper interaction trajectories than its SFT counterpart across four benchmarks.The RL model explores more solution paths and validates intermediate results before concluding.
- Interactive Scaling: 8–10 point average gains accompany increased interaction depth and frequency, defining interactive scaling as a third scaling dimension.The paper places interactive scaling alongside model size and context length.
- Limitations: RL increases tool invocation frequency, but some invocations remain marginal or redundant, limiting tool-use efficiency and action quality.The authors identify further optimization of tool use as necessary.
- Limitations: RL can produce excessively long, repetitive reasoning chains, while non-English inputs may trigger language mixing and sandbox use can cause timeouts or tool misuse.These limitations affect readability, Chinese performance, and code-execution or file-management workflows.
7. Conclusions
MiroThinker v1.0 advances tool-augmented reasoning through model, context, and interactive scaling. The paper presents deeper and more frequent agent–environment interactions as a third scaling axis, with predictable research gains across benchmarks.
- MiroThinker v1.0 advances tool-augmented reasoning through model, context, and interactive scaling.
- Deeper and more frequent agent–environment interactions enable error correction and knowledge acquisition beyond scaling model size or context length.
- Interactive scaling produces predictable research-performance gains across diverse benchmarks.
- Interaction depth is established as a third critical axis for building next-generation research agents.
A Contributions
The supplied passages identify the MiroMind Team and state that the authors are listed alphabetically by last name. They provide an author listing but no contribution descriptions.
- The paper identifies the author group as the MiroMind Team.
- The listed authors include Song Bai, Lidong Bing, Carson Chen, Guanzheng Chen, and Yuntao Chen.
- The author listing continues through contributors including Tianyu Jiao, Xiaoqi Jian, Lei Lei, Ruilin Li, and Gen Luo.