PaperlayerView exact revision on arXiv

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

arXiv:2511.11793v3cs.CL

TL;DR

Research-agent scaling has focused on model size and context length, leaving interaction depth as an additional dimension to examine. MiroThinker v1.0 uses a ReAct loop with 256K context and up to 600 tool calls per task to scale agent-environment interactions. It shows predictable interactive-scaling gains across benchmarks, including 81.9% GAIA-Text-Only accuracy, 6.2 points above MiniMax-M2 (75.7%).

  • Problem

    Whether interaction depth adds a scaling axis beyond model size and context length matters for building next-generation research agents.

    Source: §7 · p. 14

  • Method

    MiroThinker v1.0 uses a single-agent ReAct loop of reasoning, tool invocation, and observation to scale interaction depth, with 256K context supporting up to 600 tool calls per task.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Across diverse benchmarks, MiroThinker shows predictable gains from interactive scaling; GAIA-Text-Only accuracy is 81.9%, 6.2 points above MiniMax-M2 (75.7%).

    Source: §1 · p. 2, §7 · p. 14

  • Takeaways & Limitations

    Interaction depth is a third critical scaling axis, alongside model capacity and context windows, for building next-generation research agents.

    Source: §7 · p. 14

    The RL-tuned model invokes tools more often than the SFT model, but some invocations are marginal or redundant, leaving tool-use efficiency and action quality in need of further optimization.

    Source: §6.4 · p. 13

Abstract

from arXiv · show

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

Research agents need to formulate hypotheses, verify evidence, and synthesize information, but proprietary systems remain closed and open-source agents lag commercial systems. MiroThinker v1.0 advances open-source research agents through model size, context length, and interaction depth.

  • Open-weight LLM projects often lack end-to-end tools and frameworks, while specialized open-source research agents remain limited in scale, context length, and interaction depth [2] [4] [5] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22].Source: §1 · pp. 1–2
  • On BrowseComp [25], MiroThinker-v1.0-72B achieves 47.1% accuracy, surpassing MiniMax-M2 by 2.0 points; on BrowseComp-ZH [26], it improves over GLM-4.6 [4] by 6.1 points.Source: §1 · p. 2
  • With a 256K context window, MiroThinker supports up to 600 tool calls per task and is released in 8B, 30B, and 72B variants with a tool suite.Source: §1 · p. 2
  • 81.9% on GAIA-Text-Only surpasses MiniMax-M2 (75.7%) by 6.2 points, while BrowseComp-ZH improves over GLM-4.6 [4] by 6.1 points.Source: §1 · p. 2

2. Related Works

Related work moves from Agent Foundation Models trained for agentic abilities toward specialized deep-research agents that combine reasoning with information retrieval. The passages describe both proprietary and open-source systems in this progression.

Agent Foundation Models

  • Agent Foundation Models incorporate decision-making, tool use, and environmental interaction into base-model training, with GPT-5, Claude-4.5, and Kimi K2 [2] among examples.Current efforts particularly focus on code and search agents for tool-based problem solving, retrieval-augmented reasoning, and autonomous task execution.Source: §2 · p. 3

Deep Research Models

  • Deep research agents combine multi-hop reasoning, long-context retrieval, dynamic information seeking, and iterative planning to synthesize knowledge into comprehensive answers.Examples include proprietary systems such as OpenAI Deep Research and Claude Research, and open-source systems such as WebThinker [12] and Tongyi DeepResearch [11].Source: §2 · p. 3

3. Agentic Workflow

MiroThinker uses an iterative reasoning, tool-use, and observation workflow, supported by modular tools and context-management strategies for sustained interaction. Its recency-based retention preserves thoughts and actions while limiting retained tool responses, and the reported strategy does not degrade performance.

3.1. Formulation

  • Under ReAct [30], the single agent iterates through thoughts, structured tool actions, and observations until it stops acting, then produces a final answer from its trajectory.The trajectory records thoughts, actions, and observations, grounding the final response in the accumulated interaction history.Source: §3.1 · pp. 3–4

3.2. Tool Interface

  • The modular tool interface combines sandboxed shell and Python execution, bidirectional file transfer, remote-asset downloads, Google search, and targeted webpage extraction.The extraction tool uses a lightweight LLM, such as Qwen3-14B [6], to condense task-relevant content; HuggingFace access is disabled to prevent benchmark-answer leakage.Source: §3.2 · pp. 4–5

3.3. Context Management

  • With a 256K context window, MiroThinker’s two tool-response management strategies enable up to 600 tool calls; recency-based retention keeps recent responses while preserving the complete sequence of thoughts and actions.Excessively long tool outputs are truncated and marked with “[Result truncated]”.Source: §3.3 · pp. 5–6

4. Data Construction

MiroThinker’s training data combines synthetic MultiDocQA and agentic trajectories to develop factual grounding and reasoning. It also supplements these with open-source QA and post-training corpora for broader coverage and reasoning diversity.

4.1. MultiDocQA Synthesis

  • Category-balanced documents from interlinked sources are sampled into knowledge graphs, consolidated into linked articles, and used to construct MultiDocQA.The corpus includes sources such as Wikipedia, Common Crawl, and curated web repositories; sampled documents are connected by recursively following internal hyperlinks, with outside links pruned before consolidation.Source: §4.1 · pp. 6–7
  • Extracted cross-document facts are obfuscated into indirect constraints, then combined by an LLM into questions requiring multi-hop reasoning.Temporal and spatial details are generalized, while other entities and concepts are described indirectly; questions integrate constraints from different documents.Source: §4.1 · p. 7

4.2. Agentic Trajectory Synthesis

  • Trajectory synthesis uses ReAct Single-Agent [30] think–act–observe cycles and MiroFlow Multi-Agent coordination to generate complementary agent workflows.ReAct updates the agent’s understanding from tool observations; MiroFlow coordinates specialized agents through structured communication.Source: §4.2 · p. 7
  • Trajectory diversity comes from structured function calls and flexible MCP tool use, with multiple LLMs generating varied styles to mitigate single-model bias.Function calling uses predefined interfaces, whereas MCP supports context negotiation, tool composition, and dynamic discovery; synthesis models include GPT-OSS [32] and DeepSeek-V3.1 [5].Source: §4.2 · pp. 7–8

4.3. Open-Source Data Collection

  • Open-source QA pairs are converted into agentic trajectories, and post-training corpora are added to preserve conversational abilities and broaden reasoning and dialogue coverage.Included QA sources span multi-hop, science, and web tasks; AM-Thinking-v1-Distilled [43] and Nemotron-Post-Training-Dataset [44] are among the post-training corpora.Source: §4.3 · p. 8

5. Training Pipeline

MiroThinker is built on open-source Qwen2.5 and Qwen3 models [6] and trained in three stages: supervised fine-tuning, preference optimization, and reinforcement learning.

5.1. Agentic Supervised Fine-tuning

  • SFT teaches multi-hop reasoning and tool use by imitating expert thought–action–observation trajectories, after filtering and repairing noisy or invalid traces.Tool observations are prerecorded contextual inputs during training; tool execution is not performed.Source: §5.1 · p. 8

5.2. Agentic Preference Optimization

  • DPO [45] ranks trajectories by final-answer correctness without rigid agentic patterns, then combines preference optimization with an auxiliary SFT loss on preferred trajectories.Preference data is synthesized from the SFT model, and the auxiliary loss is intended to enhance stability and preserve behavioral consistency.Source: §5.2 · p. 9

5.3. Agentic Reinforcement Learning

  • GRPO [48] trains online from interactive rollouts across search, web, Python, and Linux environments, balancing answer correctness with format compliance.For each prompt, GRPO compares sampled trajectories against their group mean while keeping the policy close to a reference policy.Source: §5.3 · pp. 10–11

6. Experiments

MiroThinker v1.0 reports strong results across agentic benchmarks, with RL-driven deeper interactions accompanying accuracy gains. The experiments also identify limitations in tool-use efficiency, reasoning length, language consistency, and sandbox use.

6.1. Experimental Setup

  • Evaluation uses a ReAct-style agent with fixed 600-turn and 256K-context limits, benchmark-specific averaged runs, and LLM-as-a-Judge scoring.Models are initialized from Qwen2.5 and Qwen3; GAIA and HLE use text-only subsets, other benchmarks use full test sets, and HuggingFace access is disabled. Settings also fix temperature at 1.0, top-p at 0.95, output length at 16,384 tokens, and retention budget at 5; judges are gpt-4.1-2025-04-14 except HLE, which uses o3-mini-2025-01-31.Source: §6.1 · p. 11, §6.2 · p. 12

6.2. Overall Performance

  • 81.9% on GAIA sets a new state-of-the-art, 6.2 points above MiniMax-M2, while HLE reaches 37.7% and outperforms GPT-5-high by 2.5 points.The HLE evaluation uses its 2,158-question text-only subset.Source: §6.2 · pp. 11–12, §6.1 · p. 11, Table 1 · p. 12
  • 47.1% on BrowseComp and 51.0% on SEAL-0 are on par with advanced proprietary systems; 55.6% on BrowseComp-ZH and 77.8% on xBench-DeepSearch set open-source records.The cited proprietary systems include OpenAI DeepResearch, OpenAI o3, and Anthropic Claude 4.5.Source: §6.2 · p. 12, Table 1 · p. 12

6.3. Interactive Scaling

  • RL-tuned 30B trajectories are longer and deeper than SFT across four benchmarks, correlating with average accuracy gains of 8–10 points.Guided by verifiable rewards, RL supports exploring multiple strategies and validating intermediate results; the authors frame interaction depth as a third scaling dimension alongside model size and context length.Source: Figure 5 · p. 13, §6.3 · p. 13

6.4. Limitations

  • RL increases tool use, but some calls are marginal or redundant; code and file-management tools can time out or be misused, and sandbox initialization is sometimes missed.The model may use code execution to read web pages or PDFs instead of dedicated scraping tools.Source: §6.4 · pp. 13–14
  • RL can produce excessively long, repetitive reasoning that slows completion, while responses to Chinese queries may mix English and Chinese and impair performance.The language mixing may appear in internal reasoning or intermediate outputs.Source: §6.4 · p. 14

7. Conclusions

MiroThinker v1.0 is an open-source research agent that extends scaling to deeper and more frequent agent–environment interactions. The conclusions present interaction depth as a third critical axis for improving research capability, alongside model size and context length.

  • MiroThinker v1.0 is an open-source research agent that advances tool-augmented reasoning through model, context, and interactive scaling.Source: §7 · p. 14
  • Deeper and more frequent agent–environment interactions enable error correction and knowledge acquisition, with predictable research gains across diverse benchmarks.Source: §7 · p. 14
  • The conclusions frame interaction depth as a third critical scaling axis and present MiroThinker as a strong baseline and open platform for further exploration.Source: §7 · p. 14

Related papers

Loading 2511.11793v3…