Source-linked AI summary
Tongyi DeepResearch Technical Report
Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, Xixi Wu, Xuanzhong Chen, Yida Zhao, Zhen Zhang, Zhengwei Tao, Zhongwang Zhang, Zile Qiao, Chenxi Wang, Donglei Yu, Gang Fu, Haiyang Shen, Jiayin Yang, Jun Lin, Junkai Zhang, Kui Zeng, Li Yang, Hailong Yin, Maojia Song, Ming Yan, Minpeng Liao, Peng Xia, Qian Xiao, Rui Min, Ruixue Ding, Runnan Fang, Shaowei Chen, Shen Huang, Shihang Wang, Shihao Cai, Weizhou Shen, Xiaobin Wang, Xin Guan, Xinyu Geng, Yingcheng Shi, Yuning Wu, Zhuo Chen, Zijian Li, Yong Jiang
TL;DR
Deep-research agents need scalable ways to perform long-horizon reasoning and information seeking despite scarce, costly agentic data and limited open systems. Tongyi DeepResearch combines agentic mid-training, post-training, automated data synthesis, and stage-specific environments, achieving state-of-the-art results across multiple benchmarks with 3.3B activated parameters. The model, framework, and end-to-end solutions are released for open research.
Problem
Deep-research systems require autonomous multi-step reasoning and information seeking, but most systems are closed-source and agentic data is scarce and costly to annotate.
Method
Tongyi DeepResearch uses end-to-end agentic mid-training and post-training with fully automated data synthesis and customized environments for each training stage.
Results
Tongyi DeepResearch achieves state-of-the-art performance across multiple deep-research benchmarks while activating only 3.3 billion parameters per token.
Takeaways & Limitations
The released model, framework, and solutions provide an open foundation for reproducible research on autonomous deep-research agents.
Takeaways & Limitations
The current 128K context length remains insufficient for the most complex long-horizon tasks.
Abstract
from arXiv · showhide
We present Tongyi DeepResearch, an agentic large language model, which is specifically designed for long-horizon, deep information-seeking research tasks. To incentivize autonomous deep research agency, Tongyi DeepResearch is developed through an end-to-end training framework that combines agentic mid-training and agentic post-training, enabling scalable reasoning and information seeking across complex tasks. We design a highly scalable data synthesis pipeline that is fully automatic, without relying on costly human annotation, and empowers all training stages. By constructing customized environments for each stage, our system enables stable and consistent interactions throughout. Tongyi DeepResearch, featuring 30.5 billion total parameters, with only 3.3 billion activated per token, achieves state-of-the-art performance across a range of agentic deep research benchmarks, including Humanity's Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, xbench-DeepSearch, FRAMES and xbench-DeepSearch-2510. We open-source the model, framework, and complete solutions to empower the community.
1 Introduction
Tongyi DeepResearch targets open, autonomous deep-research agents through an end-to-end training approach, automated data synthesis, and customized environments. It reports state-of-the-art benchmark performance with an efficient model and releases the model, framework, and solutions.
- 1 Introduction: The work addresses a field where most deep-research systems remain closed-source and their intermediate research processes inaccessible.The authors position public methodology, models, and solutions as an unresolved community need.
- 1 Introduction: The end-to-end paradigm unifies agentic mid-training and post-training to develop deep reasoning and information-seeking behaviors.The stages are described as a progressive path from basic interaction skills to advanced autonomous research behaviors.
- 1 Introduction: A fully automated, scalable synthesis pipeline generates diverse agent trajectories without human annotation and supports each training stage with targeted data.The pipeline uses stage-specific synthesis strategies and is designed for fast validation and stable data distributions.
- 1 Introduction: Customized environments provide consistent interactions across training stages, spanning prior world models, simulated settings, and real-world contexts.These environments are tailored to the developmental stage of the agent.
- 1 Introduction: 32.9 on Humanity’s Last Exam, 43.4 on BrowseComp, and 90.6 on FRAMES demonstrate strong performance across deep-research benchmarks.The reported evaluation also includes BrowseComp-ZH, WebWalkerQA, GAIA, xbench-DeepSearch, and xbench-DeepSearch-2510.
2 Design Principle
The design principles center on scalable synthetic data and environments deliberately coupled to agent training. The framework balances stability, fidelity, cost, and long-horizon interaction through staged training and a simple ReAct-based formulation.
- 2 Design Principle: Agent training combines mid-training and post-training, introducing agentic learning before reinforcement learning rather than relying only on post-training.The pipeline co-designs the training algorithm and infrastructure for scalable, stable agent learning.
- 2 Design Principle: Synthetic agentic data is organized as question synthesis, behavior-data generation, and training use, enabling scalable coverage of research-level tasks.The paper motivates synthesis because agentic data is scarce and manual annotation is costly.
- 2 Design Principle: Prior-world, simulated, and real-world environments trade off stability, interaction cost, coverage, and feedback fidelity.Prior-world environments are cheapest and most scalable, simulated environments support controlled iteration, and real-world environments provide authentic feedback at higher cost and risk.
- 2 Design Principle: Mid-training primarily uses prior-world and simulated environments, while post-training validates strategies in simulation before final training in the real environment.The staged choice is intended to bootstrap capabilities cheaply and reserve costly interaction for validated policies.
- 3.1 Formulation: ReAct represents trajectories as interleaved thought, action, and observation, with actions including tool calls and the final response.Observations update the agent’s internal state and inform its next thought.
- 3.1 Formulation: The paper chooses ReAct for simplicity and scalability rather than adopting more complex single- or multi-agent designs.The rationale invokes the view that general methods leveraging scalable computation can outlast rigid, heavily engineered structures.
3.2 Overall Training Recipe
Tongyi DeepResearch is initialized from Qwen3-30B-A3B-Base and trained through agentic mid-training followed by post-training. Its mid-training synthesizes diverse workflow data, including questions, planning, reasoning, decision-making, and function-calling behaviors.
- 3.2 Overall Training Recipe: The training recipe integrates agentic mid-training and post-training, with mid-training presented as a bridge from the pretrained model to agentic post-training.The overall framework is designed for scalable reasoning and information seeking across complex research tasks.
- 3.2 Overall Training Recipe: Agentic continual pre-training begins with 32K context and expands to 128K, adding 64K–128K agentic behavior sequences for long-horizon reasoning and action.The two-stage context schedule progressively scales capability and sequence length.
- 3.2 Overall Training Recipe: The synthesis workflow covers question synthesis, planning action, reasoning action, and decision-making action across the agent lifecycle.These components model the cycle from an initial problem through iterative reflection and action to a final solution.
- 3.2 Overall Training Recipe: Planning data decomposes synthesized questions and predicts initial actions, with rejection sampling used to improve planning-output quality.The paper treats planning accuracy as closely correlated with successful task completion.
- 3.2 Overall Training Recipe: The recipe also scales function-calling data across diverse environments to strengthen general agentic capability.Environment diversity is presented as closely tied to the breadth of function-calling competence.
3.4 Agentic Post-training
Agentic post-training combines automated data synthesis, supervised cold-start fine-tuning, and reinforcement learning in customized environments. The pipeline emphasizes high-quality data and stable interactions, then merges model variants to preserve complementary capabilities.
- 3.4 Agentic Post-training: The post-training pipeline proceeds through data synthesis, supervised fine-tuning for cold start, and agentic reinforcement learning.Cold-start training retains high-quality trajectories after rejection sampling and uses mixed formulations for robustness and generalization.
- 3.4.1 High-quality Data Synthesis: Fully automated synthesis constructs difficult research QA data from interconnected knowledge structures and progressively increases question complexity.The process uses knowledge graphs, web-derived information, sampled subgraphs or subtables, and controllable atomic operations to raise uncertainty.
- 3.4.3 Agentic Reinforcement Learning: Agentic RL trains on complete rollouts, rewards final-answer correctness, and iteratively improves the policy through environment interaction.The framework supports simulated and real-world environments, while the implementation uses asynchronous rollout infrastructure and customized GRPO-based training.
- 3.4.3 Agentic Reinforcement Learning: Stable environments and high-quality data are more important to agentic RL success than the specific algorithm used.The authors therefore prioritize environment stability and data curation, with only limited algorithmic modifications for training stability.
- 3.4.4 Model Merging: Model merging forms the final model by weighted averaging parameters from variants with different capability preferences.The authors report that interpolation preserves source strengths and provides robust generalization without additional optimization costs.
4 Experiments
Tongyi DeepResearch is evaluated across diverse deep-research benchmarks using reproducible multi-run protocols and achieves strong results with sparse activation. Its Heavy Mode further scales test-time reasoning through parallel research and context-managed synthesis, while detailed analyses examine robustness, training dynamics, and tool-enabled generalization.
- 4.2 Main Results: Tongyi DeepResearch achieves the highest scores on nearly all evaluated benchmarks, surpassing open and closed systems while activating only 3.3 billion parameters per token.The evaluation spans Humanity’s Last Exam, BrowseComp, BrowseComp-ZH, GAIA, xbench-DeepSearch, WebWalkerQA, and FRAMES; the model ranks just below ChatGPT-5-Pro on xbench-DeepSearch-2510.
- 4.3 Heavy Mode: Heavy Mode deploys parallel agents along diverse tool-use and reasoning paths, then uses a synthesis model to consolidate their findings into the final answer.Compressed context-management reports preserve essential reasoning while avoiding the context cost of aggregating full trajectories.
- 4.3 Heavy Mode: 38.3% on Humanity’s Last Exam and 58.1% on BrowseComp-ZH establish state-of-the-art Heavy Mode results, while BrowseComp reaches 58.3%.Heavy Mode uses parallel exploration and intelligent aggregation to leverage test-time compute.
- 4.4 Detailed Analysis: Training rewards increase throughout reinforcement learning while policy entropy stabilizes after an initial rise, indicating sustained learning without collapse or explosion.The authors attribute this behavior to dynamic data curation, environment design, and algorithm modification.
- 4.4 Detailed Analysis: Tool use improves general-benchmark performance over the reasoning-only base model by retrieving external information and supporting mathematical computation through Python.The evaluation covers AIME25, HMMT25, and SimpleQA.
5 Discussion
The discussion highlights practical deployment considerations for Tongyi DeepResearch, including efficiency, accessibility, and current limitations. The authors also position open-source agentic systems as a route toward broader research access and capability.
- Evaluation: Figure 11 reports performance on general benchmarks, providing a broader evaluation context for the system.
- Limitations: 128K context remains insufficient for the most complex long-horizon tasks.The authors motivate extended context windows or improved context-management mechanisms as future work.
- Practical significance: Smaller agentic models can improve deployment efficiency, accessibility, and responsiveness across real-world scenarios.The authors connect smaller models with edge-device deployment and broader practical use.
- Open research: The authors advocate open-source models with emergent agency to democratize agentic intelligence and deepen understanding of agency in open systems.
6 Conclusion
The conclusion presents Tongyi DeepResearch as an open-source agent that combines scalable training and automated data synthesis to support autonomous research. It reports strong benchmark performance with only 3.3B parameters activated per token.
- Conclusion: Tongyi DeepResearch unifies agentic mid-training and post-training with automated data synthesis and stage-specific environments.This combination is described as enabling autonomous planning, searching, reasoning, and information synthesis.
- Conclusion: Activating only 3.3B parameters, Tongyi DeepResearch achieves state-of-the-art results across multiple deep research benchmarks.The conclusion states that the model surpasses strong proprietary systems.
- Conclusion: The work establishes a foundation for open, reproducible research into autonomous AI agents and more general self-improving intelligence.
Contributions
The listed contributors comprise a core-contributor group and a broader contributor group. The names are ordered alphabetically by first name.
- Contributor listing: Contributor names are listed alphabetically by first name.
- Core Contributors: The core-contributor list includes 26 named contributors.
- Contributors: The broader contributor list includes 30 named contributors.
A Rollout Details
The rollout specifies a deep-research assistant that conducts multi-source investigations and returns final answers in a prescribed format. Its available functions cover web search, webpage visits, code execution, scholarly search, and file parsing.
- Assistant behavior: The assistant is instructed to perform thorough, multi-source investigations across broad and specialized research questions.It must synthesize credible and diverse sources into comprehensive, accurate, and objective responses.
- Assistant behavior: Final responses must be enclosed within <answer></answer> tags after sufficient information has been gathered.
- Available tools: The rollout exposes search, webpage-visit, Python, Google Scholar, and file-parsing functions.The function specifications support multiple queries or URLs where applicable and define their required arguments.
- Python execution: Python execution requires an empty arguments object, code enclosed in <code> tags, and explicit printing of desired output.
- Call format: Tool calls are represented as JSON objects inside <tool_call></tool_call> XML tags.The rollout also specifies the current date field in the surrounding tool-call format.
B Evaluation Details
Evaluation uses task-specific procedures, combining judge models, manual review, and official scripts to match benchmark requirements.
- Judge models evaluate GAIA, WebWalkerQA, xbench-DeepSearch, xbench-DeepSearch-2510, BrowseComp, and BrowseComp-ZH.Qwen2.5-72B-Instruct judges GAIA and WebWalkerQA; Gemini-2.0-Flash-001 judges both xbench datasets; GPT-4o-2024-08-06 judges both BrowseComp datasets.
- Humanity’s Last Exam is evaluated on 2,154 text-only questions following prior work.
- Mathematical benchmarks use manual evaluation because outputs are detailed reports and AIME25 and HMMT25 are small.Each listed mathematical benchmark contains 30 examples.
- SimpleQA uses its official evaluation script across 4,326 knowledge-based examples.
C Post-training Synthetic Data Case
The synthetic-data cases span difficult historical, cross-source, and domain-specific questions, illustrating question–answer pairs that are difficult to produce through human annotation.
- One case requires identifying a military unit from clues about an officer, mineral discovery, commemorative coin, and later conflict.
- Another case links a southeastern English port, historical murders, patents, nobility, and manuscript-poem collecting to ask for a German princely status.The stated answer is Fürstenstand.
- The third case is a PhD-level chemistry question requiring domain expertise and multi-step reasoning.It concerns bond-length changes in trisilylamine derivatives and the orbital interaction identified as n →σ∗.
- The first two cases are synthetically generated high-quality, high-uncertainty, superhuman question–answer pairs.The passages describe these examples as exceptionally difficult to produce via human annotation.
D Environment Details
Tongyi DeepResearch uses five specialized tools for search, extraction, computation, scholarly retrieval, and file analysis, with open implementations supplied for public use.
- The environment provides Search, Visit, Python Interpreter, Google Scholar, and File Parser.
- Search retrieves concurrent Google queries with the top 10 results, while Visit extracts information relevant to page-specific goals.Visit parses pages with Jina and then uses a summary model for targeted extraction.
- Python Interpreter runs code in a secure sandbox for dynamic computation, data manipulation, and library use.Code is supplied within <code> tags, and visible outputs must be printed.
- Google Scholar retrieves academic publications, articles, papers, and citations from multiple search queries.
- File Parser converts documents, web pages, and multimedia into plain text before generating direct answers from the unified text.
- Alternative open implementations are provided for internal APIs and fallback strategies, and extensive testing verified faithful reproduction of results.