PaperlayerView exact revision on arXiv

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

arXiv:2510.24701v3cs.CLcs.AIcs.IRcs.LGcs.MA

TL;DR

Most deep-research systems remain closed-source, with their intermediate processes inaccessible, while deep research involves multi-step reasoning and internet information seeking. Tongyi DeepResearch combines agentic mid-training and post-training with automated data synthesis and stage-specific environments, achieving state-of-the-art results across multiple benchmarks.

  • Problem

    Most deep-research systems remain closed-source and their intermediate processes inaccessible, despite deep research requiring multi-step reasoning and internet information seeking.

    Source: §1 · p. 2

  • Method

    Tongyi DeepResearch combines agentic mid-training and post-training in an end-to-end training paradigm, supported by automated data synthesis and stage-specific environments.

    Source: §1 · p. 2

  • Results

    Tongyi DeepResearch achieves state-of-the-art results across multiple deep-research benchmarks, surpassing strong proprietary systems while activating only 3.3 billion parameters per token.

    Source: §4.2 · p. 12, §6 · p. 16

  • Takeaways & Limitations

    The work establishes a foundation for open, reproducible research into autonomous AI agents.

    Source: §6 · p. 16

    The current 128K context length remains insufficient for the most complex long-horizon tasks.

    Source: §5.1 · p. 16

Abstract

from arXiv · show

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 aims to give LLMs autonomous research capabilities through an end-to-end agentic training framework, automated data synthesis, and stage-specific environments; the report presents its benchmark results and releases the model, framework, and solutions.

1 Introduction

  • The work addresses the lack of systematic methodology and publicly available models by introducing Tongyi DeepResearch and releasing its model, framework, and end-to-end solutions.The stated goal is to enable LLMs to plan, search, reason, and synthesize knowledge across extended actions and diverse information sources.Source: §1 · pp. 2–3
  • The end-to-end agentic training paradigm unifies mid-training that cultivates agentic biases with post-training that unlocks potential through scalable multi-turn reinforcement learning.Together, the stages develop capabilities from basic interaction skills to advanced autonomous research behaviors.Source: §1 · p. 2
  • A fully automated, scalable data-synthesis pipeline generates diverse, high-quality agent trajectories without human annotation, using stage-specific strategies and targeted data.The passage characterizes synthetic data as scalable, fast to validate, and an engine for agent training.Source: §1 · p. 2
  • Customized environments provide consistent interactions across training stages, aligned with each stage’s objectives and spanning world models, simulations, and real-world contexts.They support data synthesis through rich, specialized interactions.Source: §1 · p. 2
  • 32.9 on Humanity’s Last Exam, 43.4 on BrowseComp, 46.7 on BrowseComp-ZH, and 72.2 on WebWalkerQA; the model outperforms strong baselines including OpenAI-o3 and Deepseek-V3.1.Tongyi DeepResearch has 30.5 billion total parameters and activates 3.3 billion per token; it is built on Qwen3-30B-A3B-Base (Yang et al., 2025), and the cited baselines are OpenAI-o3 and Deepseek-V3.1.Source: §1 · p. 2
  • 70.9 on GAIA, 75.0 on xbench-DeepSearch, 90.6 on FRAMES, and 55.0 on xbench-DeepSearch-2510.The report also presents general benchmark performance on AIME25, HMMT25, and SimpleQA, and systematic analysis of agentic reinforcement learning and synthetic data.Source: §1 · p. 2

2 Design Principle · 3.1 Formulation

The design couples agentic mid-training and post-training with synthetic-data scaling and deliberately staged environments. Its formulation uses ReAct rollouts and dynamic context reconstruction for long-horizon research tasks.

2 Design Principle

  • Mid-training supplies agentic prior knowledge before SFT and reinforcement learning in post-training, addressing the optimization conflicts of learning agentic capability and alignment simultaneously.SFT establishes a behavioral baseline through curated demonstrations; environment-coupled RL adds exploration, planning, and execution.Source: §2 · p. 3
  • Research-level agentic data is scarce and costly to annotate, so the pipeline relies on synthetic data across three steps: question synthesis, behavior-data generation, and training use.Mid-training needs large, diverse trajectories, while post-training needs high-quality, verifiable data for reward signals.Source: §2 · pp. 3–4
  • Synthetic data supports scalable question generation, diverse problem patterns, targeted meta-capability training, easier verification, and iterative data flywheels.Tasks can be decomposed into skills such as planning, information synthesis, and memory management; trained agents can then generate data with stronger reasoning and planning patterns.Source: §2 · pp. 3–4
  • Because real-world interaction is costly and non-stationary, training combines prior-world, simulated, and real-world environments, using cheap synthetic bootstrapping before validating in simulation and training verified policies in reality.Prior-world environments are stable and scalable but lack feedback; simulations enable controlled, low-cost iteration but have limited coverage and a sim-to-real gap; real-world environments offer fidelity at higher cost and risk.Source: §2 · p. 4

3.1 Formulation

  • Each rollout comprises Thought, Action, and Observation: internal analysis and planning, tool operations or the final report, and environmental feedback that informs the next thought.Available tools include Search, Visit, Python Interpreter, Google Scholar, and File Parser.Source: §3.1 · pp. 4–5
  • Tongyi DeepResearch uses vanilla ReAct (Yao et al., 2023), interleaving reasoning and actions into thought-action-observation trajectories, favoring a scalable general method over complex engineered structures.The policy conditions each step on prior interaction history; the final action is the task answer. The design cites “The Bitter Lesson” and Li et al. (2025a).Source: §3.1 · p. 5
  • Dynamic context management reconstructs a Markovian workspace from the question, evolving compressed report, and latest action and observation, avoiding full-history conditioning and context overflow.Requiring periodic synthesis and prioritization helps maintain coherent reasoning across arbitrary exploration depths.Source: §3.1 · p. 5

3.2 Overall Training Recipe

Tongyi DeepResearch starts from Qwen3-30B-A3B-Base1 and is developed through an end-to-end framework combining agentic mid-training and post-training for reasoning and information seeking across complex research tasks.

3.2 Overall Training Recipe

  • The training framework integrates agentic mid-training and post-training to support scalable reasoning and information seeking across complex research tasks.The report presents this approach as a new paradigm for training agentic models.Source: §3.2 · pp. 5–6
  • Tongyi DeepResearch is initialized from the pretrained base model Qwen3-30B-A3B-Base1.Source: §3.2 · p. 5
  • The report presents mid-training first in Section 3.3, followed by post-training in Section 3.4.Source: §3.2 · p. 6

3.3 Agentic Mid-training

Tongyi DeepResearch uses two-stage Agentic Continual Pre-training to bridge pretrained models and agentic post-training, building agentic behavior while preserving broad language competence. Its synthesized training data spans agent workflows and varied simulated environments.

3.3.1 Training Configuration

  • The two-stage Agentic CPT uses Next-Token Prediction to develop agentic inductive bias while retaining broad linguistic competence.It serves as a bridge between pretrained models and agentic post-training.Source: §3.3.1 · p. 6

3.3.2 Large-scale Agent Behavior Data Synthesis

  • Synthesized agent-workflow data covers question synthesis, planning, reasoning, and decision-making; questions draw on entity-anchored open-world memory to elicit behavioral patterns.The memory consolidates structured knowledge from sources including web-crawled data and agent interaction trajectories; question patterns include multi-hop reasoning and numerical computation.Source: §3.3.2 · p. 6
  • Planning data pairs problem decomposition and first-action prediction with rejection sampling grounded in the entities and knowledge used to construct each question.Open-source models analyze, decompose, and predict initial actions; the authors link planning accuracy to successful task completion.Source: §3.3.2 · p. 7
  • Reasoning data uses a two-stage process to generate complete chains from questions and dependent knowledge, filtered by reasoning length and answer consistency.The method targets coherent reasoning over heterogeneous data, including distilling critical knowledge from noisy, unstructured tool responses.Source: §3.3.2 · p. 7
  • Decision-making data reconstructs demonstration trajectories into multi-step decision sequences while preserving original choices after exploring feasible actions at each step.This explicitly represents the choice among potential reasoning and action paths.Source: §3.3.2 · p. 7
  • Automatically constructed, fully simulated read–write database environments broaden function-calling scenarios, and their data are incorporated into mid-training.Environment scaling is used to systematically increase environment diversity and general agentic capability.Source: §3.3.2 · p. 7

3.4 Agentic Post-training

Agentic post-training combines synthetic-data generation, supervised fine-tuning for a cold start, and agentic reinforcement learning, followed by model merging.

3.4.1 High-quality Data Synthesis

  • Synthetic-data generation scales difficult research questions through automated knowledge construction, controlled complexity increases, and iterative question upgrades.The pipeline uses web-derived knowledge graphs and real-world tables, increases uncertainty and difficulty through atomic operations and set-theoretic modeling (Wu et al., 2025a; Tao et al., 2025), and iteratively upgrades multi-source reasoning questions (Qiao et al., 2025).Source: §3.4.1 · pp. 7–8

3.4.2 Supervised Fine-tuning for Cold Start

  • Cold-start SFT trains on rigorously filtered, diverse trajectories from high-performing open-source models using both ReAct and context-management formulations.Training first uses 40K context, then extends to 128K; context-management samples train the model to summarize observations while maintaining focus across extended trajectories.Source: §3.4.2 · p. 8

3.4.3 Agentic Reinforcement Learning

  • A unified sandbox stabilizes real-world tool use, while offline Wikipedia-based simulation and asynchronous parallel rollouts support rapid, reliable RL experimentation.The sandbox uses controls including caching, retries, rate limits, and failover; the asynchronous framework separates model inference from tool invocation.Source: §3.4.3 · pp. 8–9
  • The authors find that agentic RL success depends more on data quality and environment stability than on the specific algorithm.They use strict on-policy rollouts with binary answer-correctness rewards, and dynamically filter and refresh training problems as the policy improves; selected negative rollouts are excluded to stabilize training.Source: §3.4.3 · p. 10

3.4.4 Model Merging

  • Weighted averaging or interpolation of variants from the same base model produces a merged model that generalizes robustly without additional optimization.In complex scenarios, the merged model performs comparably to the best source model in its respective area of strength.Source: §3.4.4 · p. 11

4 Experiments

Tongyi DeepResearch achieves the highest scores on nearly all evaluated benchmarks, while analyses report strong Heavy Mode results and examine training stability, context limits, and interaction scaling.

4.1 Experimental Setup

  • Evaluation spans seven public information-seeking benchmarks and compares Tongyi DeepResearch with LLM-based ReAct agents and end-to-end deep-research systems.Evaluations use fixed inference settings, allow up to 128 tool invocations and 128K context, and report average performance over three independent runs as Avg@3.Source: §4.1 · p. 11

4.2 Main Results

  • Tongyi DeepResearch scores highest on nearly all evaluated benchmarks, surpasses open and closed systems including OpenAI o3 and Gemini DeepResearch, and ranks just below ChatGPT-5-Pro on xbench-DeepSearch-2510.The report describes it as a new state of the art among open-source deep research agents, with 3.3 billion activated parameters per token.Source: §4.2 · pp. 11–12, Table 1 · p. 12

4.3 Heavy Mode

  • 38.3% on Humanity’s Last Exam and 58.1% on BrowseComp-ZH are Heavy Mode’s state-of-the-art results; it remains competitive on BrowseComp at 58.3%.Heavy Mode parallelizes agents exploring diverse tool-use and reasoning paths, then synthesizes their compressed reports to scale test-time compute without aggregating full trajectories.Source: §4.3 · pp. 12–13

4.4 Detailed Analysis

  • Pass@3 reaches 59.64 on BrowseComp, 63.67 on BrowseComp-ZH, and 45.9 on Humanity’s Last Exam; Avg@3 is consistent with best Pass@1 despite evaluation instability.The fine-grained comparison reports Avg@3, best Pass@1 over three runs, and Pass@3.Source: §4.4 · p. 13, Figure 7 · p. 13
  • Training rewards rise while policy entropy stabilizes after a brief initial increase, which the report attributes to dynamic data curation, environment design, and algorithm modification.The sustained reward improvement is described as evidence of effective policy learning without entropy collapse or explosion.Source: §4.4 · p. 13, Figure 8 · p. 14
  • With curricula curated by the 64k-context model, all 32k, 48k, and 64k variants learn stably, but the 64k model achieves the highest reward ceiling.The 32k model’s response length declines as it is implicitly incentivized to find concise solutions to problems that may require longer-than-32k-token solutions.Source: §4.4 · p. 14, Figure 9 · p. 14
  • More environment interactions consistently improve BrowseComp performance; tool use also improves AIME25, HMMT25, and SimpleQA over reasoning-only performance.Search supports knowledge-intensive tasks, while Python Interpreter provides computational support for mathematical reasoning.Source: §4.4 · p. 15, Figure 10 · p. 15
  • Over 20% of SFT samples exceed 32k tokens and involve more than 10 tool invocations; Wiki simulation’s reward curve closely matches real-environment RL.The report says this synthetic data initializes deep reasoning before RL, and the simulation enables faster algorithm iteration.Source: §4.4 · p. 15, Figure 10 · p. 15

5 Discussion

The discussion presents compact agent training as practically valuable, identifies limitations in the current system, and outlines a path from domain-specific research agents toward general-purpose agents.

5.1 Limitations

  • The authors identify limitations in context length, model scale, report generation, reinforcement-learning efficiency, and robustness beyond predefined research prompts and tools.The 128K context is insufficient for the most complex long-horizon tasks; a larger model is in progress, and report fidelity and preference alignment remain under improvement (Li et al., 2025e). They also plan to explore partial rollouts while addressing off-policy distributional shift, extend beyond predefined prompts and tools, and investigate longer context windows or improved context management (Qiao et al., 2025; Wu et al., 2025c).Source: §5.1 · p. 16

5.2 Model Scale

  • Training agentic capabilities on smaller models can improve deployment efficiency on edge devices, broaden accessibility, and enable faster, more responsive interactions (Belcak et al., 2025).The authors frame this direction as supporting powerful yet practically deployable autonomous research agents.Source: §5.2 · p. 16

5.3 What’s Next

  • The authors advocate open-source agentic models and aim to develop a unified foundation model for autonomous reasoning, planning, and action across diverse domains.Their next-generation agent foundation model is intended to combine scalable reasoning, memory, and autonomy, moving beyond domain-specific agents toward general-purpose agents with minimal human supervision.Source: §5.3 · p. 16

6 Conclusion

Tongyi DeepResearch combines agentic mid-training and post-training in an open-source, end-to-end framework for autonomous deep research. Automated data synthesis and stage-specific environments support autonomous planning, searching, reasoning, and information synthesis.

6 Conclusion

  • Activating only 3.3B parameters, Tongyi DeepResearch achieves state-of-the-art results across multiple deep research benchmarks and surpasses strong proprietary systems.The conclusion presents the model's efficiency alongside its benchmark performance.Source: §6 · p. 16
  • Tongyi DeepResearch unifies agentic mid-training and post-training in a scalable, end-to-end paradigm for deep research.The system uses automated data synthesis and stage-specific environments to train autonomous planning, searching, reasoning, and information synthesis.Source: §6 · p. 16
  • The work releases an open-source deep research agent and establishes a foundation for open, reproducible research into autonomous AI agents.The conclusion describes this as a step toward more general, self-improving intelligence.Source: §6 · p. 16

A Rollout Details

The ReAct rollout prompt defines the assistant’s research role and requires a tagged final answer, while specifying available tools and how to call them.

A Rollout Details

  • The assistant is instructed to conduct thorough, multi-source research across broad or specialized topics and deliver comprehensive, accurate, objective answers inside <answer> tags.The prompt frames this as the core function for each user request.Source: A Rollout Details · p. 21
  • Available tools include Google search, webpage visits, Python execution, Google Scholar, and parsing uploaded files.Search accepts multiple queries; webpage visits require URLs and an information goal; Scholar also returns Google results.Source: A Rollout Details · p. 21
  • Tool calls must use <tool_call> tags with a JSON object naming the function and arguments, and Python output must be printed to standard output.Python calls use an empty arguments object, with code placed in <code> tags.Source: A Rollout Details · p. 21
  • The supplied instructions constitute the system prompt used for the ReAct rollout.Source: A Rollout Details · p. 22

B Evaluation Details

The section specifies benchmark-specific judges and protocols, including consistency with official or original-paper evaluation prompts. For general benchmarks, it uses manual evaluation for mathematical reports and SimpleQA’s official script for knowledge questions.

B Evaluation Details

  • Qwen2.5-72B-Instruct judged GAIA (166 examples) and WebWalkerQA (680), Gemini-2.0-Flash-001 judged both xbench sets (100 each), and GPT-4o-2024-08-06 judged BrowseComp (1,266) and BrowseComp-ZH (289).GAIA and WebWalkerQA follow Li et al. (2025d).Source: B Evaluation Details · p. 22
  • Humanity’s Last Exam evaluation used 2,154 text-only questions following Chai et al. (2025), with o3-mini as evaluator and prompts aligned with the original-paper protocol.The authors state this alignment supports reproducibility.Source: B Evaluation Details · p. 22
  • For mathematical problems, AIME25 and HMMT25 (30 examples each) received manual evaluation, while SimpleQA (4,326 examples) used its official evaluation script.Manual scoring was chosen because the system produces detailed reports and the math datasets are small.Source: B Evaluation Details · p. 22

C Post-training Synthetic Data Case

The first two cases are synthetically generated high-quality, high-uncertainty, superhuman question–answer pairs that are exceptionally difficult to produce via human annotation; the third is a PhD-level research question requiring deep expertise and multi-step reasoning.

C Post-training Synthetic Data Case

D Environment Details

Tongyi DeepResearch uses five tools for web search, targeted page extraction, computation, scholarly retrieval, and document analysis. Alternative open-source implementations replace internal APIs and fallback strategies and were tested to reproduce the reported results.

D Environment Details

  • Search runs concurrent queries through Google and returns the top 10 ranked results per query, each with a title, snippet, and URL.A single call accepts one or more queries.Source: D Environment Details · p. 23
  • Visit extracts goal-relevant information by parsing each page with Jina (Jina.ai, 2025) and having a summary model filter its content.Each input page is paired with a dedicated information-seeking goal.Source: D Environment Details · p. 23
  • Python Interpreter runs code in a sandbox and captures standard output, supporting secure computation, data manipulation, and Python library use.Code must be enclosed in <code> tags, and desired results must be sent to print().Source: D Environment Details · p. 23
  • Google Scholar retrieves scholarly literature, including articles, papers, and citations, and accepts multiple distinct queries in one call.Source: D Environment Details · p. 23
  • File Parser answers queries over local or URL-based documents, webpages, and multimedia by converting inputs to text, transcribing audio or video when needed, and summarizing the text.Supported examples include PDF, DOCX, and MP4 files.Source: D Environment Details · p. 23
  • Alternative open implementations replace internal APIs and fallback strategies, and extensive testing verified they can faithfully reproduce the results.The report points readers to the open-source GitHub repository for these implementations.Source: D Environment Details · p. 23

Related papers

Loading 2510.24701v3…