DeepAnalyze: Agentic Large Language Models for Autonomous Data Science (2510.16872v1)

Abstract: Autonomous data science, from raw data sources to analyst-grade deep research reports, has been a long-standing challenge, and is now becoming feasible with the emergence of powerful LLMs. Recent workflow-based data agents have shown promising results on specific data tasks but remain fundamentally limited in achieving fully autonomous data science due to their reliance on predefined workflows. In this paper, we introduce DeepAnalyze-8B, the first agentic LLM designed for autonomous data science, capable of automatically completing the end-toend pipeline from data sources to analyst-grade deep research reports. To tackle high-complexity data science tasks, we propose a curriculum-based agentic training paradigm that emulates the learning trajectory of human data scientists, enabling LLMs to progressively acquire and integrate multiple capabilities in real-world environments. We also introduce a data-grounded trajectory synthesis framework that constructs high-quality training data. Through agentic training, DeepAnalyze learns to perform a broad spectrum of data tasks, ranging from data question answering and specialized analytical tasks to open-ended data research. Experiments demonstrate that, with only 8B parameters, DeepAnalyze outperforms previous workflow-based agents built on most advanced proprietary LLMs. The model, code, and training data of DeepAnalyze are open-sourced, paving the way toward autonomous data science.

• The paper presents DeepAnalyze-8B, which autonomously manages the entire data science workflow without relying on predefined orchestration frameworks.
• It employs a curriculum-based agentic training paradigm and data-grounded trajectory synthesis to enhance reasoning, code generation, and multi-turn interactions.
• Experimental results demonstrate that DeepAnalyze-8B outperforms many proprietary models on benchmarks for data analysis, report generation, and structured data understanding.

DeepAnalyze: Agentic LLMs for Autonomous Data Science

Introduction

The paper presents DeepAnalyze-8B, an agentic LLM specifically designed for autonomous data science. Unlike prior workflow-based agents and domain-specific LLMs, DeepAnalyze-8B is trained to orchestrate and optimize the entire data science pipeline, from raw data sources to analyst-grade research reports, without reliance on predefined workflows or external orchestration frameworks. The model is built upon a curriculum-based agentic training paradigm and a data-grounded trajectory synthesis framework, enabling it to acquire and integrate multiple capabilities required for complex, real-world data science tasks.

Figure 1: DeepAnalyze-8B is the first end-to-end agentic LLM that achieves autonomous data science, supporting entire data science pipeline and open-ended data research.

Model Architecture and Agentic Capabilities

DeepAnalyze-8B extends conventional LLM architectures by introducing a data-oriented interaction pattern. The model supports both passive ingestion of structured data (e.g., CSV, XLSX, Markdown) and active inspection of external data sources, emulating the workflow of human data scientists. The agentic framework is realized through five specialized actions: $\langle$Analyze$\rangle$, $\langle$Understand$\rangle$, $\langle$Code$\rangle$, $\langle$Execute$\rangle$, and $\langle$Answer$\rangle$. These actions are autonomously generated by the model, allowing for dynamic planning, reasoning, code generation, execution, and reporting within a real-world environment.

Figure 2: Architecture of DeepAnalyze.

The inference process is iterative: the model plans, interacts with data, generates and executes code, and refines its actions based on environmental feedback until the task is completed. This agentic loop is fully internalized, with no external workflow or human-defined rules.

Figure 3: Examples of DeepAnalyze-8B autonomously orchestrating and optimizing actions to complete a data science pipeline and open-ended data research.

Curriculum-based Agentic Training

To address the challenges of reward sparsity and trajectory scarcity inherent in complex data science tasks, DeepAnalyze employs a two-stage curriculum-based agentic training paradigm:

1. Single-ability Fine-tuning: The foundation LLM is fine-tuned on long-chain-of-thought (CoT) data for reasoning, structured data understanding, and code generation. This stage establishes robust single abilities.
2. Multi-ability Agentic Training: Building on the single abilities, the model undergoes agentic reinforcement learning (RL) in real-world environments, using group relative policy optimization (GRPO) and hybrid reward modeling. The RL phase leverages synthesized interaction trajectories and rewards based on both accuracy and interaction quality, as well as report quality for open-ended research.

   Figure 4: Schematic diagram of agentic RL.

This curriculum enables the model to progressively acquire composite skills, transitioning from basic reasoning and code generation to autonomous orchestration and adaptive optimization across the entire data science pipeline.

Data-grounded Trajectory Synthesis

High-quality training data is synthesized via a data-grounded trajectory synthesis framework, which consists of:

• Reasoning Trajectory Synthesis: Existing instruction-response datasets are enhanced by distilling and refining reasoning trajectories, with explicit insertion of reasoning keywords to improve structured data focus.
• Interaction Trajectory Synthesis: Multi-agent systems generate multi-turn interaction trajectories from real data sources, with rigorous validation and filtering to ensure trajectory quality.

Figure 5: The proposed data-grounded trajectory synthesis for the development of DeepAnalyze on data science tasks.

This approach yields DataScience-Instruct-500K, a large-scale dataset (500K samples) covering both single-ability and multi-ability tasks, with sequence lengths up to 32K.

Figure 6: Length distribution of training data.

Experimental Results

DeepAnalyze-8B is evaluated on 12 benchmarks, including DataSciBench, DSBench, DABStep, DABStep-Research, DS-1000, and multiple TableQA datasets. The model demonstrates strong numerical results:

• DataSciBench: DeepAnalyze-8B achieves a success rate of 59.91% and an overall score of 61.11, outperforming most advanced proprietary LLMs (e.g., GPT-4-Turbo, Claude 3.5 Sonnet) and ranking second only to GPT-4o, despite having only 8B parameters.
• DSBench (Data Analysis): Superior performance over workflow-based agents and prompting frameworks.
• DSBench (Data Modeling): Comparable or better performance than AutoGen-based workflows built on proprietary LLMs, with zero inference cost.
• DABStep: Outperforms workflow-based agents, especially on hard-level tasks requiring long-chain reasoning and multi-turn interaction.
• DABStep-Research: Consistently surpasses agent systems built on proprietary LLMs in open-ended data research and report generation, producing analyst-grade reports with high content and format scores.

  Figure 7: Performance on DSBench (data analysis).

  

  Figure 8: Performance on DABStep-Research.

  

  Figure 9: Cases in the constructed DABStep-Research benchmark, including data preparation, data analysis, data insight, report generation, and open-ended data research.

• DS-1000 and TableQA: Outperforms GPT-4-Turbo and GLM-4.5 on code generation and structured data understanding, establishing a robust foundation for autonomous data science.

Ablation and Analysis

Ablation studies confirm the necessity of the $\langle$Understand$\rangle$ action for structured data tasks and the superiority of curriculum-based agentic training over single-stage or non-curriculum approaches. Reasoning trajectory synthesis with keyword-guided refinement further enhances structured data reasoning performance.

Figure 10: Example of reasoning trajectory synthesis.

Autonomous Data Science Case Studies

Qualitative case studies demonstrate DeepAnalyze-8B's ability to autonomously complete the entire data science pipeline, from data preparation and analysis to insight extraction, report generation under constraints, and fully open-ended research.

Figure 11: A data preparation case of autonomous data science, from data sources to analyst-grade research reports.

Figure 12: A data analysis case of autonomous data science, from data sources to analyst-grade research reports.

Figure 13: A data insight case of autonomous data science, from data sources to analyst-grade research reports.

Figure 14: A case of autonomous data science with report constraints.

Figure 15: A case of autonomous data science for fully open-ended data research.

Implications and Future Directions

DeepAnalyze-8B marks a paradigm shift from workflow-based agents to agentic models in data science. The model's ability to autonomously orchestrate and optimize complex pipelines, generate analyst-grade reports, and handle open-ended research tasks has direct implications for the development of next-generation intelligent data systems. The open-sourcing of model, code, and training data facilitates reproducibility and further research.

Potential future directions include scaling agentic training to larger models, extending the framework to multimodal data environments, and integrating with data governance and management systems. The curriculum-based agentic paradigm and data-grounded trajectory synthesis are broadly applicable to other domains requiring autonomous, multi-stage reasoning and interaction.

Conclusion

DeepAnalyze-8B demonstrates that agentic LLMs, trained via curriculum-based agentic RL and data-grounded trajectory synthesis, can achieve autonomous data science at scale. The model sets new standards for open-source LLMs in data-centric tasks, outperforming most proprietary systems and enabling end-to-end automation from raw data to actionable insights. This work paves the way for future research in agentic AI for data ecosystems, with significant practical and theoretical implications for the field.