Source-linked AI summary
DeepAnalyze: Agentic Large Language Models for Autonomous Data Science
Shaolei Zhang, Ju Fan, Meihao Fan, Guoliang Li, Xiaoyong Du
TL;DR
Autonomous data science remains difficult because workflow-based agents rely on predefined procedures and existing systems lack fully autonomous, adaptive behavior. DeepAnalyze-8B trains an agentic model with a human-inspired curriculum and data-grounded trajectory synthesis, and outperforms most advanced proprietary LLMs across 12 benchmarks while supporting open-ended research and analyst-grade reports.
Problem
Existing workflow-based agents depend on predefined workflows, while prior models lack autonomous orchestration and adaptive optimization for fully autonomous data science.
Method
DeepAnalyze combines curriculum-based agentic training with data-grounded trajectory synthesis to progressively learn multiple capabilities in real-world environments.
Results
DeepAnalyze-8B outperforms most advanced proprietary LLMs across 12 data-science benchmarks and performs open-ended data research with analyst-grade report generation.
Takeaways & Limitations
DeepAnalyze advances autonomous data science from predefined workflow agents toward agentic models that can operate across data-centric tasks and open-ended research.
Takeaways & Limitations
Existing workflow-based comparisons remain dependent on manually designed heuristics and domain-specific rules, limiting autonomous and adaptive behavior.
Abstract
from arXiv · showhide
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 large language models (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.
1. Introduction
Autonomous data science requires LLMs to coordinate interdependent data tasks and adapt within real-world environments, but existing approaches remain constrained by limited autonomy and predefined workflows. DeepAnalyze addresses these gaps with curriculum-based agentic training and data-grounded trajectory synthesis, and reports strong performance across data-science benchmarks.
- Motivation: Autonomous data science spans interdependent tasks including preparation, analysis, modeling, visualization, and report generation, with open-ended research increasing the complexity.The paper frames this as a long-standing goal involving extraction of insights from structured data.
- Limitations of prior work: Existing domain-specific models lack autonomous orchestration and adaptive optimization, while workflow-based agents depend on manually designed procedures and rules.These limitations prevent fully autonomous and adaptive behavior in real-world data environments.
- Challenges: Agentic training in data science faces reward sparsity because complex tasks are difficult early in training and trajectory scarcity because long-chain solutions provide little intermediate supervision.Together, these challenges can make exploration inefficient and hinder training.
- Contributions: DeepAnalyze combines a human-inspired easy-to-difficult curriculum with data-grounded trajectory synthesis to progressively acquire capabilities and guide exploration.The approach targets both reward sparsity and scarce long-chain problem-solving trajectories.
- Contributions: DeepAnalyze is designed as an agentic LLM with autonomous orchestration and adaptive optimization for coordinating actions and refining them from environmental feedback.These capabilities are positioned as the higher-level abilities needed for complex data-centric tasks.
- Contributions: DeepAnalyze-8B surpasses most advanced proprietary LLMs on 12 data-science benchmarks while supporting open-ended data research and analyst-grade reports.The model is presented as an agentic system that extends beyond predefined workflows.
2. Related Work
Prior data-science systems target individual tasks or coordinate them through manually designed workflows, rather than internalizing autonomous behavior. DeepAnalyze instead trains an agentic model to acquire complex skills progressively and operate across the data-science pipeline.
- Existing approaches: Domain-specific LLMs address individual tasks, whereas workflow-based agents use predefined procedures to coordinate complex data-science operations.The related-work discussion distinguishes these two main approaches.
- Existing approaches: Both prior approaches remain incapable of fully autonomous data science because their capabilities are limited to task-specific modeling or manually designed workflows.DeepAnalyze is presented as avoiding prompting frameworks and predefined workflows through agentic training.
- Agentic training: Agentic training uses reinforcement learning to develop multi-step reasoning and interaction in real-world environments, with prior practical applications in coding and search.The paper positions this paradigm as the basis for training autonomous models.
- DeepAnalyze: The training setting represents a comprehensive analysis request requiring sequential actions such as loading data, inspecting structure, selecting a model, and generating a report.This example illustrates the multi-stage behavior targeted by the curriculum.
- DeepAnalyze: DeepAnalyze’s curriculum-based agentic training progresses from simpler capabilities toward complex data-science skills to address reward sparsity and trajectory scarcity.The curriculum is motivated by the difficulty of completing complex tasks during early training.
3. DeepAnalyze
DeepAnalyze is an agentic LLM for autonomous data science that interacts with real-world environments through structured actions. Its curriculum-based training and data-grounded trajectory synthesis progressively build the capabilities needed for complex tasks.
- DeepAnalyze: DeepAnalyze is designed to provide autonomous orchestration and adaptive optimization without relying on predefined workflows or prompting frameworks.These capabilities let the model coordinate interdependent actions and iteratively refine them using environmental feedback.
- Architecture: DeepAnalyze uses five actions—Analyze, Understand, Code, Execute, and Answer—to iteratively interact with data environments and produce task responses.Generated code is executed in the environment, and feedback is incorporated before the next action.
- Curriculum-based Agentic Training: Curriculum-based agentic training progresses from single-ability fine-tuning to multi-ability reinforcement learning for increasingly complex data science tasks.The stages target reasoning, structured data understanding, code generation, and their orchestration in real-world environments.
- Hybrid Reward Modeling: The training framework combines rule-based and LLM-judged rewards across data question answering, data-centric tasks, and open-ended research.Reference-answer tasks use accuracy and interaction quality, while open-ended research rewards report quality and successful environment interactions.
- Data-grounded Trajectory Synthesis: Data-grounded trajectory synthesis addresses scarce training trajectories by combining reasoning distillation and refinement for structured-data tasks.Teacher-model reasoning is verified against ground-truth responses and reformulated into Analyze and Understand components.
4. Experiments
Experiments evaluate DeepAnalyze-8B across benchmarks spanning the end-to-end data science pipeline, individual analytical tasks, structured-data abilities, and data-oriented deep research. Across these evaluations, the 8B model achieves strong results against workflow-based agents and proprietary LLM systems, including fully open-ended research and analyst-grade report generation.
- Capability on End-to-end Data Science Pipeline: DeepAnalyze-8B ranks second only to GPT-4o on DataSciBench while outperforming most compared proprietary and open-source workflow-based agents without external ReAct orchestration.DataSciBench measures coarse- and fine-grained performance across preparation, analysis, modeling, and visualization stages.
- Capability on Individual Data Science Tasks: On DSBench, DeepAnalyze-8B outperforms prompting and workflow-based agents for statistical analysis and performs comparably to AutoGen-based proprietary-LLM workflows for data modeling.The data-analysis comparison is shown in Figure 7, while data-modeling results are reported in Table 2.
- Capability on Individual Data Science Tasks: DeepAnalyze-8B outperforms previous workflow-based agents on DABStep, particularly on hard-level tasks requiring long-chain reasoning across markdown, CSV, and JSON data.The reported advantage is attributed to iterative environment interaction through autonomous orchestration and adaptive optimization.
- Capability on Data-Oriented Deep Research: DeepAnalyze-8B consistently outperforms all compared systems across DABStep-Research tasks, including fully open-ended data research and report-format quality.The benchmark covers preparation, analysis, insight, specified-outline reports, and unconstrained research.
- Capability Related to Data Science: DeepAnalyze-8B outperforms GPT-4-Turbo and GLM-4.5 on DS-1000 and surpasses Reasoning-Table on TableQA, with gains strengthened by agentic training.These evaluations target code generation and structured-data understanding that support complex data science.
5. Analysis
The analysis studies DeepAnalyze’s action design, curriculum-based agentic training, and reasoning-trajectory synthesis. Results support both the structured-data understanding action and a staged progression from single-ability tuning to multi-ability agentic training.
- Action Ablation: Removing the ⟨Understand⟩ action reduces performance on WikiTQ, MultiHiertt, and DABStep, linking structured-data understanding with both table and analysis tasks.The ablation isolates ⟨Understand⟩ from the reasoning process represented by ⟨Analyze⟩.
- Curriculum-based Agentic Training: Curriculum-based agentic training outperforms one-stage training because single-ability preparation and multi-ability interaction are scheduled from simple to complex tasks.Only single-ability fine-tuning fails on complex multi-turn DABStep tasks, while only multi-ability agentic training struggles when foundational abilities are weak.
- Reasoning Trajectory Synthesis: Distillation and refinement of reasoning trajectories both improve structured-data understanding compared with training on original Reasoning-Table data.The comparison evaluates original, distilled, and refined training data.
6. Conclusion and Future Work
DeepAnalyze demonstrates broad autonomous data-science capabilities and strong benchmark performance with an 8B-parameter model. It extends beyond predefined workflows to open-ended research and analyst-grade reporting.
- DeepAnalyze-8B outperforms state-of-the-art closed-source LLMs on 12 data science benchmarks.
- DeepAnalyze combines curriculum-based agentic training with data-grounded trajectory synthesis.
- DeepAnalyze enables open-ended data research and generates analyst-grade reports beyond predefined workflows.
- The work advances autonomous extraction of actionable insights from raw data toward agentic models for intelligent data systems.
A. Construction of DABStep-Research Benchmark
DABStep-Research is a 100-task benchmark for data-oriented deep research across five categories. Its evaluation uses an LLM judge to score generated reports on content and format.
- The benchmark addresses the need to evaluate data-oriented deep research and research-report generation beyond specific data-science tasks.
- DABStep-Research contains 100 tasks spanning data preparation, data analysis, data insight, report generation, and open-ended data research.
- Report-generation tasks specify formats such as titles, outlines, and other detailed requirements.
- An LLM judge assigns report scores from 1 to 5 for content and format using the instruction, checklist, and generated report.
- The evaluation rubric defines content as helpfulness, comprehensiveness, and relevance to the task goal.
- The format criterion evaluates whether reports use a polished academic style.
B. Keyword-guided Reasoning Trajectory Synthesis
Keyword-guided reasoning trajectory synthesis refines teacher-model traces to focus more strongly on structured data. The method addresses the tendency of general-purpose teacher models to overlook supplied data.
- Teacher-model reasoning trajectories can overlook supplied data because the models are not specifically trained for data-science domains such as structured data understanding.
- Keyword-guided refinement inserts reasoning keywords that direct trajectories toward structured data understanding.
- The method samples keywords about boundaries, prior reasoning, and closer table examination before asking the teacher model to refine its trajectory.
- The refined trajectories show stronger repeated examination and reflection on structured data.
C. Cases
The presented cases cover DeepAnalyze’s autonomous data-science pipeline from data sources to analyst-grade research reports. They span preparation, analysis, insight extraction, constrained reporting, and open-ended research.
- The cases cover the full pipeline from data sources to analyst-grade research reports.
- The examples include data preparation, data analysis, data insight extraction, report generation under constraints, and fully open-ended data research.
- Compared with previous closed-source LLMs and tool-calling frameworks, DeepAnalyze produces higher-quality analyst-level reports.
- Figure 10 presents an example of reasoning trajectory synthesis.
- Figures 11–15 present cases covering preparation, analysis, insight, constrained reporting, and open-ended research.