Guardado en:
| Autores principales: | , , , , , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | https://arxiv.org/abs/2509.09245 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866915650207744000 |
|---|---|
| author | Li, Shuocheng Liu, Yihao Du, Silin Zeng, Wenxuan Xu, Zhe Zhou, Mengyu He, Yeye Dong, Haoyu Han, Shi Zhang, Dongmei |
| author_facet | Li, Shuocheng Liu, Yihao Du, Silin Zeng, Wenxuan Xu, Zhe Zhou, Mengyu He, Yeye Dong, Haoyu Han, Shi Zhang, Dongmei |
| contents | Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks. Code and data are available at https://github.com/microsoft/Jupiter. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_09245 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search Li, Shuocheng Liu, Yihao Du, Silin Zeng, Wenxuan Xu, Zhe Zhou, Mengyu He, Yeye Dong, Haoyu Han, Shi Zhang, Dongmei Artificial Intelligence Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks. Code and data are available at https://github.com/microsoft/Jupiter. |
| title | Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2509.09245 |