EvoDS

Model Description

This model is the backbone language model used in the paper "EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management". It is initialized from Qwen3-8B and further trained to support autonomous data science tasks through multi-agent supervised fine-tuning and reinforcement learning.

Base Model

• Base Model: Qwen3-8B

About EvoDS

EvoDS is a self-evolving autonomous data science agent that continuously improves its capabilities over time. The framework introduces two key mechanisms:

Autonomous Skill Acquisition (ASA)

ASA enables the agent to autonomously synthesize, validate, cache, and reuse executable skills, allowing the agent's action space to expand over time.

Adaptive Context Compression (ACC)

ACC enables efficient long-horizon reasoning by dynamically compressing interaction history and preserving critical information under limited context budgets.

Resources

• Paper: EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management
• Code Repository: https://github.com/usail-hkust/EvoDS

Citation

@article{yang2026evods,
  title={EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management},
  author={Yang, Zherui and Liu, Fan and Ning, Yansong and Liu, Hao},
  journal={arXiv preprint arXiv:2606.03841},
  year={2026}
}