EvoDS agent learns skills and manages context for automated data science
EvoDS is a self-evolving autonomous data science agent that uses agentic reinforcement learning to dynamically acquire reusable skills and adaptively compress long-term context, outperforming state-of-the-art open-source agents by an average of 28.9% across four benchmarks.
Score breakdown
EvoDS directly addresses two core failure modes of current LLM-based data science automation — static skill sets and context overflow — with a system that learns to expand its own capabilities and manage long-horizon context, achieving a 28.9% average improvement over existing open-source agents across four benchmarks.
- 01EvoDS is a self-evolving autonomous data science agent built on agentic reinforcement learning.
- 02Autonomous Skill Acquisition (ASA) enables the agent to synthesize, validate, and reuse executable skills across tasks.
- 03Adaptive Context Compression (ACC) treats context management as a learned control problem rather than passive truncation.
Zherui Yang, Fan Liu, and Yansong Ning present EvoDS, a self-evolving autonomous data science agent that addresses two fundamental limitations of current LLM-based data science automation: static action sets that prevent experience accumulation across tasks, and the absence of principled long-horizon context management that causes failures in multi-stage, iterative pipelines.
EvoDS introduces two key mechanisms trained via agentic reinforcement learning.
EvoDS introduces two key mechanisms trained via agentic reinforcement learning. The Autonomous Skill Acquisition (ASA) mechanism enables the agent to synthesize, validate, and reuse executable skills, allowing it to build a growing library of reusable capabilities over time. The Adaptive Context Compression (ACC) strategy reframes context management as a learned control problem rather than passive truncation, enabling the agent to intelligently decide what context to retain across long task horizons. Both strategies are orchestrated within a two-stage multi-agent training scheme.
The authors provide theoretical grounding for EvoDS's design: they prove that its hierarchical structure reduces tool-selection error, and that its optimization objective aligns with an information bottleneck principle, ensuring efficient context use. Empirically, EvoDS outperforms state-of-the-art open-source data science agents by an average of 28.9% across four diverse benchmarks, while also eliminating out-of-token failures that plague existing approaches. Code and data are publicly available on GitHub.
Key facts
- 01EvoDS is a self-evolving autonomous data science agent built on agentic reinforcement learning.
- 02Autonomous Skill Acquisition (ASA) enables the agent to synthesize, validate, and reuse executable skills across tasks.
- 03Adaptive Context Compression (ACC) treats context management as a learned control problem rather than passive truncation.
- 04A two-stage multi-agent training scheme orchestrates both ASA and ACC strategies.
- 05EvoDS outperforms state-of-the-art open-source data science agents by an average of 28.9% across four diverse benchmarks.
- 06EvoDS eliminates out-of-token failures that affect existing approaches.
- 07The authors prove theoretically that EvoDS's hierarchical design reduces tool-selection error and aligns with an information bottleneck principle.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →