Title: EvoDS: A Self-Improving Autonomous Data Science Agent Integrating Skill Learning and Context Management

Abstract:

While recent breakthroughs in Large Language Model (LLM) agents have driven significant progress in automated data science, current methodologies face inherent constraints. Specifically, they are hampered by fixed action sets and an absence of rigorous long-horizon context management, which prevents agents from accumulating reusable experience across different tasks and ensures unreliable performance in complex, iterative, multi-stage data science workflows.

To overcome these limitations, we present EvoDS, an autonomous data science agent capable of self-evolution. EvoDS utilizes agentic reinforcement learning to dynamically expand its skill repertoire and manage long-term context adaptively. The system is built upon two core innovations:

1. Autonomous Skill Acquisition (ASA): This mechanism allows the agent to independently synthesize, verify, and repurpose executable skills.
2. Adaptive Context Compression (ACC): Rather than relying on passive truncation, this strategy frames context management as a learned control problem.

These components are coordinated through a two-stage multi-agent training framework, facilitating continuous autonomous improvement. From a theoretical standpoint, we demonstrate that EvoDS’s hierarchical architecture minimizes tool-selection errors. Furthermore, its optimization objective adheres to an information bottleneck principle, thereby guaranteeing efficient utilization of context.

Empirical evaluations reveal that EvoDS surpasses leading open-source data science agents by an average margin of 28.9% across four varied benchmarks. Additionally, the system successfully eradicates out-of-token failures. The source code and associated data can be accessed at https://github.com/usail-hkust/EvoDS.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC