AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI that Learns on the Job: Meet the Self-Evolving Data Science Agent

In the rapidly accelerating field of artificial intelligence, Large Language Model (LLM) agents have shown remarkable promise in automating complex coding and data analysis tasks. Yet, today’s AI agents suffer from a fundamental limitation: they have a terrible short-term memory and cannot learn from their own experiences. When given a new dataset, an agent tackles it from scratch, using a static, pre-defined set of tools. Once the task is completed, any custom code or novel problem-solving strategy it discovered is discarded.

Now, researchers from the Hong Kong University of Science and Technology (Guangzhou) have unveiled EvoDS, a self-evolving autonomous data science agent designed to break this cycle. Unlike its predecessors, EvoDS can dynamically expand its library of skills and intelligently compress its own memory, allowing it to grow smarter with every problem it solves.

Learning From Experience

To understand how EvoDS works, imagine an AI tasked with analyzing a medical database to predict stroke risk. Because stroke cases are rare—representing only about 4.5% of the data—standard machine learning algorithms struggle to make accurate predictions. This is a classic “class imbalance” problem.

Faced with this challenge, a typical AI agent might try to write custom code, but it would forget the solution once the project ended. EvoDS, however, features an Autonomous Skill Acquisition (ASA) mechanism. Recognizing that its pre-packaged tools are insufficient, EvoDS writes code to handle the imbalance, tests and validates it, and saves it in a repository as a new tool: imbalanced_binary_classification.

When the agent is later tasked with predicting default rates on credit cards—another highly imbalanced dataset—it does not waste time rediscovering the solution. Instead, it retrieves and reuses the imbalanced_binary_classification tool it created earlier, drastically reducing execution time and computing costs.

A Manager with an Active Memory

A second major hurdle for data science agents is “context explosion.” Iterative processes generate vast amounts of code snippets, database previews, and error logs, quickly overwhelming the LLM’s limited memory window.

EvoDS solves this by combining a hierarchical multi-agent architecture with Adaptive Context Compression (ACC). At the top of the hierarchy is a “Manager Agent” that delegates specific tasks to specialized sub-agents, such as a Cleaner, Modeler, or Visualizer.

Rather than passively waiting for its memory to fill up and then blindly truncating older messages, the Manager Agent acts like an efficient project lead. It dynamically decides when to summarize raw technical logs into concise progress updates, keeping only the task-critical insights while discarding the noise.

Setting New Benchmarks

To train EvoDS, the researchers used a two-stage reinforcement learning framework that teaches the agent how to coordinate tasks, build tools, and manage memory simultaneously.

The results are striking. Tested across four rigorous data science benchmarks (including DA-Code and MLE-Dojo), EvoDS outperformed existing state-of-the-art open-source agents by an average of 28.9%. Crucially, while other agents frequently crashed due to out-of-token memory failures on long-horizon tasks, EvoDS completed its workflows without a single memory-related failure.

By bridging the gap between passive execution and active learning, EvoDS represents a significant step toward truly autonomous AI colleagues capable of lifelong learning on the job.