Skip to main content
Back

HKUST's EvoDS Beats Top Data Science Agents by 28.9% With Self-Evolving Memory

EvoDS is an open-source data science agent that learns new skills on the fly and manages its own context window, beating existing agents by 28.9% across four benchmarks

Paper, Full Access
  • EvoDS is a self-evolving data science agent that learns reusable skills and manages its own context window via reinforcement learning.
  • It outperforms state-of-the-art open-source data science agents by an average of 28.9% across four benchmarks including InfiAgent-DABench and DSBench.
  • The Autonomous Skill Acquisition (ASA) mechanism lets the agent synthesize, validate, and reuse executable code skills across tasks.
  • The Adaptive Context Compression (ACC) strategy treats context management as a learned RL problem, eliminating out-of-token failures entirely.
  • The optimization objective is theoretically grounded in the information bottleneck principle, ensuring only task-relevant context is retained.
  • Code and data are open-source at github.com/usail-hkust/EvoDS.

Automated data science agents have been a hot area of LLM research, with systems like AIDE, DataInterpreter, and OpenHands all competing to replace the tedious parts of the ML pipeline. But they all share a quiet flaw: they start every task from scratch, carrying a fixed set of tools and no memory of what worked before. The longer a task runs, the more context accumulates, and eventually the model either hallucinates or crashes with an out-of-token error. A new paper from HKUST introduces EvoDS, a self-evolving agent that directly attacks both problems at once.

Why existing agents keep hitting a wall

Real data science is not a single-step problem. A typical pipeline involves exploring a dataset, engineering features, selecting models, tuning hyperparameters, debugging failures, and iterating. These systems combine LLM reasoning with structured pipelines but still face challenges in long-term dependencies, code quality, and generalization. The deeper issue is structural: most agents operate with a static action set, meaning the tools they can call are fixed at design time. If a useful pattern emerges during a task, the agent cannot save it for later.

The second problem is context. As interactions accumulate, contexts grow unbounded, creating two major challenges. First, the inference cost of transformers scales with the context length, which becomes prohibitive in long-horizon tasks. Most agents handle this by simply truncating old content, which is a passive, lossy strategy that can drop exactly the information the agent needs to finish the job. Evaluation of state-of-the-art LLMs and agents shows that they struggle with most tasks, with the best agent solving only 34.12% of data analysis tasks.

Two ideas that change the game

EvoDS tackles both problems with two complementary mechanisms trained end-to-end using reinforcement learning.