EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Management
TL;DR. EvoDS is a self-evolving data-science agent that fixes two limits of prior systems: static action sets and unprincipled long-horizon context handling. It adds Autonomous Skill Acquisition (synthesize, validate, and reuse executable skills) and Adaptive Context Compression (treat context management as a learned control problem, not passive truncation), orchestrated in a two-stage multi-agent training scheme via agentic RL. It proves the hierarchical design reduces tool-selection error and that its objective aligns with an information-bottleneck principle. Empirically it beats open-source data-science agents by 28.9% average across four benchmarks while eliminating out-of-token failures.
Source: HuggingFace Daily Papers (upvotes: 2) arxiv: 2606.03841 · Code: https://github.com/usail-hkust/EvoDS Raw: raw/huggingface/2026-06-05-evods-self-evolving-autonomous-data-science-agent-with-skill.md
Key points
- Autonomous Skill Acquisition (ASA): the agent synthesizes, validates, and reuses executable skills, replacing a fixed action set with a growing one.
- Adaptive Context Compression (ACC): context management is a learned control problem rather than passive truncation, the same "learn the memory instead of scheduling it" instinct as Echo-Infinity (06-04) and MemTrain (06-04), here applied to a data-science agent.
- Theory: hierarchical design provably reduces tool-selection error; the optimization objective aligns with an information-bottleneck principle (compress context while preserving task-relevant information).
- Results: +28.9% average over SOTA open data-science agents across four benchmarks; eliminates out-of-token failures.
Relation to prior wiki
Part of today's self-evolving-agents cluster (Continual Experience Internalization is the keystone). EvoDS learns skills via RL; the keystone paper's finding that on-policy internalization is less stable than off-policy is the caution to read alongside EvoDS's agentic-RL skill acquisition. Its learned context compression is the agent-memory analogue of the learned-memory trend the wiki logged in video (Echo-Infinity) and self-supervised memory (MemTrain). DataCOPE (also today, unsupervised verifier-guided skill discovery for data analysis, +32.3% on reasoning-style tasks) is the sibling result on the same task family.