DSWorld: A World Model for Accelerating Data Science Agents
HKUSTGZ · hf · 2026-07-20
Current autonomous data science agents rely heavily on trial-and-error mechanisms, incurring massive computational costs. To solve this, the study proposes the concept of a Data Science World Model, which predicts state transitions from candidate actions to pre-evaluate execution outcomes.
The core framework, DSWorld, combines structured state construction, cost-aware routing, lightweight real execution, and an LLM for simulating expensive operations. Key highlights include:
- Training Optimization: Constructed an 8K-scale state transition dataset and proposed an error-aware reinforcement learning strategy.
- Significant Speedup: Accelerates agent training by roughly 14x and search inference by 3-6x, while outperforming the strongest LLM baseline by 35.6% on state prediction tasks.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21