DSWorld: A Data Science World Model for Efficient Autonomous Agents
DSWorld tries to make data-science agents faster by simulating costly workflow steps before running them.
The paper defines a “Data Science World Model” that predicts state changes from a current workflow and a candidate operation. DSWorld combines structured state construction, cost-aware routing, lightweight real execution, and an LLM simulator for expensive operations. The authors report about 14x faster RL-based agent training and 3-6x faster search-based inference while keeping competitive performance. They also say it beats the strongest LLM baseline by 35.6% on transition prediction. Source: HF Daily Papers' note.
The paper defines a “Data Science World Model” that predicts state changes from a current workflow and a candidate operation. DSWorld combines structured state construction, cost-aware routing, lightweight real execution, and an LLM simulator for expensive operations. The authors report about 14x faster RL-based agent training and 3-6x faster search-based inference while keeping competitive performance. They also say it beats the strongest LLM baseline by 35.6% on transition prediction. Source: HF Daily Papers' note.
score 5