DSWorld: A Data Science World Model for Efficient Autonomous Agents

arXiv 2026 43.6 method, benchmark, system, application

TLDR

Introduces DSWorld, a data science world model for predicting execution outcomes, accelerating RL agent training and search-based inference.

Reasoning

The paper presents a novel framework (DSWorld) that models data science environment transitions, combining structured state construction, cost-aware routing, and an LLM simulator. Strengths include a clear problem motivation, a new dataset, and strong empirical results (14x training speedup, 35.6% improvement over LLM baseline). Weaknesses: the abstract lacks details on generalization or limitations of the simulator, and the 'world model' terminology is specific to data science rather than general domains.

Read-first score

Read-first score 43.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,dataset,experiment

Topical relevance 29%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 396.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

generative world model

video world model

Deep Analysis

Innovations

Introduction of the Data Science World Model concept for predicting environment state transitions in data science workflows
DSWorld framework combining structured state construction, cost-aware routing, lightweight real execution, and an LLM-based simulator for expensive operations
Construction of an 8K-scale transition trajectory dataset for training
Reflective World Model Optimization, an error-aware reinforcement learning strategy for transition prediction improvement

Methodology

DSWorld models the data science execution environment by predicting state transitions given current workflow states and candidate operations. The framework uses structured state construction, cost-aware routing to decide between lightweight real execution and LLM-based simulation, and is trained on an 8K trajectory dataset with an error-aware RL strategy (Reflective World Model Optimization).

Key Results

DSWorld accelerates RL-based agent training by approximately 14x and search-based inference by 3-6x while maintaining competitive performance, and outperforms the strongest LLM baseline by 35.6% on transition prediction tasks.

Resources

Paper