⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

CWM: An Open-Weights LLM for Research on Code Generation with World Models

arXiv 25.9 2025 64.3 method, system

TLDR

CWM is a 32B open-weights LLM mid-trained on execution trajectories to improve code generation via world models, achieving strong coding and math benchmarks.

Reasoning

The paper's strength lies in releasing a large open-weights model with novel mid-training on execution trajectories and strong benchmark results. Weaknesses include limited details on world model evaluation beyond early results and lack of explicit comparison to other world model approaches.

Read-first score

Read-first score 64.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%
80

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

Topical relevance 42%
61.4

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 25%
46

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 165.

Keyword Scores

world model
10
interactive world model
8
generative world model
7
world dynamics prediction
7
world simulator
6
model-based reinforcement learning world model
5
video world model
0

Deep Analysis

Innovations

  • Mid-training on observation-action trajectories from Python interpreter and agentic Docker environments to improve code understanding beyond static code
  • Multi-task reasoning reinforcement learning in verifiable coding, math, and multi-turn software engineering environments
  • Providing an open-weights testbed for research on world modeling for code generation with reasoning and planning
  • Step-by-step simulation of Python code execution using world models
  • Release of model checkpoints after mid-training, SFT, and RL to support further research

Methodology

CWM is a 32-billion-parameter dense decoder-only LLM trained with a context size of up to 131k tokens. It undergoes mid-training on observation-action trajectories from Python interpreter and agentic Docker environments, followed by extensive multi-task reasoning RL in verifiable coding, math, and multi-turn software engineering environments. Evaluation uses pass@1 metrics on SWE-bench Verified (with test-time scaling), LiveCodeBench, Math-500, and AIME 2024.

Key Results

CWM achieves pass@1 scores of 65.8% on SWE-bench Verified, 68.6% on LiveCodeBench, 96.6% on Math-500, and 76.0% on AIME 2024, demonstrating strong performance on general coding and math tasks independent of its world modeling capabilities.

Limitations

  • World modeling capabilities are presented as first steps and early results, indicating limited maturity and scope
  • The benefits of world modeling for agentic coding and reasoning are only preliminarily demonstrated

Resources

Paper

Tags