CWM: An Open-Weights LLM for Research on Code Generation with World Models
14 hours ago
- #AI
- #Code Generation
- #Machine Learning
- Code World Model (CWM) is a 32-billion-parameter open-weights LLM designed for code generation research.
- CWM is mid-trained on Python interpreter and Docker agentic environments to enhance code understanding beyond static code training.
- The model undergoes multi-task reasoning RL in verifiable coding, math, and software engineering environments.
- CWM supports a 131k token context size and shows strong performance in coding and math tasks.
- Performance metrics include 65.8% on SWE-bench Verified, 68.6% on LiveCodeBench, 96.6% on Math-500, and 76.0% on AIME 2024.
- Model checkpoints are released post mid-training, SFT, and RL to support further research on code world modeling.
- CWM aims to improve agentic coding, simulate Python code execution step-by-step, and enhance reasoning capabilities.