About This Research

The Paper

This blog accompanies our paper "Debugging code world models" (arXiv:2602.07672), which studies Code World Models (CWMs) through two complementary perspectives: local semantic execution and long-horizon state tracking.

Key Contributions

• We characterize two dominant failure regimes: token-budget exhaustion and string-valued state brittleness
• We show long-horizon degradation is dominated by incorrect action generation (action hallucination)
• We show teacher forcing isolates state propagation and yields strong long-horizon accuracy
• We connect string failures to subword tokenization instability rather than program structure

Resources

• Debugging code world models (arXiv:2602.07672)
• Learning State-Tracking from Code Using Linear RNNs (arXiv:2602.14814)
• Code Repository
• Dataset

Contact

For questions about this research, please contact the authors.