Code World Model releases weights and a 40-example reproduction set for treating a coding agent as a world model

GitHub·medium signalXBlueskyLinkedInCopy link

buaacyw/code-world-model went public 2026-08-26 (197 stars, paper at arXiv:2608.25927) from AGI Lab at Westlake University and Nanyang Technological University. It ships a public LoRA checkpoint on Hugging Face plus 40 compact examples that expand deterministically via `install_release_assets.sh`, and the inference CLI splits into explicit validate, prepare and generate stages against a frozen config. The reproduction discipline is the notable part: released configs run without changing inputs, prompts or seed, and every changed experiment requires a fresh cache directory.