CVRB tests how well large language models can reason and discover patterns in a two-phase self-play loop: first, Creator models invent miniature deterministic worlds; then Solver models try to answer questions about those worlds. All answers are verified automatically by running the code, not by humans - View it on GitHub