AI Codebreakers Conquer the Legendary Assembly Programming Game TIS-100

A new research project has successfully tested large language models on TIS-100, Zachtronics' notoriously difficult assembly programming puzzle game that requires writing efficient code for a fictional 1980s processor. The models tackled all 22 campaign puzzles, 25 TIS-NET puzzles, and tournament puzzles, with scoring based on pass@1 success rates where a puzzle only counts as solved if the model's single saved program replay passes in a fresh game.

The research provides concrete evidence that LLMs can generate functional, efficient assembly code for complex constraint-based challenges, with models finding 154 candidate solutions across 23 puzzles. This demonstrates significant progress in AI's ability to understand and manipulate low-level programming concepts, potentially paving the way for automated optimization tools for specialized hardware and embedded systems programming assistance.