TetrisBench: LLMs Conquer Tetris, Differently
Yoko Li's TetrisBench project reveals how LLMs, initially struggling with direct play, develop surprising, distinct strategies when tasked with generating game logic, outperforming most humans but faltering against top players' adaptive chaos.
Feb 24 at 8:28 PM3 min read
Key Takeaways
- 1LLMs initially failed at direct Tetris play, succeeding only when reframed as a task to generate game logic.
- 2Models developed distinct, unprompted strategies, revealing individual optimization horizons and intervention patterns.
- 3Top human players exploit 'controlled chaos,' exposing LLMs' brittle optimization in unusual game states.
