TetrisBench: How Language Models Develop Playing Strategies Through Code Generation

Yoko Li built TetrisBench, a benchmark where LLMs compete at Tetris by generating scoring functions rather than making move-by-move decisions. The experiment reveals distinct playing styles across models, with some prioritizing aggressive early strategies while others favor conservative, long-term approaches. Results show that strategy intervention frequency and long-horizon optimization differ significantly among models, even when performance metrics are similar.