GoKawiilBenchmark tests LLMs on Zachtronics' TIS-100 puzzle game, finds 133 new optimal solutions
A new evaluation ran multiple language models through 64 TIS-100 puzzles—covering the campaign, TIS-NET set, and 2018 tournament puzzles—giving each model one timed attempt per puzzle under a pass@1 scoring scheme. Comparing model outputs against the community leaderboard as of October 15, 2026, the project identified 154 candidate record-beating solutions, verified 133 of them across 18 puzzles as legitimate new frontier points, and published all resulting programs publicly under MIT License.