This interactive demo showcases a novel zero human knowledge approach for solving combinatorial puzzles. Here, using a compact neural network (just 100k parameters MLP, trained in 5 minutes), it guides beam search (fast/balanced/quality: beam width = 2¹⁰/2¹²/2¹⁴) to find solutions averaging 26/24/22 moves in the QTM (graph diameter is 26).
This is the first AI method to efficiently scale to larger puzzles like 4×4×4 and 5×5×5 cubes. Unlike traditional solvers that rely on human-designed heuristics, this approach learns purely from graph random walks. Thanks to the efficiency of this approach, all computation runs entirely in your browser.