Alex Wang describes reasoning as pathfinding through token space: reinforcement learning raises the total probability of paths reaching correct answers rather than guaranteeing convergence of an individual chain. With an accurate verifier and independent samples, best-of-k improves the chance of success, but this differs from lengthening one path. The related s1 study found benefits from extra thinking on limited math evaluations, not a guarantee of unlimited improvement. Rewarding only final answers does not ensure faithful explanations, and a faulty verifier can reinforce reward hacking.