The rapid progress of large language models (LLMs) has been driven primarily by scaling model size, data, and compute. While this paradigm has enabled remarkable advances in reasoning, it is increasingly constrained by physical, economic, and data limitations. At the same time, evaluating and improving reasoning systems has itself become a bottleneck: as models approach or surpass human-level performance, reliable assessment and supervision become increasingly costly, ambiguous, and fragile. This thesis investigates a complementary paradigm for advancing LLM reasoning that shifts emphasis from scaling models to scaling