Modern Architecture On Advanced LLM: Best-of-N Sampling, Learned Verification and Tree Search as a Substitute for Parameter Scale

The prevailing approach to capability in large language models couples performance to parameter count. We present an alternative architecture in which a frozen 1.5B-parameter reasoning model is paired with a structured inference-time compute layer, and show that the combination recovers a substantial fraction of the capability normally attributed to scale. On AIME 2024, the frozen base improves from 23.3% single-sample accuracy to 83.3% coverage at N=128; a trained outcome verifier converts that coverage into 52.2% realised accuracy at N=32, a 16.6-point gain over majority voting. We formalise the generation-selection decomposition, identify the selection gap as the governing bottleneck of such systems, report transfer to graduate-level science and arithmetic reasoning, and derive Best-of-N MCTS, a tree-search generalisation in which the outcome verifier supplies process-level guidance through truncated rollout. All results are reproducible from a public repository.