The research paper Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification establishes a new paradigm for improving large language model reasoning performance through inference-time scaling, rather than model retraining. By systematically sampling multiple candidate outputs and applying self-verification, the authors demonstrate that reasoning accuracy grows predictably as the number of samples and verification quality increase. This approach, termed sampling-based search, elevates existing models like Gemini v1.5 Pro beyond o1-Preview performance on reasoning benchmarks such as AIME and LiveBench. The study formalizes implicit scaling, where increasing sample diversity also enhances verification accuracy, and introduces benchmarks that expose the current weaknesses of frontier models in verifying their own reasoning.

_____

Key Point: This paper positions that scaling inference-time sampling and self-verification, not model size, emerges as a powerful, cost-efficient pathway to achieving state-of-the-art reasoning performance in large language models.