Stanford AI Lab and Berkeley Sky Computing Lab, in collaboration with NVIDIA, have unveiled a new approach called LLM-as-a-Verifier to improve the accuracy of AI programming agents. This method addresses the challenge of selecting the best solution from multiple attempts by analyzing the model's probability distribution across scoring levels, rather than relying solely on a judge's final score. The Verifier also evaluates tasks across three dimensions: task requirement fulfillment, output format correctness, and error signal presence. In experiments, the Verifier demonstrated superior performance, achieving a single-run accuracy of 74.7% compared to 57.0% for traditional methods. After 16 repetitions, accuracy increased to 77.4%, surpassing the judge's 70.2%. The Verifier also eliminated ties in solution comparisons, a common issue with traditional judges. Practical applications on Terminal-Bench 2 and SWE-Bench Verified showed significant improvements in success rates, with the Verifier achieving top rankings since its release on April 9. The framework has been open-sourced for broader use.