Meta AI Introduces DeepConf: First AI Method to Achieve 99.9% on AIME 2025 with Open-Source Models Using GPT-OSS-120B

Meta AI and UCSD researchers have developed DeepThink with Confidence (DeepConf), a novel approach that significantly enhances the efficiency of reasoning in large language models (LLMs) by leveraging the models' own confidence signals. Unlike traditional parallel thinking methods, which generate multiple reasoning paths at high computational costs and often face diminishing accuracy returns, DeepConf achieves near state-of-the-art performancesuch as 99.9% accuracy on the AIME 2025 math competitionwhile reducing token generation by up to 85%, making the process more resource-efficient. This innovation addresses the core trade-off in LLM