一个模型统一所有离线任务！微软用671B大模型重构广告推荐「推理大脑」

Core Insights - Microsoft has developed a unified offline reasoning hub called AdNanny, built on the DeepSeek-R1 model with 671 billion parameters, which significantly outperforms previous models in managing advertising systems [4][8][20] Group 1: Paradigm Shift - The advertising recommendation system is transitioning from a "model forest" approach, which involves numerous small models for different tasks, to a centralized intelligent hub that enhances performance and reduces costs [6][8] - The "model forest" approach leads to knowledge silos, high operational costs, and black-box decision-making processes, making it inefficient [6][7] Group 2: Data Transformation - AdNanny's strength lies in its innovative data processing, transforming raw advertising data into high-quality reasoning-enhanced corpora through a three-stage automated data factory [9] - The process includes reasoning generation, validation by human experts, and rejection sampling to ensure the model learns correct causal relationships [9][10] Group 3: Training Mechanisms - AdNanny employs dynamic re-weighting to focus on challenging tasks and samples, ensuring that less frequent tasks receive adequate attention during training [11][12] - Reinforcement learning is integrated to align model outputs with business metrics, ensuring that generated reasoning contributes positively to ad clicks and conversions [13] Group 4: Engineering Innovations - The model utilizes a mixed parallel architecture for efficient training, achieving high computational resource utilization across 248 GPUs [14][15] - AdNanny's inference optimization through FP8 quantization has reduced offline computational costs by approximately 50% compared to multiple smaller models [17] Group 5: Practical Applications - AdNanny has demonstrated superior performance in key tasks such as query-ad relevance and ad-user matching, significantly lowering costs and simplifying system architecture [18][19] - The model's ability to provide reliable initial assessments for ambiguous samples reduces the need for extensive manual labeling, streamlining the workflow [18] Conclusion - AdNanny represents a significant advancement in industrial AI, moving beyond mere computational power to a deeper logical framework that could influence various sectors beyond advertising, such as search, e-commerce, and financial decision-making [20]