This paper presents a case study on deploying Large Language Models (LLMs) as an advanced "annotation" mechanism to achieve nuanced content understanding (e.g., discerning content "vibe") at scale within a large-scale industrial short-form video recommendation system. Traditional machine learning classifiers for content understanding face protracted development cycles and a lack of deep, nuanced comprehension. The "LLM-as-annotators" approach addresses these by significantly shortening development times and enabling the annotation of subtle attributes. This work details an end-to-end workflow encompassing: (1) iterative definition and robust evaluation of target attributes, refined by offline metrics and online A/B testing; (2) scalable offline bulk annotation of video corpora using LLMs with multimodal features, optimized inference, and knowledge distillation for broad application; and (3) integration of these rich annotations into the online recommendation serving system, for example, through personalized restrict retrieval. Experimental results demonstrate the efficacy of this approach, with LLMs outperforming human raters in offline annotation quality for nuanced attributes and yielding significant improvements of user participation and satisfied consumption in online A/B tests. The study provides insights into designing and scaling production-level LLM pipelines for rich content evaluation, highlighting the adaptability and benefits of LLM-generated nuanced understanding for enhancing content discovery, user satisfaction, and the overall effectiveness of modern recommendation systems.

It is increasingly common in digital environments to use A/B tests to compare the performance of recommendation algorithms. However, such experiments often violate the stable unit treatment value assumption (SUTVA), particularly SUTVA's ''no hidden treatments'' assumption, due to the shared data between algorithms being compared. This results in a novel form of bias, which we term ''symbiosis bias,'' where the performance of each algorithm is influenced by the training data generated by its competitor. In this paper, we investigate three experimental designs--cluster-randomized, data-diverted, and user-corpus co-diverted experiments--aimed at mitigating symbiosis bias. We present a theoretical model of symbiosis bias and simulate the impact of each design in dynamic recommendation environments. Our results show that while each design reduces symbiosis bias to some extent, they also introduce new challenges, such as reduced training data in data-diverted experiments. We further validate the existence of symbiosis bias using data from a large-scale A/B test conducted on a global recommender system, demonstrating that symbiosis bias affects treatment effect estimates in the field. Our findings provide actionable insights for researchers and practitioners seeking to design experiments that accurately capture algorithmic performance without bias in treatment effect estimates introduced by shared data.

With the rapid growth scale of dataset and model, the training of deep neural networks (DNN) tends to be deployed in a distributed manner. In the large-scale distributed training, the bottlenecks have gradually moved from computational resources to communication process. Recent researches adopt in-network aggregation (INA) that offloads the gradient aggregation process to programmable switches, thereby reducing network traffic amount and transmission latency. Unfortunately, due to the bandwidth competition in shared training clusters, the straggler will slow down the training efficiency of INA. To address this issue, we propose an Asynchronous Control based Aggregation Transport Protocol (AC-ATP), which makes full use uncongested links to transmit gradients and the switch memory to cache gradients from the fast workers to accelerate the gradient aggregation. Meanwhile, AC-ATP performs congestion control according to the transmission progress of worker and the remaining completion time of the job. The evaluation results of real testbed and large-scale simulations show that AC-ATP reduces the aggregate time by up to 68% and speeds up training in real-world benchmark models.

Each output port in a shared memory switch can compete for shared memory pool resources. The allocation strategy of the shared buffer directly affects the ability of each output port to absorb network traffic. Due to the unpredictability of traditional network traffic, existing switch buffer management strategies take a passive approach, allocating buffers to each port only after traffic arrives. This passive response has the problem of untimely buffer allocation and cannot effectively absorb burst traffic. Distributed deep learning follows a specific training pattern, and network traffic exhibits obvious periodic characteristics during transmission. Thus, we propose a Proactive Dynamic Threshold (PDT) strategy, which realizes the pre-adjustment of switch port threshold by detecting the traffic characteristics of distributed training.