Multimodal content is crucial for click-through rate (CTR) prediction. However, directly incorporating continuous embeddings from pre-trained models into CTR models yields suboptimal results due to misaligned optimization objectives and convergence speed inconsistency during joint training. Discretizing embeddings into semantic IDs before feeding them into CTR models offers a more effective solution, yet existing methods suffer from limited codebook utilization, reconstruction accuracy, and semantic discriminability. We propose RQ-GMM (Residual Quantized Gaussian Mixture Model), which introduces probabilistic modeling to better capture the statistical structure of multimodal embedding spaces. Through Gaussian Mixture Models combined with residual quantization, RQ-GMM achieves superior codebook utilization and reconstruction accuracy. Experiments on public datasets and online A/B tests on a large-scale short-video platform serving hundreds of millions of users demonstrate substantial improvements: RQ-GMM yields a 1.502% gain in Advertiser Value over strong baselines. The method has been fully deployed, serving daily recommendations for hundreds of millions of users.

Recent methods performing well on Learning with Noisy Label (LNL) problem generally are based on semi-supervised learning and consistency regularization. It usually consists of three stages: warm-up, noisy/clean data division, and semi-supervised learning. However, these methods trained purely with classification consistency suffer from the confirmation bias problem and tend to memorize the noisy labels, resulting in accumulated error and degraded performance. Leveraging the compositional and relational peculiarities of the noisy data, we propose a graph-based Multi-Level Consistency (MLC) framework that jointly exploits multi-level relation consistencies between graphs and classification consistency which can better correct wrong labels by continuing to learn the multi-level differences between clean data and noisy data. Moreover, we propose a Dynamic Filter Module (DFM) which effectively improves the reliability of divided data by re-filtering noisy data despite its simplicity. Our method achieves the state-of-the-art performance on multiple benchmark datasets. On Cifar-100 with 90% noisy labels, our method achieves a top-1 accuracy of 49.1%, outperforming DivideMix by 17.6%.

Though Transformer-based models have achieved great success in the automatic speech recognition (ASR) field, they are generally resource-hungry and computation-intensive which makes them difficult to deploy in resource-restricted devices. In this paper, we propose LRTD, a lightweight Transformer for end-to-end ASR. LRTD compresses the model size using matrix decomposition during training and further reduces the depth and width of the trained model dynamically leveraging two structured pruning strategies. The performance of the pruned model reply on two additional structured dropout methods and two search techniques that can recognize the important layers and attention heads. The experimental results show that our proposed model can achieve competitive performance on Aishell-1 even with about 2.27 × fewer model parameters compared to the baseline transformer model. Moreover, we experimentally demonstrate that the matrix decomposition technique achieves a higher compression rate while two pruning methods can adjust the model size flexibly.