Rong Chen

Dingyan Zhang, Jinbo Han, Kaixi Zhang et al.

High-quality LLM request scheduling requires achieving two key objectives: whether the routed instance has KV$ to accelerate the request execution and whether the workload is balanced across instances. Achieving both objectives is challenging because pursuing one objective may compromise the other. Current approaches adopt various combinators (e.g., linear combinations) to compute a scheduling score combining indicators for the two objectives, which are complex in that they either require significant workload-specific hyperparameter tuning or model-hardware-aware simulator development, and could still lead to suboptimal performance. In this paper, we show that using a simple multiplication of two carefully chosen indicators-one for KV$-aware (new prefill tokens if routed to an instance) and one for load balancing-aware (current batch size of the instance)-as the scheduling score can simultaneously achieve both objectives well without any hyperparameter tuning. The key idea is that the multiplied score considers both objectives in a manner similar to a linear combination, with a nice property that the original hyperparameters are canceled out during comparison so we don't need tuning to find the best parameters. The two indicators are chosen based on our analysis of LLM characteristics, and our extensive experiments show that this simple approach can reduce TTFT by 92% and 52%, and TPOT by 21% and 20%, compared to vLLM-v1 and a production scheduler on real-world workloads covering chatbots, API calls, and coding agents. We also mathematically derive the conditions under which multiplication may fail, and find that such conditions are extremely rare in practice and can be detected (and mitigated) beforehand.

Yifan Peng, Jiafei Fan, Xingda Wei et al.

Learned top-K search is a promising approach for serving vector queries with both high accuracy and performance. However, current models trained for a specific K value fail to generalize to real-world multi-K queries: they suffer from accuracy degradation (for larger Ks) and performance loss (for smaller Ks). Training the model to generalize on different Ks requires orders of magnitude more preprocessing time and is not suitable for serving vector queries in the wild. We present OMEGA, a K-generalizable learned top-K search method that simultaneously achieves high accuracy, high performance, and low preprocessing cost for multi-K vector queries. The key idea is that a base model properly trained on K=1 with our trajectory-based features can be used to accurately predict larger Ks with a dynamic refinement procedure and smaller Ks with minimal performance loss. To make our refinements efficient, we further leverage the statistical properties of top-K searches to reduce excessive model invocations. Extensive evaluations on multiple public and production datasets show that, under the same preprocessing budgets, OMEGA achieves 6-33% lower average latency compared to state-of-the-art learned search methods, while all systems achieve the same recall target. With only 16-30% of the preprocessing time, OMEGA attains 1.01-1.28x of the optimal average latency of these baselines.

Tianyu Dai, Khaled Aljanaideh, Rong Chen et al.

MATLAB(R) releases over the last 3 years have witnessed a continuing growth in the dynamic modeling capabilities offered by the System Identification Toolbox(TM). The emphasis has been on integrating deep learning architectures and training techniques that facilitate the use of deep neural networks as building blocks of nonlinear models. The toolbox offers neural state-space models which can be extended with auto-encoding features that are particularly suited for reduced-order modeling of large systems. The toolbox contains several other enhancements that deepen its integration with the state-of-art machine learning techniques, leverage auto-differentiation features for state estimation, and enable a direct use of raw numeric matrices and timetables for training models.

Yuhao Cheng, Siru Zhang, Yiqiang Yan et al.

Multimedia compression allows us to watch videos, see pictures and hear sounds within a limited bandwidth, which helps the flourish of the internet. During the past decades, multimedia compression has achieved great success using hand-craft features and systems. With the development of artificial intelligence and video compression, there emerges a lot of research work related to using the neural network on the video compression task to get rid of the complicated system. Not only producing the advanced algorithms, but researchers also spread the compression to different content, such as User Generated Content(UGC). With the rapid development of mobile devices, screen content videos become an important part of multimedia data. In contrast, we find community lacks a large-scale dataset for screen content video compression, which impedes the fast development of the corresponding learning-based algorithms. In order to fulfill this blank and accelerate the research of this special type of videos, we propose the Large-scale Screen Content Dataset(LSCD), which contains 714 source sequences. Meanwhile, we provide the analysis of the proposed dataset to show some features of screen content videos, which will help researchers have a better understanding of how to explore new algorithms. Besides collecting and post-processing the data to organize the dataset, we also provide a benchmark containing the performance of both traditional codec and learning-based methods.

Yulong Fan, Rong Chen, Bo Li

Rain removal in images is an important task in computer vision filed and attracting attentions of more and more people. In this paper, we address a non-trivial issue of removing visual effect of rain streak from a single image. Differing from existing work, our method combines various semantic constraint task in a proposed multi-task regression model for rain removal. These tasks reinforce the model's capabilities from the content, edge-aware, and local texture similarity respectively. To further improve the performance of multi-task learning, we also present two simple but powerful dynamic weighting algorithms. The proposed multi-task enhanced network (MENET) is a powerful convolutional neural network based on U-Net for rain removal research, with a specific focus on utilize multiple tasks constraints and exploit the synergy among them to facilitate the model's rain removal capacity. It is noteworthy that the adaptive weighting scheme has further resulted in improved network capability. We conduct several experiments on synthetic and real rain images, and achieve superior rain removal performance over several selected state-of-the-art (SOTA) approaches. The overall effect of our method is impressive, even in the decomposition of heavy rain and rain streak accumulation.The source code and some results can be found at:https://github.com/SumiHui/MENET.

Shuzhi Cao, Rong Chen, Ailong He et al.

In search systems, effectively coordinating the two core objectives of search relevance matching and click-through rate (CTR) prediction is crucial for discovering users' interests and enhancing platform revenue. In our prior work PRECTR, we proposed a unified framework to integrate these two subtasks,thereby eliminating their inconsistency and leading to mutual benefit.However, our previous work still faces three main challenges. First, low-active users and new users have limited search behavioral data, making it difficult to achieve effective personalized relevance preference modeling. Second, training data for ranking models predominantly come from high-relevance exposures, creating a distribution mismatch with the broader candidate space in coarse-ranking, leading to generalization bias. Third, due to the latency constraint, the original model employs an Emb+MLP architecture with a frozen BERT encoder, which prevents joint optimization and creates misalignment between representation learning and CTR fine-tuning. To solve these issues, we further reinforce our method and propose PRECTR-V2. Specifically, we mitigate the low-activity users' sparse behavior problem by mining global relevance preferences under the specific query, which facilitates effective personalized relevance modeling for cold-start scenarios. Subsequently, we construct hard negative samples through embedding noise injection and relevance label reconstruction, and optimize their relative ranking against positive samples via pairwise loss, thereby correcting exposure bias. Finally, we pretrain a lightweight transformer-based encoder via knowledge distillation from LLM and SFT on the text relevance classification task. This encoder replaces the frozen BERT module, enabling better adaptation to CTR fine-tuning and advancing beyond the traditional Emb+MLP paradigm.

Ying Song, Yijing Wang, Hui Yang et al.

Evaluating platform-level interventions in search-based two-sided marketplaces is fundamentally challenged by systemic effects such as spillovers and network interference. While widely used for causal inference, the PSM (Propensity Score Matching) - DID (Difference-in-Differences) framework remains susceptible to selection bias and cross-unit interference from unaccounted spillovers. In this paper, we introduced Competitive Isolation PSM-DID, a novel causal framework that integrates propensity score matching with competitive isolation to enable platform-level effect measurement (e.g., order volume, GMV) instead of item-level metrics in search systems. Our approach provides theoretically guaranteed unbiased estimation under mutual exclusion conditions, with an open dataset released to support reproducible research on marketplace interference (github.com/xxxx). Extensive experiments demonstrate significant reductions in interference effects and estimation variance compared to baseline methods. Successful deployment in a large-scale marketplace confirms the framework's practical utility for platform-level causal inference.

Xiaoniu Song, Zihang Zhong, Rong Chen et al.

The promising applications of large language models are often limited by the constrained GPU memory capacity available on edge devices. Mixture-of-Experts (MoE) models help address this issue by activating only a subset of the model's parameters during computation. This approach allows the unused parameters to be offloaded to host memory, thereby reducing the overall GPU memory demand. However, existing cache-based offloading solutions handle cache misses reactively, which significantly impacts system performance. In this paper, we introduce ProMoE, a novel proactive caching system that utilizes intermediate results to predict subsequent expert usage. By proactively fetching experts in advance, ProMoE eliminates passive cache misses, removes loading time from the critical path, and reduces the performance overhead associated with offloading. Our evaluations demonstrate that ProMoE achieves an average speedup of 2.20x (up to 3.21x) and 2.07x (up to 5.02x) in the prefill and decode stages, respectively, compared to existing offloading solutions.

Daijun Ding, Rong Chen, Liwen Jing et al.

Cross-target stance detection (CTSD) is an important task, which infers the attitude of the destination target by utilizing annotated data derived from the source target. One important approach in CTSD is to extract domain-invariant features to bridge the knowledge gap between multiple targets. However, the analysis of informal and short text structure, and implicit expressions, complicate the extraction of domain-invariant knowledge. In this paper, we propose a Multi-Perspective Prompt-Tuning (MPPT) model for CTSD that uses the analysis perspective as a bridge to transfer knowledge. First, we develop a two-stage instruct-based chain-of-thought method (TsCoT) to elicit target analysis perspectives and provide natural language explanations (NLEs) from multiple viewpoints by formulating instructions based on large language model (LLM). Second, we propose a multi-perspective prompt-tuning framework (MultiPLN) to fuse the NLEs into the stance predictor. Extensive experiments results demonstrate the superiority of MPPT against the state-of-the-art baseline methods.

Jiahao Wang, Jinbo Han, Xingda Wei et al.

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

Rongxin Cheng, Yuxin Lai, Xingda Wei et al.

Serving LLMs with a cluster of GPUs is common nowadays, where the serving system must meet strict latency SLOs required by applications. However, the stateful nature of LLM serving requires maintaining huge states (i.e., KVCache) in limited GPU memory. Under spikes in real-world workloads, GPU memory can be easily throttled, leading to orders of magnitude higher response latency due to queuing introduced by waiting for KVCache to be reclaimed. Prior KVCache-centric approaches handle load throttling by dropping, migrating, or swapping KVCache. These methods fail to release sufficient memory quickly with requests still queued. This paper proposes the first parameter-centric approach to handling throttling by selectively dropping replicated parameters to instantly free memory for requests, based on an unnoticed observation that model parameters are commonly replicated across GPUs for serving LLMs. With additional memory, all requests can be served with a larger batch without queuing. To make the parameter-centric approach correct and efficient, we cooperatively execute requests on GPUs with a complete copy of parameters using pipeline parallelism, and derive an appropriate drop plan without unnecessary cooperation. We also design techniques to minimize the performance overhead due to pipeline parallelism with the execution patterns of requests under drop. Evaluations show that {\sys} reduces the tail TTFT of requests under throttling by up to 72.2 times compared to the state-of-the-art systems including Llumnix, vLLM and InferCept.

Rongxin Cheng, Kai Zhou, Xingda Wei et al.

Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. SpecActor achieves fast rollout with speculative decoding that deploys a fast path (e.g., a smaller model) to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges in speculative rollout by (1) a \emph{dynamic decoupled speculation} execution method that maximizes the GPU computational efficiency to realize speedup for large-batch execution -- a configuration common in training but unfriendly to speculative execution and (2) a \emph{dynamic Best-of-N speculation} method that selects and combines different drafting methods according to the rollout progress. It substantially improves the speculation accuracy even when the best drafting method is unknown a priori, meanwhile without requiring adding extra computation resources. {\sys} is {1.3--1.7}\,$\times$ faster than common post-training baselines, and is {1.3--1.5}\,$\times$ faster compared to naively adopting speculative decoding for rollout.

Rong Chen, Shuzhi Cao, Ailong He et al.

The two primary tasks in the search recommendation system are search relevance matching and click-through rate (CTR) prediction -- the former focuses on seeking relevant items for user queries whereas the latter forecasts which item may better match user interest. Prior research typically develops two models to predict the CTR and search relevance separately, then ranking candidate items based on the fusion of the two outputs. However, such a divide-and-conquer paradigm creates the inconsistency between different models. Meanwhile, the search relevance model mainly concentrates on the degree of objective text matching while neglecting personalized differences among different users, leading to restricted model performance. To tackle these issues, we propose a unified Personalized Search RElevance Matching and CTR Prediction Fusion Model(PRECTR). Specifically, based on the conditional probability fusion mechanism, PRECTR integrates the CTR prediction and search relevance matching into one framework to enhance the interaction and consistency of the two modules. However, directly optimizing CTR binary classification loss may bring challenges to the fusion model's convergence and indefinitely promote the exposure of items with high CTR, regardless of their search relevance. Hence, we further introduce two-stage training and semantic consistency regularization to accelerate the model's convergence and restrain the recommendation of irrelevant items. Finally, acknowledging that different users may have varied relevance preferences, we assessed current users' relevance preferences by analyzing past users' preferences for similar queries and tailored incentives for different candidate items accordingly. Extensive experimental results on our production dataset and online A/B testing demonstrate the effectiveness and superiority of our proposed PRECTR method.

Chengwei Zhang, Shan Jin, Wanli Xue et al.

The adaptive traffic signal control (ATSC) problem can be modeled as a multiagent cooperative game among urban intersections, where intersections cooperate to optimize their common goal. Recently, reinforcement learning (RL) has achieved marked successes in managing sequential decision making problems, which motivates us to apply RL in the ASTC problem. Here we use independent reinforcement learning (IRL) to solve a complex traffic cooperative control problem in this study. One of the largest challenges of this problem is that the observation information of intersection is typically partially observable, which limits the learning performance of IRL algorithms. To this, we model the traffic control problem as a partially observable weak cooperative traffic model (PO-WCTM) to optimize the overall traffic situation of a group of intersections. Different from a traditional IRL task that averages the returns of all agents in fully cooperative games, the learning goal of each intersection in PO-WCTM is to reduce the cooperative difficulty of learning, which is also consistent with the traffic environment hypothesis. We also propose an IRL algorithm called Cooperative Important Lenient Double DQN (CIL-DDQN), which extends Double DQN (DDQN) algorithm using two mechanisms: the forgetful experience mechanism and the lenient weight training mechanism. The former mechanism decreases the importance of experiences stored in the experience reply buffer, which deals with the problem of experience failure caused by the strategy change of other agents. The latter mechanism increases the weight experiences with high estimation and `leniently' trains the DDQN neural network, which improves the probability of the selection of cooperative joint strategies. Experimental results show that CIL-DDQN outperforms other methods in almost all performance indicators of the traffic control problem.

Xiatian Zhang, Xunshi He, Nan Wang et al.

For high-dimensional data, there are huge communication costs for distributed GBDT because the communication volume of GBDT is related to the number of features. To overcome this problem, we propose a novel gradient boosting algorithm, the Gradient Boosting Untrained Neural Network(GBUN). GBUN ensembles the untrained randomly generated neural network that softly distributes data samples to multiple neuron outputs and dramatically reduces the communication costs for distributed learning. To avoid creating huge neural networks for high-dimensional data, we extend Simhash algorithm to mimic forward calculation of the neural network. Our experiments on multiple public datasets show that GBUN is as good as conventional GBDT in terms of prediction accuracy and much better than it in scaling property for distributed learning. Comparing to conventional GBDT varieties, GBUN speeds up the training process up to 13 times on the cluster with 64 machines, and up to 4614 times on the cluster with 100KB/s network bandwidth. Therefore, GBUN is not only an efficient distributed learning algorithm but also has great potentials for federated learning.

Chencheng Cai, Rong Chen, Han Xiao

Discovering the underlying low dimensional structure of high dimensional data has attracted a significant amount of researches recently and has shown to have a wide range of applications. As an effective dimension reduction tool, singular value decomposition is often used to analyze high dimensional matrices, which are traditionally assumed to have a low rank matrix approximation. In this paper, we propose a new approach. We assume a high dimensional matrix can be approximated by a sum of a small number of Kronecker products of matrices with potentially different configurations, named as a hybird Kronecker outer Product Approximation (hKoPA). It provides an extremely flexible way of dimension reduction compared to the low-rank matrix approximation. Challenges arise in estimating a hKoPA when the configurations of component Kronecker products are different or unknown. We propose an estimation procedure when the set of configurations are given and a joint configuration determination and component estimation procedure when the configurations are unknown. Specifically, a least squares backfitting algorithm is used when the configuration is given. When the configuration is unknown, an iterative greedy algorithm is used. Both simulation and real image examples show that the proposed algorithms have promising performances. The hybrid Kronecker product approximation may have potentially wider applications in low dimensional representation of high dimensional data

Chencheng Cai, Rong Chen, Han Xiao

We consider the problem of matrix approximation and denoising induced by the Kronecker product decomposition. Specifically, we propose to approximate a given matrix by the sum of a few Kronecker products of matrices, which we refer to as the Kronecker product approximation (KoPA). Because the Kronecker product is an extension of the outer product from vectors to matrices, KoPA extends the low rank matrix approximation, and includes it as a special case. Comparing with the latter, KoPA also offers a greater flexibility, since it allows the user to choose the configuration, which are the dimensions of the two smaller matrices forming the Kronecker product. On the other hand, the configuration to be used is usually unknown, and needs to be determined from the data in order to achieve the optimal balance between accuracy and parsimony. We propose to use extended information criteria to select the configuration. Under the paradigm of high dimensional analysis, we show that the proposed procedure is able to select the true configuration with probability tending to one, under suitable conditions on the signal-to-noise ratio. We demonstrate the superiority of KoPA over the low rank approximations through numerical studies, and several benchmark image examples.

Renee Su, Rong Chen

This paper presents a change detection method that identifies land cover changes from aerial imagery, using semantic segmentation, a machine learning approach. We present a land cover classification training pipeline with Deeplab v3+, state-of-the-art semantic segmentation technology, including data preparation, model training for seven land cover types, and model exporting modules. In the land cover change detection system, the inputs are images retrieved from Google Earth at the same location but from different times. The system then predicts semantic segmentation results on these images using the trained model and calculates the land cover class percentage for each input image. We see an improvement in the accuracy of the land cover semantic segmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in the DeepGlobe land cover classification challenge. The land cover change detection system that leverages the state-of-the-art semantic segmentation technology is proposed and can be used for deforestation analysis, land management, and urban planning.

Chencheng Cai, Rong Chen, Han Xiao

A matrix completion problem is to recover the missing entries in a partially observed matrix. Most of the existing matrix completion methods assume a low rank structure of the underlying complete matrix. In this paper, we introduce an alternative and more general form of the underlying complete matrix, which assumes a low Kronecker rank instead of a low regular rank, but includes the latter as a special case. The extra flexibility allows for a much more parsimonious representation of the underlying matrix, but also raises the challenge of determining the proper Kronecker product configuration to be used. We find that the configuration can be identified using the mean squared error criterion as well as a modified cross-validation criterion. We establish the consistency of this procedure under suitable conditions on the signal-to-noise ratio. A aggregation procedure is also proposed to deal with special missing patterns and complex underlying structures. Both numerical and empirical studies are carried out to demonstrate the performance of the new method.

Xiaotong Luo, Rong Chen, Yuan Xie et al.

Image quality measurement is a critical problem for image super-resolution (SR) algorithms. Usually, they are evaluated by some well-known objective metrics, e.g., PSNR and SSIM, but these indices cannot provide suitable results in accordance with the perception of human being. Recently, a more reasonable perception measurement has been proposed in [1], which is also adopted by the PIRM-SR 2018 challenge. In this paper, motivated by [1], we aim to generate a high-quality SR result which balances between the two indices, i.e., the perception index and root-mean-square error (RMSE). To do so, we design a new deep SR framework, dubbed Bi-GANs-ST, by integrating two complementary generative adversarial networks (GAN) branches. One is memory residual SRGAN (MR-SRGAN), which emphasizes on improving the objective performance, such as reducing the RMSE. The other is weight perception SRGAN (WP-SRGAN), which obtains the result that favors better subjective perception via a two-stage adversarial training mechanism. Then, to produce final result with excellent perception scores and RMSE, we use soft-thresholding method to merge the results generated by the two GANs. Our method performs well on the perceptual image super-resolution task of the PIRM 2018 challenge. Experimental results on five benchmarks show that our proposal achieves highly competent performance compared with other state-of-the-art methods.

Chengwei Zhang, Xiaohong Li, Jianye Hao et al.

Although many reinforcement learning methods have been proposed for learning the optimal solutions in single-agent continuous-action domains, multiagent coordination domains with continuous actions have received relatively few investigations. In this paper, we propose an independent learner hierarchical method, named Sample Continuous Coordination with recursive Frequency Maximum Q-Value (SCC-rFMQ), which divides the cooperative problem with continuous actions into two layers. The first layer samples a finite set of actions from the continuous action spaces by a re-sampling mechanism with variable exploratory rates, and the second layer evaluates the actions in the sampled action set and updates the policy using a reinforcement learning cooperative method. By constructing cooperative mechanisms at both levels, SCC-rFMQ can handle cooperative problems in continuous action cooperative Markov games effectively. The effectiveness of SCC-rFMQ is experimentally demonstrated on two well-designed games, i.e., a continuous version of the climbing game and a cooperative version of the boat problem. Experimental results show that SCC-rFMQ outperforms other reinforcement learning algorithms.