arXiv subjects

Lifeng Sun

Publications and source records attributed to Lifeng Sun.

At least 19 recordsLinked to original sources

arXiv2026

SAIL: Perceptual Quality-Aware Rate Control for Cloud Gaming

Cloud gaming streams cloud-rendered frames under strict motion-to-photon latency, yet its at-scale viability is increasingly constrained by bandwidth cost: in our study of the T cloud gaming platform, bandwidth accounts for 30-60% of total operating expense. This high bandwidth consumption stems from a fidelity-first objective of making the stream perceptually indistinguishable from local gameplay. It drives production systems toward best-effort bitrate allocation that pushes the encoder to the highest rate allowed by congestion control. However, the bitrate-perception relationship saturates: beyond a frame-dependent perceptually lossless threshold, additional bits yield negligible perceptual improvement, creating systematic redundant quality that wastes bandwidth. We present SAIL, a production quality-aware rate control system with the goal of achieving perceptually lossless quality while avoiding unnecessary bandwidth waste. SAIL adopts a post-encoding architecture to enable millisecond-scale feedback at near-zero overhead. It comprises three key designs: (i) an encoder-driven quality assessment model that leverages zero-cost encoder outputs for real-time quality estimation; (ii) a hybrid rate control mechanism that balances steady-state adaptation with dynamic spike absorption; and (iii) a network-aware strategy that coordinates with congestion control to prevent capacity underestimation. SAIL has been fully deployed on the T cloud gaming platform and reduces bandwidth consumption by 44.27% and end-to-end latency by 8.37% without degrading perceived quality, serving tens of millions of users and accumulating billions of hours of total gameplay.

cs.NI↗

arXiv2026

Test-time Scaling over Perception: Resolving the Grounding Paradox in Thinking with Images

Recent multimodal large language models (MLLMs) support Thinking with Images, invoking visual tools such as zooming and cropping to inspect image regions during inference. Yet these systems remain brittle in fine-grained reasoning: to acquire a decisive detail, a model must ground its attention on the correct region, but knowing which region is correct presupposes having already observed that detail. We identify this circular dependency as the grounding paradox, show that grounding errors are rarely self-corrected within a single trajectory---once a misleading region is inspected, all subsequent reasoning conditions on that observation and the error propagates to the final answer---and observe that because each trajectory constructs its own evidence, answer-level aggregation discards the very information that distinguishes trajectories. We propose Test-Time Scaling over Perception (TTSP), a closed-loop framework that treats perception as the unit of scalable inference and allocates compute along two axes: Entropy-Gated Perceptual Exploration samples diverse trajectories and uses critical-token entropy to withhold evidence the model cannot commit to, while Evidence-Guided Iterative Refinement distills validated observations into a correctable Evidence Ledger that steers later rounds to re-inspect unresolved regions. Across high-resolution and general multimodal benchmarks, TTSP consistently outperforms strong test-time scaling baselines, while improving grounding quality with favorable token efficiency.

cs.CV↗

arXiv2026

MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning

Medical Vision-Language Models (VLMs) hold immense promise for complex clinical tasks, but their reasoning capabilities are often constrained by text-only paradigms that fail to ground inferences in visual evidence. This limitation not only curtails performance on tasks requiring fine-grained visual analysis but also introduces risks of visual hallucination in safety-critical applications. Thus, we introduce MedVR, a novel reinforcement learning framework that enables annotation-free visual reasoning for medical VLMs. Its core innovation lies in two synergistic mechanisms: Entropy-guided Visual Regrounding (EVR) uses model uncertainty to direct exploration, while Consensus-based Credit Assignment (CCA) distills pseudo-supervision from rollout agreement. Without any human annotations for intermediate steps, MedVR achieves state-of-the-art performance on diverse public medical VQA benchmarks, significantly outperforming existing models. By learning to reason directly with visual evidence, MedVR promotes the robustness and transparency essential for accelerating the clinical deployment of medical AI.

cs.CV↗

arXiv2026

SubFLOT: Submodel Extraction for Efficient and Personalized Federated Learning via Optimal Transport

Federated Learning (FL) enables collaborative model training while preserving data privacy, but its practical deployment is hampered by system and statistical heterogeneity. While federated network pruning offers a path to mitigate these issues, existing methods face a critical dilemma: server-side pruning lacks personalization, whereas client-side pruning is computationally prohibitive for resource-constrained devices. Furthermore, the pruning process itself induces significant parametric divergence among heterogeneous submodels, destabilizing training and hindering global convergence. To address these challenges, we propose SubFLOT, a novel framework for server-side personalized federated pruning. SubFLOT introduces an Optimal Transport-enhanced Pruning (OTP) module that treats historical client models as proxies for local data distributions, formulating the pruning task as a Wasserstein distance minimization problem to generate customized submodels without accessing raw data. Concurrently, to counteract parametric divergence, our Scaling-based Adaptive Regularization (SAR) module adaptively penalizes a submodel's deviation from the global model, with the penalty's strength scaled by the client's pruning rate. Comprehensive experiments demonstrate that SubFLOT consistently and substantially outperforms state-of-the-art methods, underscoring its potential for deploying efficient and personalized models on resource-constrained edge devices.

cs.LG↗

arXiv2026

Decoupling Defense Strategies for Robust Image Watermarking

Deep learning-based image watermarking, while robust against conventional distortions, remains vulnerable to advanced adversarial and regeneration attacks. Conventional countermeasures, which jointly optimize the encoder and decoder via a noise layer, face 2 inevitable challenges: (1) decrease of clean accuracy due to decoder adversarial training and (2) limited robustness due to simultaneous training of all three advanced attacks. To overcome these issues, we propose AdvMark, a novel two-stage fine-tuning framework that decouples the defense strategies. In stage 1, we address adversarial vulnerability via a tailored adversarial training paradigm that primarily fine-tunes the encoder while only conditionally updating the decoder. This approach learns to move the image into a non-attackable region, rather than modifying the decision boundary, thus preserving clean accuracy. In stage 2, we tackle distortion and regeneration attacks via direct image optimization. To preserve the adversarial robustness gained in stage 1, we formulate a principled, constrained image loss with theoretical guarantees, which balances the deviation from cover and previous encoded images. We also propose a quality-aware early-stop to further guarantee the lower bound of visual quality. Extensive experiments demonstrate AdvMark outperforms with the highest image quality and comprehensive robustness, i.e. up to 29\%, 33\% and 46\% accuracy improvement for distortion, regeneration and adversarial attacks, respectively.

cs.CV↗

arXiv2026

HiVid: LLM-Guided Video Saliency For Content-Aware VOD And Live Streaming

Content-aware streaming requires dynamic, chunk-level importance weights to optimize subjective quality of experience (QoE). However, direct human annotation is prohibitively expensive while vision-saliency models generalize poorly. We introduce HiVid, the first framework to leverage Large Language Models (LLMs) as a scalable human proxy to generate high-fidelity weights for both Video-on-Demand (VOD) and live streaming. We address 3 non-trivial challenges: (1) To extend LLMs' limited modality and circumvent token limits, we propose a perception module to assess frames in a local context window, autoregressively building a coherent understanding of the video. (2) For VOD with rating inconsistency across local windows, we propose a ranking module to perform global re-ranking with a novel LLM-guided merge-sort algorithm. (3) For live streaming which requires low-latency, online inference without future knowledge, we propose a prediction module to predict future weights with a multi-modal time series model, which comprises a content-aware attention and adaptive horizon to accommodate asynchronous LLM inference. Extensive experiments show HiVid improves weight prediction accuracy by up to 11.5\% for VOD and 26\% for live streaming over SOTA baselines. Real-world user study validates HiVid boosts streaming QoE correlation by 14.7\%.

cs.CV↗

arXiv2026

RRNet: Configurable Real-Time Video Enhancement with Arbitrary Local Lighting Variations

With the growing demand for real-time video enhancement in live applications, existing methods often struggle to balance speed and effective exposure control, particularly under uneven lighting. We introduce RRNet (Rendering Relighting Network), a lightweight and configurable framework that achieves a state-of-the-art tradeoff between visual quality and efficiency. By estimating parameters for a minimal set of virtual light sources, RRNet enables localized relighting through a depth-aware rendering module without requiring pixel-aligned training data. This object-aware formulation preserves facial identity and supports real-time, high-resolution performance using a streamlined encoder and lightweight prediction head. To facilitate training, we propose a generative AI-based dataset creation pipeline that synthesizes diverse lighting conditions at low cost. With its interpretable lighting control and efficient architecture, RRNet is well suited for practical applications such as video conferencing, AR-based portrait enhancement, and mobile photography. Experiments show that RRNet consistently outperforms prior methods in low-light enhancement, localized illumination adjustment, and glare removal.

cs.CV↗

arXiv2025

Crucible: Quantifying the Potential of Control Algorithms through LLM Agents

Control algorithms in production environments typically require domain experts to tune their parameters and logic for specific scenarios. However, existing research predominantly focuses on algorithmic performance under ideal or default configurations, overlooking the critical aspect of Tuning Potential. To bridge this gap, we introduce Crucible, an agent that employs an LLM-driven, multi-level expert simulation to turn algorithms and defines a formalized metric to quantitatively evaluate their Tuning Potential. We demonstrate Crucible's effectiveness across a wide spectrum of case studies, from classic control tasks to complex computer systems, and validate its findings in a real-world deployment. Our experimental results reveal that Crucible systematically quantifies the tunable space across different algorithms. Furthermore, Crucible provides a new dimension for algorithm analysis and design, which ultimately leads to performance improvements. Our code is available at https://github.com/thu-media/Crucible.

cs.AI↗

arXiv2025

Towards User-level QoE: Large-scale Practice in Personalized Optimization of Adaptive Video Streaming

Traditional optimization methods based on system-wide Quality of Service (QoS) metrics have approached their performance limitations in modern large-scale streaming systems. However, aligning user-level Quality of Experience~(QoE) with algorithmic optimization objectives remains an unresolved challenge. Therefore, we propose \texttt{LingXi}, the first large-scale deployed system for personalized adaptive video streaming based on user-level experience. \texttt{LingXi} dynamically optimizes the objectives of adaptive video streaming algorithms by analyzing user engagement. Utilizing exit rate as a key metric, we investigate the correlation between QoS indicators and exit rates based on production environment logs, subsequently developing a personalized exit rate predictor. Through Monte Carlo sampling and online Bayesian optimization, we iteratively determine optimal parameters. Large-scale A/B testing utilizing 8\% of traffic on Kuaishou, one of the largest short video platforms, demonstrates \texttt{LingXi}'s superior performance. \texttt{LingXi} achieves a 0.15\% increase in total viewing time, a 0.1\% improvement in bitrate, and a 1.3\% reduction in stall time across all users, with particularly significant improvements for low-bandwidth users who experience a 15\% reduction in stall time.

cs.MM↗

arXiv2025

Beyond Interpretability: Exploring the Comprehensibility of Adaptive Video Streaming through Large Language Models

Over the past decade, adaptive video streaming technology has witnessed significant advancements, particularly driven by the rapid evolution of deep learning techniques. However, the black-box nature of deep learning algorithms presents challenges for developers in understanding decision-making processes and optimizing for specific application scenarios. Although existing research has enhanced algorithm interpretability through decision tree conversion, interpretability does not directly equate to developers' subjective comprehensibility. To address this challenge, we introduce \texttt{ComTree}, the first bitrate adaptation algorithm generation framework that considers comprehensibility. The framework initially generates the complete set of decision trees that meet performance requirements, then leverages large language models to evaluate these trees for developer comprehensibility, ultimately selecting solutions that best facilitate human understanding and enhancement. Experimental results demonstrate that \texttt{ComTree} significantly improves comprehensibility while maintaining competitive performance, showing potential for further advancement. The source code is available at https://github.com/thu-media/ComTree.

cs.MM↗

arXiv2023

Optimizing Adaptive Video Streaming with Human Feedback

Quality of Experience~(QoE)-driven adaptive bitrate (ABR) algorithms are typically optimized using QoE models that are based on the mean opinion score~(MOS), while such principles may not account for user heterogeneity on rating scales, resulting in unexpected behaviors. In this paper, we propose Jade, which leverages reinforcement learning with human feedback~(RLHF) technologies to better align the users' opinion scores. Jade's rank-based QoE model considers relative values of user ratings to interpret the subjective perception of video sessions. We implement linear-based and Deep Neural Network (DNN)-based architectures for satisfying both accuracy and generalization ability. We further propose entropy-aware reinforced mechanisms for training policies with the integration of the proposed QoE models. Experimental results demonstrate that Jade performs favorably on conventional metrics, such as quality and stall ratio, and improves QoE by 8.09%-38.13% in different network conditions, emphasizing the importance of user heterogeneity in QoE modeling and the potential of combining linear-based and DNN-based models for performance improvement.

cs.MM↗

arXiv2021

A General Method For Automatic Discovery of Powerful Interactions In Click-Through Rate Prediction

Modeling powerful interactions is a critical challenge in Click-through rate (CTR) prediction, which is one of the most typical machine learning tasks in personalized advertising and recommender systems. Although developing hand-crafted interactions is effective for a small number of datasets, it generally requires laborious and tedious architecture engineering for extensive scenarios. In recent years, several neural architecture search (NAS) methods have been proposed for designing interactions automatically. However, existing methods only explore limited types and connections of operators for interaction generation, leading to low generalization ability. To address these problems, we propose a more general automated method for building powerful interactions named AutoPI. The main contributions of this paper are as follows: AutoPI adopts a more general search space in which the computational graph is generalized from existing network connections, and the interactive operators in the edges of the graph are extracted from representative hand-crafted works. It allows searching for various powerful feature interactions to produce higher AUC and lower Logloss in a wide variety of applications. Besides, AutoPI utilizes a gradient-based search strategy for exploration with a significantly low computational cost. Experimentally, we evaluate AutoPI on a diverse suite of benchmark datasets, demonstrating the generalizability and efficiency of AutoPI over hand-crafted architectures and state-of-the-art NAS algorithms.

cs.IR↗

arXiv2021

Multimedia Edge Computing

In this paper, we investigate the recent studies on multimedia edge computing, from sensing not only traditional visual/audio data but also individuals' geographical preference and mobility behaviors, to performing distributed machine learning over such data using the joint edge and cloud infrastructure and using evolutional strategies like reinforcement learning and online learning at edge devices to optimize the quality of experience for multimedia services at the last mile proactively. We provide both a retrospective view of recent rapid migration (resp. merge) of cloud multimedia to (resp. and) edge-aware multimedia and insights on the fundamental guidelines for designing multimedia edge computing strategies that target satisfying the changing demand of quality of experience. By showing the recent research studies and industrial solutions, we also provide future directions towards high-quality multimedia services over edge computing.

cs.MM↗

arXiv2020

Federated Learning with Unbiased Gradient Aggregation and Controllable Meta Updating

Federated learning (FL) aims to train machine learning models in the decentralized system consisting of an enormous amount of smart edge devices. Federated averaging (FedAvg), the fundamental algorithm in FL settings, proposes on-device training and model aggregation to avoid the potential heavy communication costs and privacy concerns brought by transmitting raw data. However, through theoretical analysis we argue that 1) the multiple steps of local updating will result in gradient biases and 2) there is an inconsistency between the expected target distribution and the optimization objectives following the training paradigm in FedAvg. To tackle these problems, we first propose an unbiased gradient aggregation algorithm with the keep-trace gradient descent and the gradient evaluation strategy. Then we introduce an additional controllable meta updating procedure with a small set of data samples, indicating the expected target distribution, to provide a clear and consistent optimization objective. Both the two improvements are model- and task-agnostic and can be applied individually or together. Experimental results demonstrate that the proposed methods are faster in convergence and achieve higher accuracy with different network architectures in various FL settings.

cs.LG↗

arXiv2020

Continual Local Training for Better Initialization of Federated Models

Federated learning (FL) refers to the learning paradigm that trains machine learning models directly in the decentralized systems consisting of smart edge devices without transmitting the raw data, which avoids the heavy communication costs and privacy concerns. Given the typical heterogeneous data distributions in such situations, the popular FL algorithm \emph{Federated Averaging} (FedAvg) suffers from weight divergence and thus cannot achieve a competitive performance for the global model (denoted as the \emph{initial performance} in FL) compared to centralized methods. In this paper, we propose the local continual training strategy to address this problem. Importance weights are evaluated on a small proxy dataset on the central server and then used to constrain the local training. With this additional term, we alleviate the weight divergence and continually integrate the knowledge on different local clients into the global model, which ensures a better generalization ability. Experiments on various FL settings demonstrate that our method significantly improves the initial performance of federated models with few extra communication costs.

cs.LG↗

arXiv2020

Self-play Reinforcement Learning for Video Transmission

Video transmission services adopt adaptive algorithms to ensure users' demands. Existing techniques are often optimized and evaluated by a function that linearly combines several weighted metrics. Nevertheless, we observe that the given function fails to describe the requirement accurately. Thus, such proposed methods might eventually violate the original needs. To eliminate this concern, we propose \emph{Zwei}, a self-play reinforcement learning algorithm for video transmission tasks. Zwei aims to update the policy by straightforwardly utilizing the actual requirement. Technically, Zwei samples a number of trajectories from the same starting point and instantly estimates the win rate w.r.t the competition outcome. Here the competition result represents which trajectory is closer to the assigned requirement. Subsequently, Zwei optimizes the strategy by maximizing the win rate. To build Zwei, we develop simulation environments, design adequate neural network models, and invent training methods for dealing with different requirements on various video transmission scenarios. Trace-driven analysis over two representative tasks demonstrates that Zwei optimizes itself according to the assigned requirement faithfully, outperforming the state-of-the-art methods under all considered scenarios.

cs.MM↗

arXiv2019

Comyco: Quality-Aware Adaptive Video Streaming via Imitation Learning

Learning-based Adaptive Bit Rate~(ABR) method, aiming to learn outstanding strategies without any presumptions, has become one of the research hotspots for adaptive streaming. However, it typically suffers from several issues, i.e., low sample efficiency and lack of awareness of the video quality information. In this paper, we propose Comyco, a video quality-aware ABR approach that enormously improves the learning-based methods by tackling the above issues. Comyco trains the policy via imitating expert trajectories given by the instant solver, which can not only avoid redundant exploration but also make better use of the collected samples. Meanwhile, Comyco attempts to pick the chunk with higher perceptual video qualities rather than video bitrates. To achieve this, we construct Comyco's neural network architecture, video datasets and QoE metrics with video quality features. Using trace-driven and real-world experiments, we demonstrate significant improvements of Comyco's sample efficiency in comparison to prior work, with 1700x improvements in terms of the number of samples required and 16x improvements on training time required. Moreover, results illustrate that Comyco outperforms previously proposed methods, with the improvements on average QoE of 7.5% - 16.79%. Especially, Comyco also surpasses state-of-the-art approach Pensieve by 7.37% on average video quality under the same rebuffering time.

cs.MM↗

arXiv2019

Adversarial Feature Alignment: Avoid Catastrophic Forgetting in Incremental Task Lifelong Learning

Human beings are able to master a variety of knowledge and skills with ongoing learning. By contrast, dramatic performance degradation is observed when new tasks are added to an existing neural network model. This phenomenon, termed as \emph{Catastrophic Forgetting}, is one of the major roadblocks that prevent deep neural networks from achieving human-level artificial intelligence. Several research efforts, e.g. \emph{Lifelong} or \emph{Continual} learning algorithms, have been proposed to tackle this problem. However, they either suffer from an accumulating drop in performance as the task sequence grows longer, or require to store an excessive amount of model parameters for historical memory, or cannot obtain competitive performance on the new tasks. In this paper, we focus on the incremental multi-task image classification scenario. Inspired by the learning process of human students, where they usually decompose complex tasks into easier goals, we propose an adversarial feature alignment method to avoid catastrophic forgetting. In our design, both the low-level visual features and high-level semantic features serve as soft targets and guide the training process in multiple stages, which provide sufficient supervised information of the old tasks and help to reduce forgetting. Due to the knowledge distillation and regularization phenomenons, the proposed method gains even better performance than finetuning on the new tasks, which makes it stand out from other methods. Extensive experiments in several typical lifelong learning scenarios demonstrate that our method outperforms the state-of-the-art methods in both accuracies on new tasks and performance preservation on old tasks.

cs.LG↗