Mingyu Yang

Daizong Liu, Mingyu Yang, Xiaoye Qu et al.

With the significant development of large models in recent years, Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a wide range of multimodal understanding and reasoning tasks. Compared to traditional Large Language Models (LLMs), LVLMs present great potential and challenges due to its closer proximity to the multi-resource real-world applications and the complexity of multi-modal processing. However, the vulnerability of LVLMs is relatively underexplored, posing potential security risks in daily usage. In this paper, we provide a comprehensive review of the various forms of existing LVLM attacks. Specifically, we first introduce the background of attacks targeting LVLMs, including the attack preliminary, attack challenges, and attack resources. Then, we systematically review the development of LVLM attack methods, such as adversarial attacks that manipulate model outputs, jailbreak attacks that exploit model vulnerabilities for unauthorized actions, prompt injection attacks that engineer the prompt type and pattern, and data poisoning that affects model training. Finally, we discuss promising research directions in the future. We believe that our survey provides insights into the current landscape of LVLM vulnerabilities, inspiring more researchers to explore and mitigate potential safety issues in LVLM developments. The latest papers on LVLM attacks are continuously collected in https://github.com/liudaizong/Awesome-LVLM-Attack.

Mingyu Yang, Jian Zhao, Xunhan Hu et al.

Cooperative multi-agent reinforcement learning (MARL) has made prominent progress in recent years. For training efficiency and scalability, most of the MARL algorithms make all agents share the same policy or value network. However, in many complex multi-agent tasks, different agents are expected to possess specific abilities to handle different subtasks. In those scenarios, sharing parameters indiscriminately may lead to similar behavior across all agents, which will limit the exploration efficiency and degrade the final performance. To balance the training complexity and the diversity of agent behavior, we propose a novel framework to learn dynamic subtask assignment (LDSA) in cooperative MARL. Specifically, we first introduce a subtask encoder to construct a vector representation for each subtask according to its identity. To reasonably assign agents to different subtasks, we propose an ability-based subtask selection strategy, which can dynamically group agents with similar abilities into the same subtask. In this way, agents dealing with the same subtask share their learning of specific abilities and different subtasks correspond to different specific abilities. We further introduce two regularizers to increase the representation difference between subtasks and stabilize the training by discouraging agents from frequently changing subtasks, respectively. Empirical results show that LDSA learns reasonable and effective subtask assignment for better collaboration and significantly improves the learning performance on the challenging StarCraft II micromanagement benchmark and Google Research Football.

Wenxuan Xia, Mingyu Yang, Wentao Li et al.

Approximate k-Nearest Neighbor (AKNN) search is widely used in vector databases. When vectors carry additional attributes (e.g., labels or numerical values), filtered AKNN search retrieves the nearest vectors to a query vector under attribute constraints. Most existing methods use a fixed termination condition, searching the entire index while respecting attribute filters. However, this leads to substantial redundant computations, since different queries require different amounts of search effort, and thus misses early termination opportunities for easy queries. This paper proposes a lightweight model to estimate the search cost of filtered AKNN queries and enable adaptive termination: For easy queries, the search stops early to reduce latency, while for hard queries, it continues longer to preserve accuracy. The key challenge is accurate cost prediction under attribute filters. To address this, we show that information collected during an early probing phase (e.g., attribute distributions and intermediate distance statistics) can effectively predict the overall search cost. Experiments on six real-world datasets demonstrate 1.1-3.7 speedup over state-of-the-art baselines at 95% recall, while maintaining search accuracy.

Mingyu Yang, Yu Chen, Hun-Seok Kim

In recent years, deep learning-based approaches for visual-inertial odometry (VIO) have shown remarkable performance outperforming traditional geometric methods. Yet, all existing methods use both the visual and inertial measurements for every pose estimation incurring potential computational redundancy. While visual data processing is much more expensive than that for the inertial measurement unit (IMU), it may not always contribute to improving the pose estimation accuracy. In this paper, we propose an adaptive deep-learning based VIO method that reduces computational redundancy by opportunistically disabling the visual modality. Specifically, we train a policy network that learns to deactivate the visual feature extractor on the fly based on the current motion state and IMU readings. A Gumbel-Softmax trick is adopted to train the policy network to make the decision process differentiable for end-to-end system training. The learned strategy is interpretable, and it shows scenario-dependent decision patterns for adaptive complexity reduction. Experiment results show that our method achieves a similar or even better performance than the full-modality baseline with up to 78.8% computational complexity reduction for KITTI dataset evaluation. The code is available at https://github.com/mingyuyng/Visual-Selective-VIO.

Wenxuan Xia, Mingyu Yang, Wentao Li et al.

Reverse k-nearest neighbor (RkNN) search returns all data points that regard a query vector as one of their k-nearest neighbors (kNNs). Existing RkNN methods typically follow a filter-and-verification framework: vectors near the query vector are first collected as candidates and then verified against their kNN-radius (i.e., the distance to their k-th nearest neighbor). However, existing methods face two key limitations in high-dimensional spaces. First, nearby vectors often do not belong to the query's true RkNN set, resulting in excessive candidate expansion overhead. Second, existing methods compute kNN-radius online during verification, incurring substantial query-processing cost. To address these limitations, we propose HRNN, a hybrid graph index for approximate RkNN search. (1) Rather than directly treating nearby vectors as RkNN candidates, HRNN uses them as proxy points based on the assumption that a query's RkNN results can often be discovered through the RkNN results of its nearby vectors. (2) To reduce verification cost, HRNN materializes high-fidelity kNN-radius offline, eliminating expensive online reconstruction while preserving accuracy. HRNN combines a navigation graph, a ranked KNN graph, and reverse-neighbor lists into a hybrid index that supports efficient proxy retrieval, candidate generation, and kNN-radius access. We also develop efficient index construction and append-only maintenance algorithms. Extensive experiments show that HRNN consistently outperforms existing methods, achieving up to one order of magnitude higher throughput. Moreover, HRNN scales to datasets containing up to 10 million high-dimensional vectors while supporting efficient dynamic index maintenance.

Jian Zhao, Xunhan Hu, Mingyu Yang et al.

Due to the partial observability and communication constraints in many multi-agent reinforcement learning (MARL) tasks, centralized training with decentralized execution (CTDE) has become one of the most widely used MARL paradigms. In CTDE, centralized information is dedicated to learning the allocation of the team reward with a mixing network, while the learning of individual Q-values is usually based on local observations. The insufficient utility of global observation will degrade performance in challenging environments. To this end, this work proposes a novel Centralized Teacher with Decentralized Student (CTDS) framework, which consists of a teacher model and a student model. Specifically, the teacher model allocates the team reward by learning individual Q-values conditioned on global observation, while the student model utilizes the partial observations to approximate the Q-values estimated by the teacher model. In this way, CTDS balances the full utilization of global observation during training and the feasibility of decentralized execution for online inference. Our CTDS framework is generic which is ready to be applied upon existing CTDE methods to boost their performance. We conduct experiments on a challenging set of StarCraft II micromanagement tasks to test the effectiveness of our method and the results show that CTDS outperforms the existing value-based MARL methods.

Wentao Zhang, Zeang Sheng, Mingyu Yang et al. · pku, tencent-ai

Recently, graph neural networks (GNNs) have shown prominent performance in graph representation learning by leveraging knowledge from both graph structure and node features. However, most of them have two major limitations. First, GNNs can learn higher-order structural information by stacking more layers but can not deal with large depth due to the over-smoothing issue. Second, it is not easy to apply these methods on large graphs due to the expensive computation cost and high memory usage. In this paper, we present node-adaptive feature smoothing (NAFS), a simple non-parametric method that constructs node representations without parameter learning. NAFS first extracts the features of each node with its neighbors of different hops by feature smoothing, and then adaptively combines the smoothed features. Besides, the constructed node representation can further be enhanced by the ensemble of smoothed features extracted via different smoothing strategies. We conduct experiments on four benchmark datasets on two different application scenarios: node clustering and link prediction. Remarkably, NAFS with feature ensemble outperforms the state-of-the-art GNNs on these tasks and mitigates the aforementioned two limitations of most learning-based GNN counterparts.

Jian Zhao, Youpeng Zhao, Weixun Wang et al.

Multi-agent reinforcement learning is difficult to be applied in practice, which is partially due to the gap between the simulated and real-world scenarios. One reason for the gap is that the simulated systems always assume that the agents can work normally all the time, while in practice, one or more agents may unexpectedly "crash" during the coordination process due to inevitable hardware or software failures. Such crashes will destroy the cooperation among agents, leading to performance degradation. In this work, we present a formal formulation of a cooperative multi-agent reinforcement learning system with unexpected crashes. To enhance the robustness of the system to crashes, we propose a coach-assisted multi-agent reinforcement learning framework, which introduces a virtual coach agent to adjust the crash rate during training. We design three coaching strategies and the re-sampling strategy for our coach agent. To the best of our knowledge, this work is the first to study the unexpected crashes in the multi-agent system. Extensive experiments on grid-world and StarCraft II micromanagement tasks demonstrate the efficacy of adaptive strategy compared with the fixed crash rate strategy and curriculum learning strategy. The ablation study further illustrates the effectiveness of our re-sampling strategy.

Zheng Chen, Jian Zhao, Mingyu Yang et al.

In this work, we are dedicated to multi-target active object tracking (AOT), where there are multiple targets as well as multiple cameras in the environment. The goal is maximize the overall target coverage of all cameras. Previous work makes a strong assumption that each camera is fixed in a location and only allowed to rotate, which limits its application. In this work, we relax the setting by allowing all cameras to both move along the boundary lines and rotate. In our setting, the action space becomes much larger, which leads to much higher computational complexity to identify the optimal action. To this end, we propose to leverage the action selection from multi-agent reinforcement learning (MARL) network to prune the search tree of Monte Carlo Tree Search (MCTS) method, so as to find the optimal action more efficiently. Besides, we model the motion of the targets to predict the future position of the targets, which makes a better estimation of the future environment state in the MCTS process. We establish a multi-target 2D environment to simulate the sports games, and experimental results demonstrate that our method can effectively improve the target coverage.

Yangbin Yu, Mingyu Yang, Junyou Li et al.

Existing Large Language Model (LLM) agents struggle in interactive environments requiring long-horizon planning, primarily due to compounding errors when simulating future states. To address this, we propose ProAct, a framework that enables agents to internalize accurate lookahead reasoning through a two-stage training paradigm. First, we introduce Grounded LookAhead Distillation (GLAD), where the agent undergoes supervised fine-tuning on trajectories derived from environment-based search. By compressing complex search trees into concise, causal reasoning chains, the agent learns the logic of foresight without the computational overhead of inference-time search. Second, to further refine decision accuracy, we propose the Monte-Carlo Critic (MC-Critic), a plug-and-play auxiliary value estimator designed to enhance policy-gradient algorithms like PPO and GRPO. By leveraging lightweight environment rollouts to calibrate value estimates, MC-Critic provides a low-variance signal that facilitates stable policy optimization without relying on expensive model-based value approximation. Experiments on both stochastic (e.g., 2048) and deterministic (e.g., Sokoban) environments demonstrate that ProAct significantly improves planning accuracy. Notably, a 4B parameter model trained with ProAct outperforms all open-source baselines and rivals state-of-the-art closed-source models, while demonstrating robust generalization to unseen environments. The codes and models are available at https://github.com/GreatX3/ProAct

Kejia Zhang, Mingyu Yang, Stephen D. J. Lang et al.

African penguins (Spheniscus demersus) are an endangered species. Little is known regarding their underwater hunting strategies and associated predation success rates, yet this is essential for guiding conservation. Modern bio-logging technology has the potential to provide valuable insights, but manually analysing large amounts of data from animal-borne video recorders (AVRs) is time-consuming. In this paper, we publish an animal-borne underwater video dataset of penguins and introduce a ready-to-deploy deep learning system capable of robustly detecting penguins (mAP50@98.0%) and also instances of fish (mAP50@73.3%). We note that the detectors benefit explicitly from air-bubble learning to improve accuracy. Extending this detector towards a dual-stream behaviour recognition network, we also provide the first results for identifying predation behaviour in penguin underwater videos. Whilst results are promising, further work is required for useful applicability of predation behaviour detection in field scenarios. In summary, we provide a highly reliable underwater penguin detector, a fish detector, and a valuable first attempt towards an automated visual detection of complex behaviours in a marine predator. We publish the networks, the DivingWithPenguins video dataset, annotations, splits, and weights for full reproducibility and immediate usability by practitioners.

Sara Shoouri, Mingyu Yang, Zichen Fan et al.

Solving multiple visual tasks using individual models can be resource-intensive, while multi-task learning can conserve resources by sharing knowledge across different tasks. Despite the benefits of multi-task learning, such techniques can struggle with balancing the loss for each task, leading to potential performance degradation. We present a novel computation- and parameter-sharing framework that balances efficiency and accuracy to perform multiple visual tasks utilizing individually-trained single-task transformers. Our method is motivated by transfer learning schemes to reduce computational and parameter storage costs while maintaining the desired performance. Our approach involves splitting the tasks into a base task and the other sub-tasks, and sharing a significant portion of activations and parameters/weights between the base and sub-tasks to decrease inter-task redundancies and enhance knowledge sharing. The evaluation conducted on NYUD-v2 and PASCAL-context datasets shows that our method is superior to the state-of-the-art transformer-based multi-task learning techniques with higher accuracy and reduced computational resources. Moreover, our method is extended to video stream inputs, further reducing computational costs by efficiently sharing information across the temporal domain as well as the task domain. Our codes and models will be publicly available.

Zichuan Lin, Feiyu Liu, Yijun Yang et al.

Autonomous mobile GUI agents have attracted increasing attention along with the advancement of Multimodal Large Language Models (MLLMs). However, existing methods still suffer from inefficient learning from failed trajectories and ambiguous credit assignment under sparse rewards for long-horizon GUI tasks. To that end, we propose UI-Voyager, a novel two-stage self-evolving mobile GUI agent. In the first stage, we employ Rejection Fine-Tuning (RFT), which enables the continuous co-evolution of data and models in a fully autonomous loop. The second stage introduces Group Relative Self-Distillation (GRSD), which identifies critical fork points in group rollouts and constructs dense step-level supervision from successful trajectories to correct failed ones. Extensive experiments on AndroidWorld show that our 4B model achieves an 81.0% Pass@1 success rate, outperforming numerous recent baselines and exceeding human-level performance. Ablation and case studies further verify the effectiveness of GRSD. Our method represents a significant leap toward efficient, self-evolving, and high-performance mobile GUI automation without expensive manual data annotation.

Mingyu Yang, Junyou Li, Zhongbin Fang et al.

In recent years, Artificial Intelligence Generated Content (AIGC) has advanced from text-to-image generation to text-to-video and multimodal video synthesis. However, generating playable games presents significant challenges due to the stringent requirements for real-time interaction, high visual quality, and accurate simulation of game mechanics. Existing approaches often fall short, either lacking real-time capabilities or failing to accurately simulate interactive mechanics. To tackle the playability issue, we propose a novel method called \emph{PlayGen}, which encompasses game data generation, an autoregressive DiT-based diffusion model, and a comprehensive playability-based evaluation framework. Validated on well-known 2D and 3D games, PlayGen achieves real-time interaction, ensures sufficient visual quality, and provides accurate interactive mechanics simulation. Notably, these results are sustained even after over 1000 frames of gameplay on an NVIDIA RTX 2060 GPU. Our code is publicly available: https://github.com/GreatX3/Playable-Game-Generation. Our playable demo generated by AI is: http://124.156.151.207.

Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge, Mingyu Yang et al.

Cross-tokenizer knowledge distillation allows a student model to learn from teachers with incompatible vocabularies. Prior work operates on hidden states or logits; the latter is preferred as a drop-in replacement requiring no auxiliary components. Logit-based methods either use only the correct-token probability, missing the full 'dark knowledge' in the teacher's distribution, or operate on the full output distribution, relying on strict token partitioning and/or unprincipled heuristic ranking. We identify two key shortcomings of full-distribution, logit-based methods: (i) an uncommon-token failure, where critical tokens fall into the unmatched subset (e.g., Llama's 1100 multi-digit numerals under digit-splitting Qwen supervision) and are suppressed during training, reducing GSM8k from 12.89 to 2.56 compared to same-tokenizer KD from a weaker teacher; and (ii) over-conservative matching, where strict 1-to-1 matching excludes near-equivalent tokens across surface forms. These failures require distinct remedies: eliminating the partition when critical tokens are misaligned, and refining it when alignment is reliable. We propose X-Token, an approach with two complementary loss formulations targeting these issues. P-KL removes partitioning and aligns the student's distribution with the teacher's via a sparse projection matrix W (initialized from tokenizer-level string rules) to address the uncommon-token failure. H-KL retains the hybrid form while relaxing matching to align each student token with its top-ranked teacher mapping under W. Both objectives share W and extend naturally to multiple teachers. Empirically, on Llama-3.2-1B, X-Token outperforms the current state of the art GOLD by +3.82 average points with a Qwen3-4B teacher and by +0.5 with a Phi-4-Mini teacher. Further, a two-teacher setup (Phi-4-mini + Llama-3B) improves over single-teacher distillation by +1.3 points.

Ivannia Gomez Moreno, Yi Yao, Ye Tian et al.

LiDAR semantic segmentation plays a pivotal role in 3D scene understanding for edge applications such as autonomous driving. However, significant challenges remain for real-world deployments, particularly for on-device post-deployment adaptation. Real-world environments can shift as the system navigates through different locations, leading to substantial performance degradation without effective and timely model adaptation. Furthermore, edge systems operate under strict computational and energy constraints, making it infeasible to adapt conventional segmentation models (based on large neural networks) directly on-device. To address the above challenges, we introduce HyperLiDAR, the first lightweight, post-deployment LiDAR segmentation framework based on Hyperdimensional Computing (HDC). The design of HyperLiDAR fully leverages the fast learning and high efficiency of HDC, inspired by how the human brain processes information. To further improve the adaptation efficiency, we identify the high data volume per scan as a key bottleneck and introduce a buffer selection strategy that focuses learning on the most informative points. We conduct extensive evaluations on two state-of-the-art LiDAR segmentation benchmarks and two representative devices. Our results show that HyperLiDAR outperforms or achieves comparable adaptation performance to state-of-the-art segmentation methods, while achieving up to a 13.8x speedup in retraining.

Yue Jiang, Mingyu Yang, Liuyuxin Yang et al.

Recent advances in generative motion synthesis have enabled the production of realistic human motions from diverse input modalities. However, synthesizing compound actions from texts, which integrate multiple concurrent actions into coherent full-body sequences, remains a major challenge. We identify two key limitations in current text-to-motion diffusion models: (i) catastrophic neglect, where earlier actions are overwritten by later ones due to improper handling of temporal information, and (ii) attention collapse, which arises from excessive feature fusion in cross-attention mechanisms. As a result, existing approaches often depend on overly detailed textual descriptions (e.g., raising right hand), explicit body-part specifications (e.g., editing the upper body), or the use of large language models (LLMs) for body-part interpretation. These strategies lead to deficient semantic representations of physical structures and kinematic mechanisms, limiting the ability to incorporate natural behaviors such as greeting while walking. To address these issues, we propose the Motion-Adapter, a plug-and-play module that guides text-to-motion diffusion models in generating compound actions by computing decoupled cross-attention maps, which serve as structural masks during the denoising process. Extensive experiments demonstrate that our method consistently produces more faithful and coherent compound motions across diverse textual prompts, surpassing state-of-the-art approaches.

Guihong Li, Mehdi Rezagholizadeh, Mingyu Yang et al.

Multi-head latent attention (MLA) is designed to optimize KV cache memory through low-rank key-value joint compression. Rather than caching keys and values separately, MLA stores their compressed latent representations, reducing memory overhead while maintaining the performance. While MLA improves memory efficiency without compromising language model accuracy, its major limitation lies in its integration during the pre-training phase, requiring models to be trained from scratch. This raises a key question: can we use MLA's benefits fully or partially in models that have already been pre-trained with different attention mechanisms? In this paper, we propose X-EcoMLA to deploy post training distillation to enable the upcycling of Transformer-based attention into an efficient hybrid MLA variant through lightweight post-training adaptation, bypassing the need for extensive pre-training. We demonstrate that leveraging the dark knowledge of a well-trained model can enhance training accuracy and enable extreme KV cache compression in MLA without compromising model performance. The experimental results show that our proposed method can effectively compress the KV cache while preserving the performance on the benchmarks; specifically, for Llama3.2-1B-Instruct baseline, a 6.4x compression achieves the same average score by using only 3.6B training tokens and 70 GPU hours on AMD MI300, whereas a 10.6x compression have less than 0.1% average score drop with 7B training tokens and 140 GPU hours. The code for this work is available at https://github.com/AMD-AGI/AMD-Hybrid-Models.

Mingyu Yang, Hun-Seok Kim

We present a novel adaptive deep joint source-channel coding (JSCC) scheme for wireless image transmission. The proposed scheme supports multiple rates using a single deep neural network (DNN) model and learns to dynamically control the rate based on the channel condition and image contents. Specifically, a policy network is introduced to exploit the tradeoff space between the rate and signal quality. To train the policy network, the Gumbel-Softmax trick is adopted to make the policy network differentiable and hence the whole JSCC scheme can be trained end-to-end. To the best of our knowledge, this is the first deep JSCC scheme that can automatically adjust its rate using a single network model. Experiments show that our scheme successfully learns a reasonable policy that decreases channel bandwidth utilization for high SNR scenarios or simple image contents. For an arbitrary target rate, our rate-adaptive scheme using a single model achieves similar performance compared to an optimized model specifically trained for that fixed target rate. To reproduce our results, we make the source code publicly available at https://github.com/mingyuyng/Dynamic_JSCC.

Duyi Pan, Tianao Lou, Xin Li et al.

Large Language Models (LLMs) exhibit hallucinations in knowledge-intensive tasks. Graph-based retrieval augmented generation (RAG) has emerged as a promising solution, yet existing approaches suffer from fundamental recall and precision limitations when operating over black-box knowledge graphs -- graphs whose schema and structure are unknown in advance. We identify three core challenges that cause recall loss (semantic instantiation uncertainty and structural path uncertainty) and precision loss (evidential comparison uncertainty). To address these challenges, we formalize the retrieval task as the Optimal Informative Subgraph Retrieval (OISR) problem -- a variant of Group Steiner Tree -- and prove it to be NP-hard and APX-hard. We propose BubbleRAG, a training-free pipeline that systematically optimizes for both recall and precision through semantic anchor grouping, heuristic bubble expansion to discover candidate evidence graphs (CEGs), composite ranking, and reasoning-aware expansion. Experiments on multi-hop QA benchmarks demonstrate that BubbleRAG achieves state-of-the-art results, outperforming strong baselines in both F1 and accuracy while remaining plug-and-play.

Mingyu Yang, Bowen Liu, Boyang Wang et al.

Deep learning-based joint source-channel coding (deep JSCC) has been demonstrated to be an effective approach for wireless image transmission. Nevertheless, most existing work adopts an autoencoder framework to optimize conventional criteria such as Mean Squared Error (MSE) and Structural Similarity Index (SSIM) which do not suffice to maintain the perceptual quality of reconstructed images. Such an issue is more prominent under stringent bandwidth constraints or low signal-to-noise ratio (SNR) conditions. To tackle this challenge, we propose DiffJSCC, a novel framework that leverages the prior knowledge of the pre-trained Statble Diffusion model to produce high-realism images via the conditional diffusion denoising process. Our DiffJSCC first extracts multimodal spatial and textual features from the noisy channel symbols in the generation phase. Then, it produces an initial reconstructed image as an intermediate representation to aid robust feature extraction and a stable training process. In the following diffusion step, DiffJSCC uses the derived multimodal features, together with channel state information such as the signal-to-noise ratio (SNR), as conditions to guide the denoising diffusion process, which converts the initial random noise to the final reconstruction. DiffJSCC employs a novel control module to fine-tune the Stable Diffusion model and adjust it to the multimodal conditions. Extensive experiments on diverse datasets reveal that our method significantly surpasses prior deep JSCC approaches on both perceptual metrics and downstream task performance, showcasing its ability to preserve the semantics of the original transmitted images. Notably, DiffJSCC can achieve highly realistic reconstructions for 768x512 pixel Kodak images with only 3072 symbols (<0.008 symbols per pixel) under 1dB SNR channels.

Parsa Ashrafi Fashi, Utkarsh Saxena, Mehdi Rezagholizadeh et al.

Hybrid sequence models that combine efficient Transformer components with linear sequence modeling blocks are a promising alternative to pure Transformers, but most are still pretrained from scratch and therefore fail to reuse existing Transformer checkpoints. We study upcycling as a practical path to convert pretrained Transformer LLMs into hybrid architectures while preserving short-context quality and improving long-context capability. We call our solution \emph{HyLo} (HYbrid LOng-context): a long-context upcycling recipe that combines architectural adaptation with efficient Transformer blocks, Multi-Head Latent Attention (MLA), and linear blocks (Mamba2 or Gated DeltaNet), together with staged long-context training and teacher-guided distillation for stable optimization. HyLo extends usable context length by up to $32\times$ through efficient post-training and reduces KV-cache memory by more than $90\%$, enabling up to 2M-token prefill and decoding in our \texttt{vLLM} inference stack, while comparable Llama baselines run out of memory beyond 64K context. Across 1B- and 3B-scale settings (Llama- and Qwen-based variants), HyLo delivers consistently strong short- and long-context performance and significantly outperforms state-of-the-art upcycled hybrid baselines on long-context evaluations such as RULER. Notably, at similar scale, HyLo-Qwen-1.7B trained on only 10B tokens significantly outperforms JetNemotron (trained on 400B tokens) on GSM8K, Lm-Harness common sense reasoning and RULER-64K.

Seungbeom Choi, Jeonghoe Goo, Eunjoo Jeon et al.

We propose ELIS, a serving system for Large Language Models (LLMs) featuring an Iterative Shortest Remaining Time First (ISRTF) scheduler designed to efficiently manage inference tasks with the shortest remaining tokens. Current LLM serving systems often employ a first-come-first-served scheduling strategy, which can lead to the "head-of-line blocking" problem. To overcome this limitation, it is necessary to predict LLM inference times and apply a shortest job first scheduling strategy. However, due to the auto-regressive nature of LLMs, predicting the inference latency is challenging. ELIS addresses this challenge by training a response length predictor for LLMs using the BGE model, an encoder-based state-of-the-art model. Additionally, we have devised the ISRTF scheduling strategy, an optimization of shortest remaining time first tailored to existing LLM iteration batching. To evaluate our work in an industrial setting, we simulate streams of requests based on our study of real-world user LLM serving trace records. Furthermore, we implemented ELIS as a cloud-native scheduler system on Kubernetes to evaluate its performance in production environments. Our experimental results demonstrate that ISRTF reduces the average job completion time by up to 19.6%.

Mingyu Yang, Mehdi Rezagholizadeh, Guihong Li et al.

With the growing demand for deploying large language models (LLMs) across diverse applications, improving their inference efficiency is crucial for sustainable and democratized access. However, retraining LLMs to meet new user-specific requirements is prohibitively expensive and environmentally unsustainable. In this work, we propose a practical and scalable alternative: composing efficient hybrid language models from existing pre-trained models. Our approach, Zebra-Llama, introduces a family of 1B, 3B, and 8B hybrid models by combining State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers, using a refined initialization and post-training pipeline to efficiently transfer knowledge from pre-trained Transformers. Zebra-Llama achieves Transformer-level accuracy with near-SSM efficiency using only 7-11B training tokens (compared to trillions of tokens required for pre-training) and an 8B teacher. Moreover, Zebra-Llama dramatically reduces KV cache size -down to 3.9%, 2%, and 2.73% of the original for the 1B, 3B, and 8B variants, respectively-while preserving 100%, 100%, and >97% of average zero-shot performance on LM Harness tasks. Compared to models like MambaInLLaMA, X-EcoMLA, Minitron, and Llamba, Zebra-Llama consistently delivers competitive or superior accuracy while using significantly fewer tokens, smaller teachers, and vastly reduced KV cache memory. Notably, Zebra-Llama-8B surpasses Minitron-8B in few-shot accuracy by 7% while using 8x fewer training tokens, over 12x smaller KV cache, and a smaller teacher (8B vs. 15B). It also achieves 2.6x-3.8x higher throughput (tokens/s) than MambaInLlama up to a 32k context length. We will release code and model checkpoints upon acceptance.

Mingyu Yang, Wentao Li, Wei Wang

Hybrid queries combining high-dimensional vector similarity search with spatio-temporal filters are increasingly critical for modern retrieval-augmented generation (RAG) systems. Existing systems typically handle these workloads by nesting vector indices within low-dimensional spatial structures, such as R-trees. However, this decoupled architecture fragments the vector space, forcing the query engine to invoke multiple disjoint sub-indices per query. This fragmentation destroys graph routing connectivity, incurs severe traversal overhead, and struggles to optimize for complex spatial boundaries. In this paper, we propose CubeGraph, a novel indexing framework designed to natively integrate vector search with arbitrary spatial constraints. CubeGraph partitions the spatial domain using a hierarchical grid, maintaining modular vector graphs within each cell. During query execution, CubeGraph dynamically stitches together adjacent cube-level indices on the fly whenever their spatial cells intersect with the query filter. This dynamic graph integration restores global connectivity, enabling a unified, single-pass nearest-neighbor traversal that eliminates the overhead of fragmented sub-index invocations. Extensive evaluations on real-world datasets demonstrate that CubeGraph significantly outperforms state-of-the-art baselines, offering superior query execution performance, scalability, and flexibility for complex hybrid workloads.

Deheng Ye, Fangyun Zhou, Jiacheng Lv et al.

We present Yan, a foundational framework for interactive video generation, covering the entire pipeline from simulation and generation to editing. Specifically, Yan comprises three core modules. AAA-level Simulation: We design a highly-compressed, low-latency 3D-VAE coupled with a KV-cache-based shift-window denoising inference process, achieving real-time 1080P/60FPS interactive simulation. Multi-Modal Generation: We introduce a hierarchical autoregressive caption method that injects game-specific knowledge into open-domain multi-modal video diffusion models (VDMs), then transforming the VDM into a frame-wise, action-controllable, real-time infinite interactive video generator. Notably, when the textual and visual prompts are sourced from different domains, the model demonstrates strong generalization, allowing it to blend and compose the style and mechanics across domains flexibly according to user prompts. Multi-Granularity Editing: We propose a hybrid model that explicitly disentangles interactive mechanics simulation from visual rendering, enabling multi-granularity video content editing during interaction through text. Collectively, Yan offers an integration of these modules, pushing interactive video generation beyond isolated capabilities toward a comprehensive AI-driven interactive creation paradigm, paving the way for the next generation of creative tools, media, and entertainment. The project page is: https://greatx3.github.io/Yan/.

Mingyu Yang, Jitong Lu, Hun-Seok Kim

Multi-modal 3D semantic segmentation is vital for applications such as autonomous driving and virtual reality (VR). To effectively deploy these models in real-world scenarios, it is essential to employ cross-domain adaptation techniques that bridge the gap between training data and real-world data. Recently, self-training with pseudo-labels has emerged as a predominant method for cross-domain adaptation in multi-modal 3D semantic segmentation. However, generating reliable pseudo-labels necessitates stringent constraints, which often result in sparse pseudo-labels after pruning. This sparsity can potentially hinder performance improvement during the adaptation process. We propose an image-guided pseudo-label enhancement approach that leverages the complementary 2D prior knowledge from the Segment Anything Model (SAM) to introduce more reliable pseudo-labels, thereby boosting domain adaptation performance. Specifically, given a 3D point cloud and the SAM masks from its paired image data, we collect all 3D points covered by each SAM mask that potentially belong to the same object. Then our method refines the pseudo-labels within each SAM mask in two steps. First, we determine the class label for each mask using majority voting and employ various constraints to filter out unreliable mask labels. Next, we introduce Geometry-Aware Progressive Propagation (GAPP) which propagates the mask label to all 3D points within the SAM mask while avoiding outliers caused by 2D-3D misalignment. Experiments conducted across multiple datasets and domain adaptation scenarios demonstrate that our proposed method significantly increases the quantity of high-quality pseudo-labels and enhances the adaptation performance over baseline methods.

Mingyu Yang, Jae-Young Choi, Kihyo Moon et al.

Speculative decoding accelerates large language model inference, but its reliance on a fixed speculation length is suboptimal in large-batch serving environments with diverse requests. This paper explores a new direction for dynamic adaptation by investigating a novel class of post-hoc, diagnostic signals. We propose Dynamic Speculative Decoding Engine (DSDE), a training-free framework built on two primary components: (1) a predictive signal based on the variance of the Kullback-Leibler (KLD) divergence, which diagnoses the generation's regional stability, and (2) an adaptive speculation length cap to mitigate the straggler problem in per-sequence decoding. Experiments demonstrate the potential of using KLD-based stability signals for dynamic adaptation. An algorithm guided by these signals achieves end-to-end latency competitive with leading baselines and exhibits superior robustness across diverse workloads. This robustness is particularly valuable in challenging low-acceptance-rate regimes, where the proposed signal maintains its diagnostic utility. Collectively, these findings validate post-hoc signals as a valuable component for building more robust and intelligent LLM inference systems, and highlight a promising direction for future research on dynamic speculation length adaptation.

Zeyu Fang, Jian Zhao, Mingyu Yang et al.

Active Multi-Object Tracking (AMOT) is a task where cameras are controlled by a centralized system to adjust their poses automatically and collaboratively so as to maximize the coverage of targets in their shared visual field. In AMOT, each camera only receives partial information from its observation, which may mislead cameras to take locally optimal action. Besides, the global goal, i.e., maximum coverage of objects, is hard to be directly optimized. To address the above issues, we propose a coordinate-aligned multi-camera collaboration system for AMOT. In our approach, we regard each camera as an agent and address AMOT with a multi-agent reinforcement learning solution. To represent the observation of each agent, we first identify the targets in the camera view with an image detector, and then align the coordinates of the targets in 3D environment. We define the reward of each agent based on both global coverage as well as four individual reward terms. The action policy of the agents is derived with a value-based Q-network. To the best of our knowledge, we are the first to study the AMOT task. To train and evaluate the efficacy of our system, we build a virtual yet credible 3D environment, named "Soccer Court", to mimic the real-world AMOT scenario. The experimental results show that our system achieves a coverage of 71.88%, outperforming the baseline method by 8.9%.

Jian Zhao, Mingyu Yang, Youpeng Zhao et al.

In cooperative multi-agent tasks, a team of agents jointly interact with an environment by taking actions, receiving a team reward and observing the next state. During the interactions, the uncertainty of environment and reward will inevitably induce stochasticity in the long-term returns and the randomness can be exacerbated with the increasing number of agents. However, such randomness is ignored by most of the existing value-based multi-agent reinforcement learning (MARL) methods, which only model the expectation of Q-value for both individual agents and the team. Compared to using the expectations of the long-term returns, it is preferable to directly model the stochasticity by estimating the returns through distributions. With this motivation, this work proposes a novel value-based MARL framework from a distributional perspective, \emph{i.e.}, parameterizing value function via \underline{M}ixture of \underline{C}ategorical distributions for MARL. Specifically, we model both individual Q-values and global Q-value with categorical distribution. To integrate categorical distributions, we define five basic operations on the distribution, which allow the generalization of expected value function factorization methods (\emph{e.g.}, VDN and QMIX) to their MCMARL variants. We further prove that our MCMARL framework satisfies \emph{Distributional-Individual-Global-Max} (DIGM) principle with respect to the expectation of distribution, which guarantees the consistency between joint and individual greedy action selections in the global Q-value and individual Q-values. Empirically, we evaluate MCMARL on both a stochastic matrix game and a challenging set of StarCraft II micromanagement tasks, showing the efficacy of our framework.

Wentao Zhang, Mingyu Yang, Zeang Sheng et al.

Recent works reveal that feature or label smoothing lies at the core of Graph Neural Networks (GNNs). Concretely, they show feature smoothing combined with simple linear regression achieves comparable performance with the carefully designed GNNs, and a simple MLP model with label smoothing of its prediction can outperform the vanilla GCN. Though an interesting finding, smoothing has not been well understood, especially regarding how to control the extent of smoothness. Intuitively, too small or too large smoothing iterations may cause under-smoothing or over-smoothing and can lead to sub-optimal performance. Moreover, the extent of smoothness is node-specific, depending on its degree and local structure. To this end, we propose a novel algorithm called node-dependent local smoothing (NDLS), which aims to control the smoothness of every node by setting a node-specific smoothing iteration. Specifically, NDLS computes influence scores based on the adjacency matrix and selects the iteration number by setting a threshold on the scores. Once selected, the iteration number can be applied to both feature smoothing and label smoothing. Experimental results demonstrate that NDLS enjoys high accuracy -- state-of-the-art performance on node classifications tasks, flexibility -- can be incorporated with any models, scalability and efficiency -- can support large scale graphs with fast training.

Mingyu Yang, Chenghong Bian, Hun-Seok Kim

We present a deep learning based joint source channel coding (JSCC) scheme for wireless image transmission over multipath fading channels with non-linear signal clipping. The proposed encoder and decoder use convolutional neural networks (CNN) and directly map the source images to complex-valued baseband samples for orthogonal frequency division multiplexing (OFDM) transmission. The proposed model-driven machine learning approach eliminates the need for separate source and channel coding while integrating an OFDM datapath to cope with multipath fading channels. The end-to-end JSCC communication system combines trainable CNN layers with non-trainable but differentiable layers representing the multipath channel model and OFDM signal processing blocks. Our results show that injecting domain expert knowledge by incorporating OFDM baseband processing blocks into the machine learning framework significantly enhances the overall performance compared to an unstructured CNN. Our method outperforms conventional schemes that employ state-of-the-art but separate source and channel coding such as BPG and LDPC with OFDM. Moreover, our method is shown to be robust against non-linear signal clipping in OFDM for various channel conditions that do not match the model parameter used during the training.

Mingyu Yang, Roger Hsiao, Gordy Carichner et al.

Details of Monarch butterfly migration from the U.S. to Mexico remain a mystery due to lack of a proper localization technology to accurately localize and track butterfly migration. In this paper, we propose a deep learning based butterfly localization algorithm that can estimate a butterfly's daily location by analyzing a light and temperature sensor data log continuously obtained from an ultra-low power, mm-scale sensor attached to the butterfly. To train and test the proposed neural network based multi-sensor fusion localization algorithm, we collected over 1500 days of real world sensor measurement data with 82 volunteers all over the U.S. The proposed algorithm exhibits a mean absolute error of <1.5 degree in latitude and <0.5 degree in longitude Earth coordinate, satisfying our target goal for the Monarch butterfly migration study.