MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection

KV cache has become a de facto technique for the inference of large language models (LLMs), where tensors of shape (layer number, head number, sequence length, feature dimension) are introduced to cache historical information for self-attention. As the size of the model and data grows, the KV cache can quickly become a bottleneck within the system in both storage and memory transfer. To address this, prior studies usually focus on the first three axes of the cache tensors for compression. This paper supplements them, focusing on the feature dimension axis, by utilizing low-rank projection matrices to transform the cache features into spaces with reduced dimensions. We begin by investigating the canonical orthogonal projection method for data compression through principal component analysis (PCA). We observe the issue with PCA projection where significant performance degradation is observed at low compression rates. To bridge the gap, we propose to directly tune the orthogonal projection matrices with a distillation objective using an elaborate Matryoshka training strategy. After training, we adaptively search for the optimal compression rates for various layers and heads given varying compression budgets. Compared to previous works, our method can easily embrace pre-trained LLMs and hold a smooth tradeoff between performance and compression rate. We empirically witness the high data efficiency of our training procedure and find that our method can sustain over 90% performance with an average KV cache compression rate of 60% (and up to 75% in certain extreme scenarios) for popular LLMs like LLaMA2-7B-base and Mistral-7B-v0.3-base.

T2V-Turbo-v2: Enhancing Video Generation Model Post-Training through Data, Reward, and Conditional Guidance Design

In this paper, we focus on enhancing a diffusion-based text-to-video (T2V) model during the post-training phase by distilling a highly capable consistency model from a pretrained T2V model. Our proposed method, T2V-Turbo-v2, introduces a significant advancement by integrating various supervision signals, including high-quality training data, reward model feedback, and conditional guidance, into the consistency distillation process. Through comprehensive ablation studies, we highlight the crucial importance of tailoring datasets to specific learning objectives and the effectiveness of learning from diverse reward models for enhancing both the visual quality and text-video alignment. Additionally, we highlight the vast design space of conditional guidance strategies, which centers on designing an effective energy function to augment the teacher ODE solver. We demonstrate the potential of this approach by extracting motion guidance from the training datasets and incorporating it into the ODE solver, showcasing its effectiveness in improving the motion quality of the generated videos with the improved motion-related metrics from VBench and T2V-CompBench. Empirically, our T2V-Turbo-v2 establishes a new state-of-the-art result on VBench, with a Total score of 85.13, surpassing proprietary systems such as Gen-3 and Kling.

An Efficient and Generalizable Symbolic Regression Method for Time Series Analysis

Time series analysis and prediction methods currently excel in quantitative analysis, offering accurate future predictions and diverse statistical indicators, but generally falling short in elucidating the underlying evolution patterns of time series. To gain a more comprehensive understanding and provide insightful explanations, we utilize symbolic regression techniques to derive explicit expressions for the non-linear dynamics in the evolution of time series variables. However, these techniques face challenges in computational efficiency and generalizability across diverse real-world time series data. To overcome these challenges, we propose \textbf{N}eural-\textbf{E}nhanced \textbf{Mo}nte-Carlo \textbf{T}ree \textbf{S}earch (NEMoTS) for time series. NEMoTS leverages the exploration-exploitation balance of Monte-Carlo Tree Search (MCTS), significantly reducing the search space in symbolic regression and improving expression quality. Furthermore, by integrating neural networks with MCTS, NEMoTS not only capitalizes on their superior fitting capabilities to concentrate on more pertinent operations post-search space reduction, but also replaces the complex and time-consuming simulation process, thereby substantially improving computational efficiency and generalizability in time series analysis. NEMoTS offers an efficient and comprehensive approach to time series analysis. Experiments with three real-world datasets demonstrate NEMoTS's significant superiority in performance, efficiency, reliability, and interpretability, making it well-suited for large-scale real-world time series data.

Adaptive Two-Stage Cloud Resource Scaling via Hierarchical Multi-Indicator Forecasting and Bayesian Decision-Making

The surging demand for cloud computing resources, driven by the rapid growth of sophisticated large-scale models and data centers, underscores the critical importance of efficient and adaptive resource allocation. As major tech enterprises deploy massive infrastructures with thousands of GPUs, existing cloud platforms still struggle with low resource utilization due to key challenges: capturing hierarchical indicator structures, modeling non-Gaussian distributions, and decision-making under uncertainty. To address these challenges, we propose HRAMONY, an adaptive Hierarchical Attention-based Resource Modeling and Decision-Making System. HARMONY combines hierarchical multi-indicator distribution forecasting and uncertainty-aware Bayesian decision-making. It introduces a novel hierarchical attention mechanism that comprehensively models complex inter-indicator dependencies, enabling accurate predictions that can adapt to evolving environment states. By transforming Gaussian projections into adaptive non-Gaussian distributions via Normalizing Flows. Crucially, HARMONY leverages the full predictive distributions in an adaptive Bayesian process, proactively incorporating uncertainties to optimize resource allocation while robustly meeting SLA constraints under varying conditions. Extensive evaluations across four large-scale cloud datasets demonstrate HARMONY's state-of-the-art performance, significantly outperforming nine established methods. A month-long real-world deployment validated HARMONY's substantial practical impact, realizing over 35,000 GPU hours in savings and translating to $100K+ in cost reduction, showcasing its remarkable economic value through adaptive, uncertainty-aware scaling. Our code is available at https://github.com/Floating-LY/HARMONY1.

Towards Personalized Federated Multi-scenario Multi-task Recommendation

In modern recommender system applications, such as e-commerce, predicting multiple targets like click-through rate (CTR) and post-view click-through \& conversion rate (CTCVR) is common. Multi-task recommender systems are gaining traction in research and practical use. Existing multi-task recommender systems tackle diverse business scenarios, merging and modeling these scenarios unlocks shared knowledge to boost overall performance. As new and more complex real-world recommendation scenarios have emerged, data privacy issues make it difficult to train a single global multi-task recommendation model that processes multiple separate scenarios. In this paper, we propose a novel framework for personalized federated multi-scenario multi-task recommendation, called PF-MSMTrec. We assign each scenario to a dedicated client, with each client utilizing the Mixture-of-Experts (MMoE) structure. Our proposed method aims to tackle the unique challenge posed by multiple optimization conflicts in this setting. We introduce a bottom-up joint learning mechanism. Firstly, we design a parameter template to decouple the parameters of the expert network. Thus, scenario parameters are shared knowledge for federated parameter aggregation, while task-specific parameters are personalized local parameters. Secondly, we conduct personalized federated learning for the parameters of each expert network through a federated communication round, utilizing three modules: federated batch normalization, conflict coordination, and personalized aggregation. Finally, we perform another round of personalized federated parameter aggregation on the task tower network to obtain the prediction results for multiple tasks. We conduct extensive experiments on two public datasets, and the results demonstrate that our proposed method surpasses state-of-the-art methods.

ATFNet: Adaptive Time-Frequency Ensembled Network for Long-term Time Series Forecasting

The intricate nature of time series data analysis benefits greatly from the distinct advantages offered by time and frequency domain representations. While the time domain is superior in representing local dependencies, particularly in non-periodic series, the frequency domain excels in capturing global dependencies, making it ideal for series with evident periodic patterns. To capitalize on both of these strengths, we propose ATFNet, an innovative framework that combines a time domain module and a frequency domain module to concurrently capture local and global dependencies in time series data. Specifically, we introduce Dominant Harmonic Series Energy Weighting, a novel mechanism for dynamically adjusting the weights between the two modules based on the periodicity of the input time series. In the frequency domain module, we enhance the traditional Discrete Fourier Transform (DFT) with our Extended DFT, designed to address the challenge of discrete frequency misalignment. Additionally, our Complex-valued Spectrum Attention mechanism offers a novel approach to discern the intricate relationships between different frequency combinations. Extensive experiments across multiple real-world datasets demonstrate that our ATFNet framework outperforms current state-of-the-art methods in long-term time series forecasting.

GROUNDHOG: Grounding Large Language Models to Holistic Segmentation

Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases.

Mixture of Link Predictors

Link prediction, which aims to forecast unseen connections in graphs, is a fundamental task in graph machine learning. Heuristic methods, leveraging a range of different pairwise measures such as common neighbors and shortest paths, often rival the performance of vanilla Graph Neural Networks (GNNs). Therefore, recent advancements in GNNs for link prediction (GNN4LP) have primarily focused on integrating one or a few types of pairwise information. In this work, we reveal that different node pairs within the same dataset necessitate varied pairwise information for accurate prediction and models that only apply the same pairwise information uniformly could achieve suboptimal performance. As a result, we propose a simple mixture of experts model Link-MoE for link prediction. Link-MoE utilizes various GNNs as experts and strategically selects the appropriate expert for each node pair based on various types of pairwise information. Experimental results across diverse real-world datasets demonstrate substantial performance improvement from Link-MoE. Notably, Link-MoE achieves a relative improvement of 18.82\% on the MRR metric for the Pubmed dataset and 10.8\% on the Hits@100 metric for the ogbl-ppa dataset, compared to the best baselines.

Pareto-based Multi-Objective Recommender System with Forgetting Curve

Recommender systems with cascading architecture play an increasingly significant role in online recommendation platforms, where the approach to dealing with negative feedback is a vital issue. For instance, in short video platforms, users tend to quickly slip away from candidates that they feel aversive, and recommender systems are expected to receive these explicit negative feedbacks and make adjustments to avoid these recommendations. Considering recency effect in memories, we propose a forgetting model based on Ebbinghaus Forgetting Curve to cope with negative feedback. In addition, we introduce a Pareto optimization solver to guarantee a better trade-off between recency and model performance. In conclusion, we propose Pareto-based Multi-Objective Recommender System with forgetting curve (PMORS), which can be applied to any multi-objective recommendation and show sufficiently superiority when facing explicit negative feedback. We have conducted evaluations of PMORS and achieved favorable outcomes in short-video scenarios on both public dataset and industrial dataset. After being deployed on an online short video platform named WeChat Channels in May, 2023, PMORS has not only demonstrated promising results for both consistency and recency but also achieved an improvement of up to +1.45% GMV.

Planning as In-Painting: A Diffusion-Based Embodied Task Planning Framework for Environments under Uncertainty

Task planning for embodied AI has been one of the most challenging problems where the community does not meet a consensus in terms of formulation. In this paper, we aim to tackle this problem with a unified framework consisting of an end-to-end trainable method and a planning algorithm. Particularly, we propose a task-agnostic method named 'planning as in-painting'. In this method, we use a Denoising Diffusion Model (DDM) for plan generation, conditioned on both language instructions and perceptual inputs under partially observable environments. Partial observation often leads to the model hallucinating the planning. Therefore, our diffusion-based method jointly models both state trajectory and goal estimation to improve the reliability of the generated plan, given the limited available information at each step. To better leverage newly discovered information along the plan execution for a higher success rate, we propose an on-the-fly planning algorithm to collaborate with the diffusion-based planner. The proposed framework achieves promising performances in various embodied AI tasks, including vision-language navigation, object manipulation, and task planning in a photorealistic virtual environment. The code is available at: https://github.com/joeyy5588/planning-as-inpainting.