Continual Learning (CL) epitomizes an advanced training paradigm wherein prior data samples remain inaccessible during the acquisition of new tasks. Numerous investigations have delved into leveraging a pre-trained Vision Transformer (ViT) to enhance model efficacy in continual learning. Nonetheless, these approaches typically utilize a singular, static backbone, which inadequately adapts to novel tasks, particularly when engaging with diverse data domains, due to a substantial number of inactive parameters. This paper addresses this limitation by introducing an innovative Self-Controlled Dynamic Expansion Model (SCDEM), which orchestrates multiple distinct trainable pre-trained ViT backbones to furnish diverse and semantically enriched representations. Specifically, by employing the multi-backbone architecture as a shared module, the proposed SCDEM dynamically generates a new expert with minimal parameters to accommodate a new task. A novel Collaborative Optimization Mechanism (COM) is introduced to synergistically optimize multiple backbones by harnessing prediction signals from historical experts, thereby facilitating new task learning without erasing previously acquired knowledge. Additionally, a novel Feature Distribution Consistency (FDC) approach is proposed to align semantic similarity between previously and currently learned representations through an optimal transport distance-based mechanism, effectively mitigating negative knowledge transfer effects. Furthermore, to alleviate over-regularization challenges, this paper presents a novel Dynamic Layer-Wise Feature Attention Mechanism (DLWFAM) to autonomously determine the penalization intensity on each trainable representation layer. An extensive series of experiments have been conducted to evaluate the proposed methodology's efficacy, with empirical results corroborating that the approach attains state-of-the-art performance.
Time series analysis is crucial in fields like finance, transportation, and industry. However, traditional models often focus solely on temporal features, limiting their ability to capture underlying information. This paper proposes a novel time series multitask framework, called LTM, which integrates temporal features with textual descriptions to enhance analytical and predictive capabilities. LTM combines pre-trained time series model, large language model (LLM), and knowledge graph to tackle time series tasks, including forecasting, imputation, and anomaly detection. LTM achieves improved performance with a few trainable parameters. It is very efficient and practical. LTM encodes time series data into patches and enriches user-provided prompts using knowledge graphs to generate enhanced prompts. A novel feature fusion method embeds prompts into each patch encoding, which is processed by a frozen LLM, followed by a feature enhancement module and a time decoder module. During fine-tuning stage, cosine similarity between prompts and temporal patches is integrated into the loss function to boost performance. Experiments on benchmark datasets show that LTM significantly outperforms existing methods. It provides a robust and versatile solution for time series tasks.
Attention mechanisms have become a foundational component in diffusion models, significantly influencing their capacity across a wide range of generative and discriminative tasks. This paper presents a comprehensive survey of attention within diffusion models, systematically analysing its roles, design patterns, and operations across different modalities and tasks. We propose a unified taxonomy that categorises attention-related modifications into parts according to the structural components they affect, offering a clear lens through which to understand their functional diversity. In addition to reviewing architectural innovations, we examine how attention mechanisms contribute to performance improvements in diverse applications. We also identify current limitations and underexplored areas, and outline potential directions for future research. Our study provides valuable insights into the evolving landscape of diffusion models, with a particular focus on the integrative and ubiquitous role of attention.
One global model in federated learning (FL) might not be sufficient to serve many clients with non-IID tasks and distributions. While there has been advances in FL to train multiple global models for better personalization, they only provide limited choices to clients so local finetuning is still indispensable. In this paper, we propose a novel ``FedMerge'' approach that can create a personalized model per client by simply merging multiple global models with automatically optimized and customized weights. In FedMerge, a few global models can serve many non-IID clients, even without further local finetuning. We formulate this problem as a joint optimization of global models and the merging weights for each client. Unlike existing FL approaches where the server broadcasts one or multiple global models to all clients, the server only needs to send a customized, merged model to each client. Moreover, instead of periodically interrupting the local training and re-initializing it to a global model, the merged model aligns better with each client's task and data distribution, smoothening the local-global gap between consecutive rounds caused by client drift. We evaluate FedMerge on three different non-IID settings applied to different domains with diverse tasks and data types, in which FedMerge consistently outperforms existing FL approaches, including clustering-based and mixture-of-experts (MoE) based methods.
The inference-time resource costs of large language and vision models present a growing challenge in production deployments. We propose the use of foundation model programs, i.e., programs that can invoke foundation models with varying resource costs and performance, as an approach to this problem. Specifically, we present a method that translates a task into a program, then learns a policy for resource allocation that, on each input, selects foundation model "backends" for each program module. The policy uses smaller, cheaper backends to handle simpler subtasks, while allowing more complex subtasks to leverage larger, more capable models. We evaluate the method on two new "streaming" visual question-answering tasks in which a system answers a question on a sequence of inputs, receiving ground-truth feedback after each answer. Compared to monolithic multi-modal models, our implementation achieves up to 98% resource savings with minimal accuracy loss, demonstrating its potential for scalable and resource-efficient multi-modal inference.
In many estimation theory and statistical analysis problems, the true data model is unknown, or partially unknown. To describe the model generating the data, parameterized models of some degree are used. A question that arises is which model should be used to best approximate the true model, a.k.a. model selection. In the field of machine learning, it is encountered in the form of architecture types of neural networks, number of model parameters, etc. In this paper, we propose a new model selection criterion, based on the misspecified Cramer-Rao bound (MCRB) for mean-squared-error (MSE) performance. The criterion selects the model in which the bound on the estimated parameters MSE is the lowest, compared to other candidate models. Its goal is to minimize the MSE with-respect-to (w.r.t.) the model. The criterion is applied to the problems of direction-of-arrival (DOA) estimation under unknown clutter / interference, and spectrum estimation of auto-regressive (AR) model. It is shown to incorporate the bias-variance trade-off, and outperform the Akaike information criterion (AIC), finite-sample corrected AIC (AICc), and minimum description length (MDL) in terms of MSE performance.
Recently, diffusion-based recommendation methods have achieved impressive results. However, existing approaches predominantly treat each user's historical interactions as independent training samples, overlooking the potential of higher-order collaborative signals between users and items. Such signals, which encapsulate richer and more nuanced relationships, can be naturally captured using graph-based data structures. To address this limitation, we extend diffusion-based recommendation methods to the graph domain by directly modeling user-item bipartite graphs with diffusion models. This enables better modeling of the higher-order connectivity inherent in complex interaction dynamics. However, this extension introduces two primary challenges: (1) Noise Heterogeneity, where interactions are influenced by various forms of continuous and discrete noise, and (2) Relation Explosion, referring to the high computational costs of processing large-scale graphs. To tackle these challenges, we propose a Graph-based Diffusion Model for Collaborative Filtering (GDMCF). To address noise heterogeneity, we introduce a multi-level noise corruption mechanism that integrates both continuous and discrete noise, effectively simulating real-world interaction complexities. To mitigate relation explosion, we design a user-active guided diffusion process that selectively focuses on the most meaningful edges and active users, reducing inference costs while preserving the graph's topological integrity. Extensive experiments on three benchmark datasets demonstrate that GDMCF consistently outperforms state-of-the-art methods, highlighting its effectiveness in capturing higher-order collaborative signals and improving recommendation performance.
Persistence of excitation (PE) of the system input is a fundamental requirement for the successful operation of data-driven model predictive control, as it ensures that the input--output data contains sufficient information about the underlying system dynamics. Nonetheless, this property is usually assumed rather than guaranteed. This paper introduces a novel data-driven predictive control formulation that maintains persistence of excitation. The technical development that permits this is the characterization of the nonexciting input set i.e. the set of inputs that lead to loss of PE, and the consequent derivation of a pair of disjoint, linear inequality constraints on the input that, if satisfied, maintain PE. When used in the predictive control formulation, these constraints lead to a mixed-integer optimal control problem with a single binary variable or, equivalently, a pair of disjoint quadratic programming problems that can be efficiently and reliably solved in parallel. Numerical examples show how these constraints are able to maintain persistence of excitation on the input during the controller's operation.
The rapid advances in Foundation Models and agentic Artificial Intelligence are transforming multimedia analytics by enabling richer, more sophisticated interactions between humans and analytical systems. Existing conceptual models for visual and multimedia analytics, however, do not adequately capture the complexity introduced by these powerful AI paradigms. To bridge this gap, we propose a comprehensive multimedia analytics model specifically designed for the foundation model era. Building upon established frameworks from visual analytics, multimedia analytics, knowledge generation, analytic task definition, mixed-initiative guidance, and human-in-the-loop reinforcement learning, our model emphasizes integrated human-AI teaming based on visual analytics agents from both technical and conceptual perspectives. Central to the model is a seamless, yet explicitly separable, interaction channel between expert users and semi-autonomous analytical processes, ensuring continuous alignment between user intent and AI behavior. The model addresses practical challenges in sensitive domains such as intelligence analysis, investigative journalism, and other fields handling complex, high-stakes data. We illustrate through detailed case studies how our model facilitates deeper understanding and targeted improvement of multimedia analytics solutions. By explicitly capturing how expert users can optimally interact with and guide AI-powered multimedia analytics systems, our conceptual framework sets a clear direction for system design, comparison, and future research.
Chain-of-Thought (CoT) reasoning has demonstrated remarkable effectiveness in enhancing the reasoning abilities of large language models (LLMs). However, its efficiency remains a challenge due to the generation of excessive intermediate reasoning tokens, which introduce semantic redundancy and overly detailed reasoning steps. Moreover, computational expense and latency are significant concerns, as the cost scales with the number of output tokens, including those intermediate steps. In this work, we observe that most CoT tokens are unnecessary, and retaining only a small portion of them is sufficient for producing high-quality responses. Inspired by this, we propose HAWKEYE, a novel post-training and inference framework where a large model produces concise CoT instructions to guide a smaller model in response generation. HAWKEYE quantifies redundancy in CoT reasoning and distills high-density information via reinforcement learning. By leveraging these concise CoTs, HAWKEYE is able to expand responses while reducing token usage and computational cost significantly. Our evaluation shows that HAWKEYE can achieve comparable response quality using only 35% of the full CoTs, while improving clarity, coherence, and conciseness by approximately 10%. Furthermore, HAWKEYE can accelerate end-to-end reasoning by up to 3.4x on complex math tasks while reducing inference cost by up to 60%. HAWKEYE will be open-sourced and the models will be available soon.
Prior work on in-context copying has shown the existence of induction heads, which attend to and promote individual tokens during copying. In this work we introduce a new type of induction head: concept-level induction heads, which copy entire lexical units instead of individual tokens. Concept induction heads learn to attend to the ends of multi-token words throughout training, working in parallel with token-level induction heads to copy meaningful text. We show that these heads are responsible for semantic tasks like word-level translation, whereas token induction heads are vital for tasks that can only be done verbatim, like copying nonsense tokens. These two "routes" operate independently: in fact, we show that ablation of token induction heads causes models to paraphrase where they would otherwise copy verbatim. In light of these findings, we argue that although token induction heads are vital for specific tasks, concept induction heads may be more broadly relevant for in-context learning.
Accurate correspondence matching in coronary angiography images is crucial for reconstructing 3D coronary artery structures, which is essential for precise diagnosis and treatment planning of coronary artery disease (CAD). Traditional matching methods for natural images often fail to generalize to X-ray images due to inherent differences such as lack of texture, lower contrast, and overlapping structures, compounded by insufficient training data. To address these challenges, we propose a novel pipeline that generates realistic paired coronary angiography images using a diffusion model conditioned on 2D projections of 3D reconstructed meshes from Coronary Computed Tomography Angiography (CCTA), providing high-quality synthetic data for training. Additionally, we employ large-scale image foundation models to guide feature aggregation, enhancing correspondence matching accuracy by focusing on semantically relevant regions and keypoints. Our approach demonstrates superior matching performance on synthetic datasets and effectively generalizes to real-world datasets, offering a practical solution for this task. Furthermore, our work investigates the efficacy of different foundation models in correspondence matching, providing novel insights into leveraging advanced image foundation models for medical imaging applications.
Continual Learning (CL) strives to learn incrementally across tasks while mitigating catastrophic forgetting. A key challenge in CL is balancing stability (retaining prior knowledge) and plasticity (learning new tasks). While representative gradient projection methods ensure stability, they often limit plasticity. Model merging techniques offer promising solutions, but prior methods typically rely on empirical assumptions and carefully selected hyperparameters. In this paper, we explore the potential of model merging to enhance the stability-plasticity trade-off, providing theoretical insights that underscore its benefits. Specifically, we reformulate the merging mechanism using Bayesian continual learning principles and derive a closed-form solution for the optimal merging coefficient that adapts to the diverse characteristics of tasks. To validate our approach, we introduce a two-stage framework named BECAME, which synergizes the expertise of gradient projection and adaptive merging. Extensive experiments show that our approach outperforms state-of-the-art CL methods and existing merging strategies.
Hybrid neural-physics modeling frameworks through differentiable programming have emerged as powerful tools in scientific machine learning, enabling the integration of known physics with data-driven learning to improve prediction accuracy and generalizability. However, most existing hybrid frameworks rely on explicit recurrent formulations, which suffer from numerical instability and error accumulation during long-horizon forecasting. In this work, we introduce Im-PiNDiff, a novel implicit physics-integrated neural differentiable solver for stable and accurate modeling of spatiotemporal dynamics. Inspired by deep equilibrium models, Im-PiNDiff advances the state using implicit fixed-point layers, enabling robust long-term simulation while remaining fully end-to-end differentiable. To enable scalable training, we introduce a hybrid gradient propagation strategy that integrates adjoint-state methods with reverse-mode automatic differentiation. This approach eliminates the need to store intermediate solver states and decouples memory complexity from the number of solver iterations, significantly reducing training overhead. We further incorporate checkpointing techniques to manage memory in long-horizon rollouts. Numerical experiments on various spatiotemporal PDE systems, including advection-diffusion processes, Burgers' dynamics, and multi-physics chemical vapor infiltration processes, demonstrate that Im-PiNDiff achieves superior predictive performance, enhanced numerical stability, and substantial reductions in memory and runtime cost relative to explicit and naive implicit baselines. This work provides a principled, efficient, and scalable framework for hybrid neural-physics modeling.
We present the evaluation methodology, datasets and results of the BOP Challenge 2024, the 6th in a series of public competitions organized to capture the state of the art in 6D object pose estimation and related tasks. In 2024, our goal was to transition BOP from lab-like setups to real-world scenarios. First, we introduced new model-free tasks, where no 3D object models are available and methods need to onboard objects just from provided reference videos. Second, we defined a new, more practical 6D object detection task where identities of objects visible in a test image are not provided as input. Third, we introduced new BOP-H3 datasets recorded with high-resolution sensors and AR/VR headsets, closely resembling real-world scenarios. BOP-H3 include 3D models and onboarding videos to support both model-based and model-free tasks. Participants competed on seven challenge tracks. Notably, the best 2024 method for model-based 6D localization of unseen objects (FreeZeV2.1) achieves 22% higher accuracy on BOP-Classic-Core than the best 2023 method (GenFlow), and is only 4% behind the best 2023 method for seen objects (GPose2023) although being significantly slower (24.9 vs 2.7s per image). A more practical 2024 method for this task is Co-op which takes only 0.8s per image and is 13% more accurate than GenFlow. Methods have similar rankings on 6D detection as on 6D localization but higher run time. On model-based 2D detection of unseen objects, the best 2024 method (MUSE) achieves 21--29% relative improvement compared to the best 2023 method (CNOS). However, the 2D detection accuracy for unseen objects is still -35% behind the accuracy for seen objects (GDet2023), and the 2D detection stage is consequently the main bottleneck of existing pipelines for 6D localization/detection of unseen objects. The online evaluation system stays open and is available at http://bop.felk.cvut.cz/
Diffusion models (DMs) have emerged as the new state-of-the-art family of deep generative models. To gain deeper insights into the limitations of diffusion models in recommender systems, we investigate the fundamental structural disparities between images and items. Consequently, items often exhibit distinct anisotropic and directional structures that are less prevalent in images. However, the traditional forward diffusion process continuously adds isotropic Gaussian noise, causing anisotropic signals to degrade into noise, which impairs the semantically meaningful representations in recommender systems. Inspired by the advancements in hyperbolic spaces, we propose a novel \textit{\textbf{H}yperbolic} \textit{\textbf{D}iffusion} \textit{\textbf{R}ecommender} \textit{\textbf{M}odel} (named HDRM). Unlike existing directional diffusion methods based on Euclidean space, the intrinsic non-Euclidean structure of hyperbolic space makes it particularly well-adapted for handling anisotropic diffusion processes. In particular, we begin by formulating concepts to characterize latent directed diffusion processes within a geometrically grounded hyperbolic space. Subsequently, we propose a novel hyperbolic latent diffusion process specifically tailored for users and items. Drawing upon the natural geometric attributes of hyperbolic spaces, we impose structural restrictions on the space to enhance hyperbolic diffusion propagation, thereby ensuring the preservation of the intrinsic topology of user-item graphs. Extensive experiments on three benchmark datasets demonstrate the effectiveness of HDRM.
Computational methods in drug repositioning can help to conserve resources. In particular, methods based on biological networks are showing promise. Considering only the network topology and knowledge on drug target genes is not sufficient for quantitative predictions or predictions involving drug combinations. We propose an iterative procedure alternating between system identification and drug response experiments. Data from experiments are used to improve the model and drug effect knowledge, which is then used to select drugs for the next experiments. Using simulated data, we show that the procedure can identify nearly optimal drug combinations.
We propose a numerical method of searching for parameters with experimental constraints in generic flavor models by utilizing diffusion models, which are classified as a type of generative artificial intelligence (generative AI). As a specific example, we consider the $S_4^\prime$ modular flavor model and construct a neural network that reproduces quark masses, the CKM matrix, and the Jarlskog invariant by treating free parameters in the flavor model as generating targets. By generating new parameters with the trained network, we find various phenomenologically interesting parameter regions where an analytical evaluation of the $S_4^\prime$ model is challenging. Additionally, we confirm that the spontaneous CP violation occurs in the $S_4^\prime$ model. The diffusion model enables an inverse problem approach, allowing the machine to provide a series of plausible model parameters from given experimental data. Moreover, it can serve as a versatile analytical tool for extracting new physical predictions from flavor models.
3D graphics editing is crucial in applications like movie production and game design, yet it remains a time-consuming process that demands highly specialized domain expertise. Automating this process is challenging because graphical editing requires performing a variety of tasks, each requiring distinct skill sets. Recently, vision-language models (VLMs) have emerged as a powerful framework for automating the editing process, but their development and evaluation are bottlenecked by the lack of a comprehensive benchmark that requires human-level perception and presents real-world editing complexity. In this work, we present BlenderGym, the first comprehensive VLM system benchmark for 3D graphics editing. BlenderGym evaluates VLM systems through code-based 3D reconstruction tasks. We evaluate closed- and open-source VLM systems and observe that even the state-of-the-art VLM system struggles with tasks relatively easy for human Blender users. Enabled by BlenderGym, we study how inference scaling techniques impact VLM's performance on graphics editing tasks. Notably, our findings reveal that the verifier used to guide the scaling of generation can itself be improved through inference scaling, complementing recent insights on inference scaling of LLM generation in coding and math tasks. We further show that inference compute is not uniformly effective and can be optimized by strategically distributing it between generation and verification.
Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.