Pan-Tilt-Zoom (PTZ) cameras with wide-angle lenses are widely used in surveillance but often require image rectification due to their inherent nonlinear distortions. Current deep learning approaches typically struggle to maintain fine-grained geometric details, resulting in inaccurate rectification. This paper presents a Forward Distortion and Backward Warping Network (FDBW-Net), a novel framework for wide-angle image rectification. It begins by using a forward distortion model to synthesize barrel-distorted images, reducing pixel redundancy and preventing blur. The network employs a pyramid context encoder with attention mechanisms to generate backward warping flows containing geometric details. Then, a multi-scale decoder is used to restore distorted features and output rectified images. FDBW-Net's performance is validated on diverse datasets: public benchmarks, AirSim-rendered PTZ camera imagery, and real-scene PTZ camera datasets. It demonstrates that FDBW-Net achieves SOTA performance in distortion rectification, boosting the adaptability of PTZ cameras for practical visual applications.
Sparsely-activated Mixture of Experts (MoE) transformers are promising architectures for foundation models. Compared to dense transformers that require the same amount of floating point operations (FLOPs) per forward pass, MoEs benefit from improved sample efficiency at training time and achieve much stronger performance. Many closed-source and open-source frontier language models have thus adopted an MoE architecture. Naturally, practitioners will want to extend the capabilities of these models with large amounts of newly collected data without completely re-training them. Prior work has shown that a simple combination of replay and learning rate re-warming and re-decaying can enable the continual pre-training (CPT) of dense decoder-only transformers with minimal performance degradation compared to full re-training. In the case of decoder-only MoE transformers, however, it is unclear how the routing algorithm will impact continual pre-training performance: 1) do the MoE transformer's routers exacerbate forgetting relative to a dense model?; 2) do the routers maintain a balanced load on previous distributions after CPT?; 3) are the same strategies applied to dense models sufficient to continually pre-train MoE LLMs? In what follows, we conduct a large-scale (>2B parameter switch and DeepSeek MoE LLMs trained for 600B tokens) empirical study across four MoE transformers to answer these questions. Our results establish a surprising robustness to distribution shifts for both Sinkhorn-Balanced and Z-and-Aux-loss-balanced routing algorithms, even in MoEs continually pre-trained without replay. Moreover, we show that MoE LLMs maintain their sample efficiency (relative to a FLOP-matched dense model) during CPT and that they can match the performance of a fully re-trained MoE at a fraction of the cost.
We propose LauraTSE, an Auto-Regressive Decoder-Only Language Model for Target Speaker Extraction (TSE) based on the LauraGPT backbone. It employs a small-scale auto-regressive decoder-only language model which takes the continuous representations for both the mixture and the reference speeches and produces the first few layers of the target speech's discrete codec representations. In addition, a one-step encoder-only language model reconstructs the sum of the predicted codec embeddings using both the mixture and the reference information. Our approach achieves superior or comparable performance to existing generative and discriminative TSE models. To the best of our knowledge, LauraTSE is the first single-task TSE model to leverage an auto-regressive decoder-only language model as the backbone.
Large Language Models (LLMs) can encode complex relationships in their latent spaces, yet harnessing them for optimization under uncertainty remains challenging. We address this gap with a novel architecture that reframes LLM finetuning as Gaussian process (GP) marginal likelihood optimization via deep kernel methods. We introduce LLM-based deep kernels, jointly optimized with GPs to preserve the benefits of both - LLMs to provide a rich and flexible input space for Bayesian optimization and - GPs to model this space with predictive uncertainty for more efficient sampling. Applied to Buchwald-Hartwig reaction optimization, our method nearly doubles the discovery rate of high-performing reactions compared to static LLM embeddings (from 24% to 43% coverage of the top 5% reactions in just 50 optimization iterations). We also observe a 14% improvement over domain-specific representations without requiring specialized features. Extensive empirical evaluation across 19 benchmarks - ranging from general chemistry to reaction and molecular property optimization - demonstrates our method's robustness, generality, and consistent improvements across: (1) tasks, (2) LLM architectures (encoder, decoder, encoder-decoder), (3) pretraining domains (chemistry-related or general-purpose) and (4) hyperparameter settings (tuned once on a single dataset). Finally, we explain these improvements: joint LLM-GP optimization through marginal likelihood implicitly performs contrastive learning, aligning representations to produce (1) better-structured embedding spaces, (2) improved uncertainty calibration, and (3) more efficient sampling - without requiring any external loss. This work provides both practical advances in sample-efficient optimization and insights into what makes effective Bayesian optimization.
Accurate surgical instrument segmentation is essential in cataract surgery for tasks such as skill assessment and workflow optimization. However, limited annotated data makes it difficult to develop fully automatic models. Prompt-based methods like SAM2 offer flexibility yet remain highly sensitive to the point prompt placement, often leading to inconsistent segmentations. We address this issue by introducing RP-SAM2, which incorporates a novel shift block and a compound loss function to stabilize point prompts. Our approach reduces annotator reliance on precise point positioning while maintaining robust segmentation capabilities. Experiments on the Cataract1k dataset demonstrate that RP-SAM2 improves segmentation accuracy, with a 2% mDSC gain, a 21.36% reduction in mHD95, and decreased variance across random single-point prompt results compared to SAM2. Additionally, on the CaDIS dataset, pseudo masks generated by RP-SAM2 for fine-tuning SAM2's mask decoder outperformed those generated by SAM2. These results highlight RP-SAM2 as a practical, stable and reliable solution for semi-automatic instrument segmentation in data-constrained medical settings. The code is available at https://github.com/BioMedIA-MBZUAI/RP-SAM2.
Continual learning in large language models (LLMs) is prone to catastrophic forgetting, where adapting to new tasks significantly degrades performance on previously learned ones. Existing methods typically rely on low-rank, parameter-efficient updates that limit the model's expressivity and introduce additional parameters per task, leading to scalability issues. To address these limitations, we propose a novel continual full fine-tuning approach leveraging adaptive singular value decomposition (SVD). Our method dynamically identifies task-specific low-rank parameter subspaces and constrains updates to be orthogonal to critical directions associated with prior tasks, thus effectively minimizing interference without additional parameter overhead or storing previous task gradients. We evaluate our approach extensively on standard continual learning benchmarks using both encoder-decoder (T5-Large) and decoder-only (LLaMA-2 7B) models, spanning diverse tasks including classification, generation, and reasoning. Empirically, our method achieves state-of-the-art results, up to 7% higher average accuracy than recent baselines like O-LoRA, and notably maintains the model's general linguistic capabilities, instruction-following accuracy, and safety throughout the continual learning process by reducing forgetting to near-negligible levels. Our adaptive SVD framework effectively balances model plasticity and knowledge retention, providing a practical, theoretically grounded, and computationally scalable solution for continual learning scenarios in large language models.
Unlocking the next generation of biotechnology and therapeutic innovation demands overcoming the inherent complexity and resource-intensity of conventional protein engineering methods. Recent GenAI-powered computational techniques often rely on the availability of the target protein's 3D structures and specific binding sites to generate high-affinity binders, constraints exhibited by models such as AlphaProteo and RFdiffusion. In this work, we explore the use of Protein Language Models (pLMs) for high-affinity binder generation. We introduce Prot42, a novel family of Protein Language Models (pLMs) pretrained on vast amounts of unlabeled protein sequences. By capturing deep evolutionary, structural, and functional insights through an advanced auto-regressive, decoder-only architecture inspired by breakthroughs in natural language processing, Prot42 dramatically expands the capabilities of computational protein design based on language only. Remarkably, our models handle sequences up to 8,192 amino acids, significantly surpassing standard limitations and enabling precise modeling of large proteins and complex multi-domain sequences. Demonstrating powerful practical applications, Prot42 excels in generating high-affinity protein binders and sequence-specific DNA-binding proteins. Our innovative models are publicly available, offering the scientific community an efficient and precise computational toolkit for rapid protein engineering.
Supervised machine learning (ML) methods are emerging as valid alternatives to standard mathematical methods for identifying knots in long, collapsed polymers. Here, we introduce a hybrid supervised/unsupervised ML approach for knot classification based on a variational autoencoder enhanced with a knot type classifier (VAEC). The neat organization of knots in its latent representation suggests that the VAEC, only based on an arbitrary labeling of three-dimensional configurations, has grasped complex topological concepts such as chirality, unknotting number, braid index, and the grouping in families such as achiral, torus, and twist knots. The understanding of topological concepts is confirmed by the ability of the VAEC to distinguish the chirality of knots $9_{42}$ and $10_{71}$ not used for its training and with a notoriously undetected chirality to standard tools. The well-organized latent space is also key for generating configurations with the decoder that reliably preserves the topology of the input ones. Our findings demonstrate the ability of a hybrid supervised-generative ML algorithm to capture different topological features of entangled filaments and to exploit this knowledge to faithfully reconstruct or produce new knotted configurations without simulations.
Unmanned aerial vehicles (UAVs) have been widely applied in military and civilian fields, but they also pose great threats to restricted areas, such as densely populated areas and airports. Thermal infrared (IR) imaging technology is capable of monitoring UAVs at a long range in both day and night conditions. Therefore, the anti-UAV technology based on thermal IR imaging has attracted growing attention. However, the images acquired by IR sensors often suffer from small and dim targets, as well as heavy background clutter and noise. Conventional detection methods usually have a high false alarm rate and low detection accuracy. This article proposes a detection method that formulates the UAV detection as predicting the residual image (i.e., background, clutter, and noise) by learning the nonlinear mapping from the input image to the residual image. The UAV target image is obtained by subtracting the residual image from the input IR image. The constructed end-to-end U-shaped network exploits the depthwise separable residual dense blocks in the encoder stage to extract the abundant hierarchical features. Besides, the multiscale feature fusion and representation block is introduced to fully aggregate multiscale features from the encoder layers and intermediate connection layers at the same scale, as well as the decoder layers at different scales, to better reconstruct the residual image in the decoder stage. In addition, the global residual connection is adopted in the proposed network to provide long-distance information compensation and promote gradient backpropagation, which further improves the performance in reconstructing the image. The experimental results show that the proposed method achieves favorable detection performance in real-world IR images and outperforms other state-of-the-art methods in terms of quantitative and qualitative evaluation metrics.
Machine learning-based embedded systems for safety-critical applications, such as aerospace and autonomous driving, must be robust to perturbations caused by soft errors. As transistor geometries shrink and voltages decrease, modern electronic devices become more susceptible to background radiation, increasing the concern about failures produced by soft errors. The resilience of deep neural networks (DNNs) to these errors depends not only on target device technology but also on model structure and the numerical representation and arithmetic precision of their parameters. Compression techniques like pruning and quantization, used to reduce memory footprint and computational complexity, alter both model structure and representation, affecting soft error robustness. In this regard, although often overlooked, the choice of activation functions (AFs) impacts not only accuracy and trainability but also compressibility and error resilience. This paper explores the use of bounded AFs to enhance robustness against parameter perturbations, while evaluating their effects on model accuracy, compressibility, and computational load with a technology-agnostic approach. We focus on encoder-decoder convolutional models developed for semantic segmentation of hyperspectral images with application to autonomous driving systems. Experiments are conducted on an AMD-Xilinx's KV260 SoM.
High-resolution segmentation is critical for precise disease diagnosis by extracting micro-imaging information from medical images. Existing transformer-based encoder-decoder frameworks have demonstrated remarkable versatility and zero-shot performance in medical segmentation. While beneficial, they usually require huge memory costs when handling large-size segmentation mask predictions, which are expensive to apply to real-world scenarios. To address this limitation, we propose a memory-efficient framework for high-resolution medical image segmentation, called HRMedSeg. Specifically, we first devise a lightweight gated vision transformer (LGViT) as our image encoder to model long-range dependencies with linear complexity. Then, we design an efficient cross-multiscale decoder (ECM-Decoder) to generate high-resolution segmentation masks. Moreover, we utilize feature distillation during pretraining to unleash the potential of our proposed model. Extensive experiments reveal that HRMedSeg outperforms state-of-the-arts in diverse high-resolution medical image segmentation tasks. In particular, HRMedSeg uses only 0.59GB GPU memory per batch during fine-tuning, demonstrating low training costs. Besides, when HRMedSeg meets the Segment Anything Model (SAM), our HRMedSegSAM takes 0.61% parameters of SAM-H. The code is available at https://github.com/xq141839/HRMedSeg.
We introduce \textbf{VISTA-OCR} (Vision and Spatially-aware Text Analysis OCR), a lightweight architecture that unifies text detection and recognition within a single generative model. Unlike conventional methods that require separate branches with dedicated parameters for text recognition and detection, our approach leverages a Transformer decoder to sequentially generate text transcriptions and their spatial coordinates in a unified branch. Built on an encoder-decoder architecture, VISTA-OCR is progressively trained, starting with the visual feature extraction phase, followed by multitask learning with multimodal token generation. To address the increasing demand for versatile OCR systems capable of advanced tasks, such as content-based text localization \ref{content_based_localization}, we introduce new prompt-controllable OCR tasks during pre-training.To enhance the model's capabilities, we built a new dataset composed of real-world examples enriched with bounding box annotations and synthetic samples. Although recent Vision Large Language Models (VLLMs) can efficiently perform these tasks, their high computational cost remains a barrier for practical deployment. In contrast, our VISTA$_{\text{omni}}$ variant processes both handwritten and printed documents with only 150M parameters, interactively, by prompting. Extensive experiments on multiple datasets demonstrate that VISTA-OCR achieves better performance compared to state-of-the-art specialized models on standard OCR tasks while showing strong potential for more sophisticated OCR applications, addressing the growing need for interactive OCR systems. All code and annotations for VISTA-OCR will be made publicly available upon acceptance.
Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing the performance of large language models (LLMs) by integrating external knowledge into the generation process. A key component of RAG pipelines is the reranker, which selects the most relevant documents from a pool of retrieved candidates and significantly improves the quality of the generated responses. While rerankers refine the selection of retrieved documents in RAG pipelines, they introduce computational challenges that hinder high throughput and low latency. To address this problem, we propose HyperRAG, a system that optimizes the trade-off between quality and efficiency in RAG pipelines by leveraging KV-cache reuse for efficient reranker inference. By reusing document-side KV-cache, HyperRAG achieves both high-quality generation and system-level efficiency. To fully realize the benefits of KV-cache reuse, HyperRAG incorporates a range of system-level optimizations designed to enhance efficiency and scalability. Experiments show that HyperRAG achieves a 2 - 3 throughput improvement with decoder-only rerankers while also delivering higher downstream performance compared with traditional RAG service.
In this work, we introduce implicit Finite Operator Learning (iFOL) for the continuous and parametric solution of partial differential equations (PDEs) on arbitrary geometries. We propose a physics-informed encoder-decoder network to establish the mapping between continuous parameter and solution spaces. The decoder constructs the parametric solution field by leveraging an implicit neural field network conditioned on a latent or feature code. Instance-specific codes are derived through a PDE encoding process based on the second-order meta-learning technique. In training and inference, a physics-informed loss function is minimized during the PDE encoding and decoding. iFOL expresses the loss function in an energy or weighted residual form and evaluates it using discrete residuals derived from standard numerical PDE methods. This approach results in the backpropagation of discrete residuals during both training and inference. iFOL features several key properties: (1) its unique loss formulation eliminates the need for the conventional encode-process-decode pipeline previously used in operator learning with conditional neural fields for PDEs; (2) it not only provides accurate parametric and continuous fields but also delivers solution-to-parameter gradients without requiring additional loss terms or sensitivity analysis; (3) it can effectively capture sharp discontinuities in the solution; and (4) it removes constraints on the geometry and mesh, making it applicable to arbitrary geometries and spatial sampling (zero-shot super-resolution capability). We critically assess these features and analyze the network's ability to generalize to unseen samples across both stationary and transient PDEs. The overall performance of the proposed method is promising, demonstrating its applicability to a range of challenging problems in computational mechanics.
Existing makeup techniques often require designing multiple models to handle different inputs and align features across domains for different makeup tasks, e.g., beauty filter, makeup transfer, and makeup removal, leading to increased complexity. Another limitation is the absence of text-guided makeup try-on, which is more user-friendly without needing reference images. In this study, we make the first attempt to use a single model for various makeup tasks. Specifically, we formulate different makeup tasks as cross-domain translations and leverage a cross-domain diffusion model to accomplish all tasks. Unlike existing methods that rely on separate encoder-decoder configurations or cycle-based mechanisms, we propose using different domain embeddings to facilitate domain control. This allows for seamless domain switching by merely changing embeddings with a single model, thereby reducing the reliance on additional modules for different tasks. Moreover, to support precise text-to-makeup applications, we introduce the MT-Text dataset by extending the MT dataset with textual annotations, advancing the practicality of makeup technologies.
Speech separation (SS) seeks to disentangle a multi-talker speech mixture into single-talker speech streams. Although SS can be generally achieved using offline methods, such a processing paradigm is not suitable for real-time streaming applications. Causal separation models, which rely only on past and present information, offer a promising solution for real-time streaming. However, these models typically suffer from notable performance degradation due to the absence of future context. In this paper, we introduce a novel frontend that is designed to mitigate the mismatch between training and run-time inference by implicitly incorporating future information into causal models through predictive patterns. The pretrained frontend employs a transformer decoder network with a causal convolutional encoder as the backbone and is pretrained in a self-supervised manner with two innovative pretext tasks: autoregressive hybrid prediction and contextual knowledge distillation. These tasks enable the model to capture predictive patterns directly from mixtures in a self-supervised manner. The pretrained frontend subsequently serves as a feature extractor to generate high-quality predictive patterns. Comprehensive evaluations on synthetic and real-world datasets validated the effectiveness of the proposed pretrained frontend.
The advent of Large Language Models (LLMs) has significantly reshaped the landscape of machine translation (MT), particularly for low-resource languages and domains that lack sufficient parallel corpora, linguistic tools, and computational infrastructure. This survey presents a comprehensive overview of recent progress in leveraging LLMs for MT. We analyze techniques such as few-shot prompting, cross-lingual transfer, and parameter-efficient fine-tuning that enable effective adaptation to under-resourced settings. The paper also explores synthetic data generation strategies using LLMs, including back-translation and lexical augmentation. Additionally, we compare LLM-based translation with traditional encoder-decoder models across diverse language pairs, highlighting the strengths and limitations of each. We discuss persistent challenges such as hallucinations, evaluation inconsistencies, and inherited biases while also evaluating emerging LLM-driven metrics for translation quality. This survey offers practical insights and outlines future directions for building robust, inclusive, and scalable MT systems in the era of large-scale generative models.
We present ILLUME+ that leverages dual visual tokenization and a diffusion decoder to improve both deep semantic understanding and high-fidelity image generation. Existing unified models have struggled to simultaneously handle the three fundamental capabilities in a unified model: understanding, generation, and editing. Models like Chameleon and EMU3 utilize VQGAN for image discretization, due to the lack of deep semantic interaction, they lag behind specialist models like LLaVA in visual understanding tasks. To mitigate this, LaViT and ILLUME employ semantic encoders for tokenization, but they struggle with image editing due to poor texture preservation. Meanwhile, Janus series decouples the input and output image representation, limiting their abilities to seamlessly handle interleaved image-text understanding and generation. In contrast, ILLUME+ introduces a unified dual visual tokenizer, DualViTok, which preserves both fine-grained textures and text-aligned semantics while enabling a coarse-to-fine image representation strategy for multimodal understanding and generation. Additionally, we employ a diffusion model as the image detokenizer for enhanced generation quality and efficient super-resolution. ILLUME+ follows a continuous-input, discrete-output scheme within the unified MLLM and adopts a progressive training procedure that supports dynamic resolution across the vision tokenizer, MLLM, and diffusion decoder. This design allows for flexible and efficient context-aware image editing and generation across diverse tasks. ILLUME+ (3B) exhibits competitive performance against existing unified MLLMs and specialized models across multimodal understanding, generation, and editing benchmarks. With its strong performance, ILLUME+ provides a scalable and versatile foundation for future multimodal applications. Project Page: https://illume-unified-mllm.github.io/.
Unsupervised domain adaptation (UDA) frameworks have shown good generalization capabilities for 3D point cloud semantic segmentation models on clean data. However, existing works overlook adversarial robustness when the source domain itself is compromised. To comprehensively explore the robustness of the UDA frameworks, we first design a stealthy adversarial point cloud generation attack that can significantly contaminate datasets with only minor perturbations to the point cloud surface. Based on that, we propose a novel dataset, AdvSynLiDAR, comprising synthesized contaminated LiDAR point clouds. With the generated corrupted data, we further develop the Adversarial Adaptation Framework (AAF) as the countermeasure. Specifically, by extending the key point sensitive (KPS) loss towards the Robust Long-Tail loss (RLT loss) and utilizing a decoder branch, our approach enables the model to focus on long-tail classes during the pre-training phase and leverages high-confidence decoded point cloud information to restore point cloud structures during the adaptation phase. We evaluated our AAF method on the AdvSynLiDAR dataset, where the results demonstrate that our AAF method can mitigate performance degradation under source adversarial perturbations for UDA in the 3D point cloud segmentation application.
Epilepsy is a common neurological disorder that affects around 65 million people worldwide. Detecting seizures quickly and accurately is vital, given the prevalence and severity of the associated complications. Recently, deep learning-based automated seizure detection methods have emerged as solutions; however, most existing methods require extensive post-processing and do not effectively handle the crucial long-range patterns in EEG data. In this work, we propose SeizureTransformer, a simple model comprised of (i) a deep encoder comprising 1D convolutions (ii) a residual CNN stack and a transformer encoder to embed previous output into high-level representation with contextual information, and (iii) streamlined decoder which converts these features into a sequence of probabilities, directly indicating the presence or absence of seizures at every time step. Extensive experiments on public and private EEG seizure detection datasets demonstrate that our model significantly outperforms existing approaches (ranked in the first place in the 2025 "seizure detection challenge" organized in the International Conference on Artificial Intelligence in Epilepsy and Other Neurological Disorders), underscoring its potential for real-time, precise seizure detection.