Variational Autoencoders (VAEs) typically rely on a probabilistic decoder with a predefined likelihood, most commonly an isotropic Gaussian, to model the data conditional on latent variables. While convenient for optimization, this choice often leads to likelihood misspecification, resulting in blurry reconstructions and poor data fidelity, especially for high-dimensional data such as images. In this work, we propose EnVAE, a novel likelihood-free generative framework that has a deterministic decoder and employs the energy score--a proper scoring rule--to build the reconstruction loss. This enables likelihood-free inference without requiring explicit parametric density functions. To address the computational inefficiency of the energy score, we introduce a fast variant, FEnVAE, based on the local smoothness of the decoder and the sharpness of the posterior distribution of latent variables. This yields an efficient single-sample training objective that integrates seamlessly into existing VAE pipelines with minimal overhead. Empirical results on standard benchmarks demonstrate that EnVAE achieves superior reconstruction and generation quality compared to likelihood-based baselines. Our framework offers a general, scalable, and statistically principled alternative for flexible and nonparametric distribution learning in generative modeling.
We introduce a novel generative framework for functions by integrating Implicit Neural Representations (INRs) and Transformer-based hypernetworks into latent variable models. Unlike prior approaches that rely on MLP-based hypernetworks with scalability limitations, our method employs a Transformer-based decoder to generate INR parameters from latent variables, addressing both representation capacity and computational efficiency. Our framework extends latent diffusion models (LDMs) to INR generation by replacing standard decoders with a Transformer-based hypernetwork, which can be trained either from scratch or via hyper-transforming-a strategy that fine-tunes only the decoder while freezing the pre-trained latent space. This enables efficient adaptation of existing generative models to INR-based representations without requiring full retraining.
Current Audio-Visual Source Separation methods primarily adopt two design strategies. The first strategy involves fusing audio and visual features at the bottleneck layer of the encoder, followed by processing the fused features through the decoder. However, when there is a significant disparity between the two modalities, this approach may lead to the loss of critical information. The second strategy avoids direct fusion and instead relies on the decoder to handle the interaction between audio and visual features. Nonetheless, if the encoder fails to integrate information across modalities adequately, the decoder may be unable to effectively capture the complex relationships between them. To address these issues, this paper proposes a dynamic fusion method based on a gating mechanism that dynamically adjusts the modality fusion degree. This approach mitigates the limitations of solely relying on the decoder and facilitates efficient collaboration between audio and visual features. Additionally, an audio attention module is introduced to enhance the expressive capacity of audio features, thereby further improving model performance. Experimental results demonstrate that our method achieves significant performance improvements on two benchmark datasets, validating its effectiveness and advantages in Audio-Visual Source Separation tasks.
Arabic text diacritization remains a persistent challenge in natural language processing due to the language's morphological richness. In this paper, we introduce Sadeed, a novel approach based on a fine-tuned decoder-only language model adapted from Kuwain 1.5B Hennara et al. [2025], a compact model originally trained on diverse Arabic corpora. Sadeed is fine-tuned on carefully curated, high-quality diacritized datasets, constructed through a rigorous data-cleaning and normalization pipeline. Despite utilizing modest computational resources, Sadeed achieves competitive results compared to proprietary large language models and outperforms traditional models trained on similar domains. Additionally, we highlight key limitations in current benchmarking practices for Arabic diacritization. To address these issues, we introduce SadeedDiac-25, a new benchmark designed to enable fairer and more comprehensive evaluation across diverse text genres and complexity levels. Together, Sadeed and SadeedDiac-25 provide a robust foundation for advancing Arabic NLP applications, including machine translation, text-to-speech, and language learning tools.
This paper presents a neural architecture search method based on Transformer architecture, searching cross multihead attention computation ways for different number of encoder and decoder combinations. In order to search for neural network structures with better translation results, we considered perplexity as an auxiliary evaluation metric for the algorithm in addition to BLEU scores and iteratively improved each individual neural network within the population by a multi-objective genetic algorithm. Experimental results show that the neural network structures searched by the algorithm outperform all the baseline models, and that the introduction of the auxiliary evaluation metric can find better models than considering only the BLEU score as an evaluation metric.
Frequency-domain compression has proven effective in reducing redundancies for spatial signals. In this work, we propose FreqKV, a novel frequency domain key-value (KV) compression technique that enables efficient context window extension for decoder-only large language models (LLMs). Our approach is motivated by a key observation that, in the frequency domain, the energy distribution of the KV cache is predominantly concentrated in low-frequency components. By discarding high-frequency components, we achieve efficient compression of the KV cache with minimal information loss. FreqKV iteratively compresses the increasing KV cache to a fixed size in the frequency domain, allowing models to process lengthy contexts efficiently. Introducing no additional parameters or architectural modifications, FreqKV is applicable to both fine-tuning and inference. With minimal fine-tuning, LLMs can learn to leverage the limited cache that is compressed in the frequency domain and extend the context window. Experiments on a range of long context language modeling and understanding tasks demonstrate the efficiency and effectiveness of the proposed method.
Recent advances in digital watermarking make use of deep neural networks for message embedding and extraction. They typically follow the ``encoder-noise layer-decoder''-based architecture. By deliberately establishing a differentiable noise layer to simulate the distortion of the watermarked signal, they jointly train the deep encoder and decoder to fit the noise layer to guarantee robustness. As a result, they are usually weak against unknown distortions that are not used in their training pipeline. In this paper, we propose a novel watermarking framework to resist unknown distortions, namely Adversarial Shallow Watermarking (ASW). ASW utilizes only a shallow decoder that is randomly parameterized and designed to be insensitive to distortions for watermarking extraction. During the watermark embedding, ASW freezes the shallow decoder and adversarially optimizes a host image until its updated version (i.e., the watermarked image) stably triggers the shallow decoder to output the watermark message. During the watermark extraction, it accurately recovers the message from the watermarked image by leveraging the insensitive nature of the shallow decoder against arbitrary distortions. Our ASW is training-free, encoder-free, and noise layer-free. Experiments indicate that the watermarked images created by ASW have strong robustness against various unknown distortions. Compared to the existing ``encoder-noise layer-decoder'' approaches, ASW achieves comparable results on known distortions and better robustness on unknown distortions.
Magnetic resonance spectroscopy (MRS) is a non-invasive technique to measure the metabolic composition of tissues, offering valuable insights into neurological disorders, tumor detection, and other metabolic dysfunctions. However, accurate metabolite quantification is hindered by challenges such as spectral overlap, low signal-to-noise ratio, and various artifacts. Traditional methods like linear-combination modeling are susceptible to ambiguities and commonly only provide a theoretical lower bound on estimation accuracy in the form of the Cram\'er-Rao bound. This work introduces a Bayesian inference framework using Sylvester normalizing flows (SNFs) to approximate posterior distributions over metabolite concentrations, enhancing quantification reliability. A physics-based decoder incorporates prior knowledge of MRS signal formation, ensuring realistic distribution representations. We validate the method on simulated 7T proton MRS data, demonstrating accurate metabolite quantification, well-calibrated uncertainties, and insights into parameter correlations and multi-modal distributions.
We study the problem of distributed multi-view representation learning. In this problem, $K$ agents observe each one distinct, possibly statistically correlated, view and independently extracts from it a suitable representation in a manner that a decoder that gets all $K$ representations estimates correctly the hidden label. In the absence of any explicit coordination between the agents, a central question is: what should each agent extract from its view that is necessary and sufficient for a correct estimation at the decoder? In this paper, we investigate this question from a generalization error perspective. First, we establish several generalization bounds in terms of the relative entropy between the distribution of the representations extracted from training and "test" datasets and a data-dependent symmetric prior, i.e., the Minimum Description Length (MDL) of the latent variables for all views and training and test datasets. Then, we use the obtained bounds to devise a regularizer; and investigate in depth the question of the selection of a suitable prior. In particular, we show and conduct experiments that illustrate that our data-dependent Gaussian mixture priors with judiciously chosen weights lead to good performance. For single-view settings (i.e., $K=1$), our experimental results are shown to outperform existing prior art Variational Information Bottleneck (VIB) and Category-Dependent VIB (CDVIB) approaches. Interestingly, we show that a weighted attention mechanism emerges naturally in this setting. Finally, for the multi-view setting, we show that the selection of the joint prior as a Gaussians product mixture induces a Gaussian mixture marginal prior for each marginal view and implicitly encourages the agents to extract and output redundant features, a finding which is somewhat counter-intuitive.
We present Variational Self-Supervised Learning (VSSL), a novel framework that combines variational inference with self-supervised learning to enable efficient, decoder-free representation learning. Unlike traditional VAEs that rely on input reconstruction via a decoder, VSSL symmetrically couples two encoders with Gaussian outputs. A momentum-updated teacher network defines a dynamic, data-dependent prior, while the student encoder produces an approximate posterior from augmented views. The reconstruction term in the ELBO is replaced with a cross-view denoising objective, preserving the analytical tractability of Gaussian KL divergence. We further introduce cosine-based formulations of KL and log-likelihood terms to enhance semantic alignment in high-dimensional latent spaces. Experiments on CIFAR-10, CIFAR-100, and ImageNet-100 show that VSSL achieves competitive or superior performance to leading self-supervised methods, including BYOL and MoCo V3. VSSL offers a scalable, probabilistically grounded approach to learning transferable representations without generative reconstruction, bridging the gap between variational modeling and modern self-supervised techniques.
The growing applications of AR/VR increase the demand for real-time full-body pose estimation from Head-Mounted Displays (HMDs). Although HMDs provide joint signals from the head and hands, reconstructing a full-body pose remains challenging due to the unconstrained lower body. Recent advancements often rely on conventional neural networks and generative models to improve performance in this task, such as Transformers and diffusion models. However, these approaches struggle to strike a balance between achieving precise pose reconstruction and maintaining fast inference speed. To overcome these challenges, a lightweight and efficient model, SSD-Poser, is designed for robust full-body motion estimation from sparse observations. SSD-Poser incorporates a well-designed hybrid encoder, State Space Attention Encoders, to adapt the state space duality to complex motion poses and enable real-time realistic pose reconstruction. Moreover, a Frequency-Aware Decoder is introduced to mitigate jitter caused by variable-frequency motion signals, remarkably enhancing the motion smoothness. Comprehensive experiments on the AMASS dataset demonstrate that SSD-Poser achieves exceptional accuracy and computational efficiency, showing outstanding inference efficiency compared to state-of-the-art methods.
We propose the Compact Clustering Attention (COCA) layer, an effective building block that introduces a hierarchical strategy for object-centric representation learning, while solving the unsupervised object discovery task on single images. COCA is an attention-based clustering module capable of extracting object-centric representations from multi-object scenes, when cascaded into a bottom-up hierarchical network architecture, referred to as COCA-Net. At its core, COCA utilizes a novel clustering algorithm that leverages the physical concept of compactness, to highlight distinct object centroids in a scene, providing a spatial inductive bias. Thanks to this strategy, COCA-Net generates high-quality segmentation masks on both the decoder side and, notably, the encoder side of its pipeline. Additionally, COCA-Net is not bound by a predetermined number of object masks that it generates and handles the segmentation of background elements better than its competitors. We demonstrate COCA-Net's segmentation performance on six widely adopted datasets, achieving superior or competitive results against the state-of-the-art models across nine different evaluation metrics.
Pipeline Parallelism (PP) enables large neural network training on small, interconnected devices by splitting the model into multiple stages. To maximize pipeline utilization, asynchronous optimization is appealing as it offers 100% pipeline utilization by construction. However, it is inherently challenging as the weights and gradients are no longer synchronized, leading to stale (or delayed) gradients. To alleviate this, we introduce a variant of Nesterov Accelerated Gradient (NAG) for asynchronous optimization in PP. Specifically, we modify the look-ahead step in NAG to effectively address the staleness in gradients. We theoretically prove that our approach converges at a sublinear rate in the presence of fixed delay in gradients. Our experiments on large-scale language modelling tasks using decoder-only architectures with up to 1B parameters, demonstrate that our approach significantly outperforms existing asynchronous methods, even surpassing the synchronous baseline.
Real-time, intelligent, and natural speech interaction is an essential part of the next-generation human-computer interaction. Recent advancements have showcased the potential of building intelligent spoken chatbots based on large language models (LLMs). In this paper, we introduce LLaMA-Omni 2, a series of speech language models (SpeechLMs) ranging from 0.5B to 14B parameters, capable of achieving high-quality real-time speech interaction. LLaMA-Omni 2 is built upon the Qwen2.5 series models, integrating a speech encoder and an autoregressive streaming speech decoder. Despite being trained on only 200K multi-turn speech dialogue samples, LLaMA-Omni 2 demonstrates strong performance on several spoken question answering and speech instruction following benchmarks, surpassing previous state-of-the-art SpeechLMs like GLM-4-Voice, which was trained on millions of hours of speech data.
Large Language Models (LLMs) based on autoregressive, decoder-only Transformers generate text one token at a time, where a token represents a discrete unit of text. As each newly produced token is appended to the partial output sequence, the length grows and so does the memory and compute load, due to the expanding key-value caches, which store intermediate representations of all previously generated tokens in the multi-head attention (MHA) layer. As this iterative process steadily increases memory and compute demands, layer-based partitioning in resource-constrained edge environments often results in memory overload or high inference latency. To address this and reduce inference latency, we propose a resource-aware Transformer architecture partitioning algorithm, where the partitioning decision is updated at regular intervals during token generation. The approach is myopic in that it is based on instantaneous information about device resource availability and network link bandwidths. When first executed, the algorithm places blocks on devices, and in later executions, it migrates these blocks among devices so that the sum of migration delay and inference delay remains low. Our approach partitions the decoder at the attention head level, co-locating each attention head with its key-value cache and allowing dynamic migrations whenever resources become tight. By allocating different attention heads to different devices, we exploit parallel execution of attention heads and thus achieve substantial reductions in inference delays. Our experiments show that in small-scale settings (3-5 devices), the proposed method achieves within 15 to 20 percent of an exact optimal solver's latency, while in larger-scale tests it achieves notable improvements in inference speed and memory usage compared to state-of-the-art layer-based partitioning approaches.
We present Rapid Attention Distillation to Linear Attention Decoders at Scale (RADLADS), a protocol for rapidly converting softmax attention transformers into linear attention decoder models, along with two new RWKV-variant architectures, and models converted from popular Qwen2.5 open source models in 7B, 32B, and 72B sizes. Our conversion process requires only 350-700M tokens, less than 0.005% of the token count used to train the original teacher models. Converting to our 72B linear attention model costs less than \$2,000 USD at today's prices, yet quality at inference remains close to the original transformer. These models achieve state-of-the-art downstream performance across a set of standard benchmarks for linear attention models of their size. We release all our models on HuggingFace under the Apache 2.0 license, with the exception of our 72B models which are also governed by the Qwen License Agreement. Models at https://huggingface.co/collections/recursal/radlads-6818ee69e99e729ba8a87102 Training Code at https://github.com/recursal/RADLADS-paper
Modern Search Engine Results Pages (SERPs) present complex layouts where multiple elements compete for visibility. Attention modelling is crucial for optimising web design and computational advertising, whereas attention metrics can inform ad placement and revenue strategies. We introduce AdSight, a method leveraging mouse cursor trajectories to quantify in a scalable and accurate manner user attention in multi-slot environments like SERPs. AdSight uses a novel Transformer-based sequence-to-sequence architecture where the encoder processes cursor trajectory embeddings, and the decoder incorporates slot-specific features, enabling robust attention prediction across various SERP layouts. We evaluate our approach on two Machine Learning tasks: (1) regression, to predict fixation times and counts; and (2) classification, to determine some slot types were noticed. Our findings demonstrate the model's ability to predict attention with unprecedented precision, offering actionable insights for researchers and practitioners.
Large Language Models (LLMs) pruning seeks to remove unimportant weights for inference speedup with minimal performance impact. However, existing methods often suffer from performance loss without full-model sparsity-aware fine-tuning. This paper presents Wanda++, a novel pruning framework that outperforms the state-of-the-art methods by utilizing decoder-block-level \textbf{regional} gradients. Specifically, Wanda++ improves the pruning score with regional gradients for the first time and proposes an efficient regional optimization method to minimize pruning-induced output discrepancies between the dense and sparse decoder output. Notably, Wanda++ improves perplexity by up to 32\% over Wanda in the language modeling task and generalizes effectively to downstream tasks. Further experiments indicate our proposed method is orthogonal to sparsity-aware fine-tuning, where Wanda++ can be combined with LoRA fine-tuning to achieve a similar perplexity improvement as the Wanda method. The proposed method is lightweight, pruning a 7B LLaMA model in under 10 minutes on a single NVIDIA H100 GPU.
Encoder-decoder Large Language Models (LLMs), such as BERT and RoBERTa, require that all categories in an annotation task be sufficiently represented in the training data for optimal performance. However, it is often difficult to find sufficient examples for all categories in a task when building a high-quality training set. In this article, I describe this problem and propose a solution, the synthetic imputation approach. Leveraging a generative LLM (GPT-4o), this approach generates synthetic texts based on careful prompting and five original examples drawn randomly with replacement from the sample. This approach ensures that new synthetic texts are sufficiently different from the original texts to reduce overfitting, but retain the underlying substantive meaning of the examples to maximize out-of-sample performance. With 75 original examples or more, synthetic imputation's performance is on par with a full sample of original texts, and overfitting remains low, predictable and correctable with 50 original samples. The synthetic imputation approach provides a novel role for generative LLMs in research and allows applied researchers to balance their datasets for best performance.
Temporal Action Detection and Moment Retrieval constitute two pivotal tasks in video understanding, focusing on precisely localizing temporal segments corresponding to specific actions or events. Recent advancements introduced Moment Detection to unify these two tasks, yet existing approaches remain confined to closed-set scenarios, limiting their applicability in open-world contexts. To bridge this gap, we present Grounding-MD, an innovative, grounded video-language pre-training framework tailored for open-world moment detection. Our framework incorporates an arbitrary number of open-ended natural language queries through a structured prompt mechanism, enabling flexible and scalable moment detection. Grounding-MD leverages a Cross-Modality Fusion Encoder and a Text-Guided Fusion Decoder to facilitate comprehensive video-text alignment and enable effective cross-task collaboration. Through large-scale pre-training on temporal action detection and moment retrieval datasets, Grounding-MD demonstrates exceptional semantic representation learning capabilities, effectively handling diverse and complex query conditions. Comprehensive evaluations across four benchmark datasets including ActivityNet, THUMOS14, ActivityNet-Captions, and Charades-STA demonstrate that Grounding-MD establishes new state-of-the-art performance in zero-shot and supervised settings in open-world moment detection scenarios. All source code and trained models will be released.