This paper provides the first discourse parsing experiments with a large language model(LLM) finetuned on corpora annotated in the style of SDRT (Segmented Discourse Representation Theory Asher, 1993; Asher and Lascarides, 2003). The result is a discourse parser, Llamipa (Llama Incremental Parser), that leverages discourse context, leading to substantial performance gains over approaches that use encoder-only models to provide local, context-sensitive representations of discourse units. Furthermore, it can process discourse data incrementally, which is essential for the eventual use of discourse information in downstream tasks.

Building on the Joint-Embedding Predictive Architecture (JEPA) paradigm, a recent self-supervised learning framework that predicts latent representations of masked regions in high-level feature spaces, we propose Audio-JEPA (Audio Joint-Embedding Predictive Architecture), tailored specifically for audio data. Audio-JEPA uses a simple Vision Transformer backbone to predict latent representations of masked spectrogram patches rather than reconstructing raw audio. We pre-train on unlabeled AudioSet clips (10s, 32kHz) with random patch masking on mel-spectrograms. We evaluate on the X-ARES suite covering speech, music, and environmental sound tasks. Although our implementation is a straightforward translation of the original model to audio, the results still show comparable performance to wav2vec 2.0 and data2vec while using less than one-fifth of their training data and with no hyper-parameter tuning. All code and pretrained checkpoints will be released on GitHub.

Whitening is a classical technique in unsupervised learning that can facilitate estimation tasks by standardizing data. An important application is the estimation of latent variable models via the decomposition of tensors built from high-order moments. In particular, whitening orthogonalizes the means of a spherical Gaussian mixture model (GMM), thereby making the corresponding moment tensor orthogonally decomposable, hence easier to decompose. However, in the large-dimensional regime (LDR) where data are high-dimensional and scarce, the standard whitening matrix built from the sample covariance becomes ineffective because the latter is spectrally distorted. Consequently, whitened means of a spherical GMM are no longer orthogonal. Using random matrix theory, we derive exact limits for their dot products, which are generally nonzero in the LDR. As our main contribution, we then construct a corrected whitening matrix that restores asymptotic orthogonality, allowing for performance gains in spherical GMM estimation.

Our paper challenges claims from prior research that transformer-based models, when learning in context, implicitly implement standard learning algorithms. We present empirical evidence inconsistent with this view and provide a mathematical analysis demonstrating that transformers cannot achieve general predictive accuracy due to inherent architectural limitations.

Neural implicit surfaces are a promising tool for geometry processing that represent a solid object as the zero level set of a neural network. Usually trained to approximate a signed distance function of the considered object, these methods exhibit great visual fidelity and quality near the surface, yet their properties tend to degrade with distance, making geometrical queries hard to perform without the help of complex range analysis techniques. Based on recent advancements in Lipschitz neural networks, we introduce a new method for approximating the signed distance function of a given object. As our neural function is made 1- Lipschitz by construction, it cannot overestimate the distance, which guarantees robustness even far from the surface. Moreover, the 1-Lipschitz constraint allows us to use a different loss function, called the hinge-Kantorovitch-Rubinstein loss, which pushes the gradient as close to unit-norm as possible, thus reducing computation costs in iterative queries. As this loss function only needs a rough estimate of occupancy to be optimized, this means that the true distance function need not to be known. We are therefore able to compute neural implicit representations of even bad quality geometry such as noisy point clouds or triangle soups. We demonstrate that our methods is able to approximate the distance function of any closed or open surfaces or curves in the plane or in space, while still allowing sphere tracing or closest point projections to be performed robustly.

Speaker segmentation consists in partitioning a conversation between one or more speakers into speaker turns. Usually addressed as the late combination of three sub-tasks (voice activity detection, speaker change detection, and overlapped speech detection), we propose to train an end-to-end segmentation model that does it directly. Inspired by the original end-to-end neural speaker diarization approach (EEND), the task is modeled as a multi-label classification problem using permutation-invariant training. The main difference is that our model operates on short audio chunks (5 seconds) but at a much higher temporal resolution (every 16ms). Experiments on multiple speaker diarization datasets conclude that our model can be used with great success on both voice activity detection and overlapped speech detection. Our proposed model can also be used as a post-processing step, to detect and correctly assign overlapped speech regions. Relative diarization error rate improvement over the best considered baseline (VBx) reaches 17% on AMI, 13% on DIHARD 3, and 13% on VoxConverse.

While Sharpness-Aware Minimization (SAM) improves generalization in deep neural networks by minimizing both loss and sharpness, it suffers from inefficiency in distributed large-batch training. We present Landscape-Smoothed SAM (LSAM), a novel optimizer that preserves SAM's generalization advantages while offering superior efficiency. LSAM integrates SAM's adversarial steps with an asynchronous distributed sampling strategy, generating an asynchronous distributed sampling scheme, producing a smoothed sharpness-aware loss landscape for optimization. This design eliminates synchronization bottlenecks, accelerates large-batch convergence, and delivers higher final accuracy compared to data-parallel SAM.

Large Language Models (LLMs) are gaining traction as a method to generate consensus statements and aggregate preferences in digital democracy experiments. Yet, LLMs may introduce critical vulnerabilities in these systems. Here, we explore the impact of prompt-injection attacks targeting consensus generating systems by introducing a four-dimensional taxonomy of attacks. We test these attacks using LLaMA 3.1 8B and Chat GPT 4.1 Nano finding the LLMs more vulnerable to criticism attacks -- attacks using disagreeable prompts -- and more effective at tilting ambiguous consensus statements. We also find evidence of more effective manipulation when using explicit imperatives and rational-sounding arguments compared to emotional language or fabricated statistics. To mitigate these vulnerabilities, we apply Direct Preference Optimization (DPO), an alignment method that fine-tunes LLMs to prefer unperturbed consensus statements. While DPO significantly improves robustness, it still offers limited protection against attacks targeting ambiguous consensus. These results advance our understanding of the vulnerability and robustness of consensus generating LLMs in digital democracy applications.

In this paper we introduce Tale, Task-Aware Layer Elimination, an inference-time algorithm that prunes entire transformer layers in an LLM by directly optimizing task-specific validation performance. We evaluate TALE on 9 tasks and 5 models, including LLaMA 3.1 8B, Qwen 2.5 7B, Qwen 2.5 0.5B, Mistral 7B, and Lucie 7B, under both zero-shot and few-shot settings. Unlike prior approaches, TALE requires no retraining and consistently improves accuracy while reducing computational cost across all benchmarks. Furthermore, applying TALE during finetuning leads to additional performance gains. Finally, TALE provides flexible user control over trade-offs between accuracy and efficiency. Mutual information analysis shows that certain layers act as bottlenecks, degrading task-relevant representations. Tale's selective layer removal remedies this problem, producing smaller, faster, and more accurate models that are also faster to fine-tune while offering new insights into transformer interpretability.

This paper introduces a versatile paradigm for integrating multi-view reflectance (optional) and normal maps acquired through photometric stereo. Our approach employs a pixel-wise joint re-parameterization of reflectance and normal, considering them as a vector of radiances rendered under simulated, varying illumination. This re-parameterization enables the seamless integration of reflectance and normal maps as input data in neural volume rendering-based 3D reconstruction while preserving a single optimization objective. In contrast, recent multi-view photometric stereo (MVPS) methods depend on multiple, potentially conflicting objectives. Despite its apparent simplicity, our proposed approach outperforms state-of-the-art approaches in MVPS benchmarks across F-score, Chamfer distance, and mean angular error metrics. Notably, it significantly improves the detailed 3D reconstruction of areas with high curvature or low visibility.

Achieving high-fidelity 3D surface reconstruction while preserving fine details remains challenging, especially in the presence of materials with complex reflectance properties and without a dense-view setup. In this paper, we introduce a versatile framework that incorporates multi-view normal and optionally reflectance maps into radiance-based surface reconstruction. Our approach employs a pixel-wise joint re-parametrization of reflectance and surface normals, representing them as a vector of radiances under simulated, varying illumination. This formulation enables seamless incorporation into standard surface reconstruction pipelines, such as traditional multi-view stereo (MVS) frameworks or modern neural volume rendering (NVR) ones. Combined with the latter, our approach achieves state-of-the-art performance on multi-view photometric stereo (MVPS) benchmark datasets, including DiLiGenT-MV, LUCES-MV and Skoltech3D. In particular, our method excels in reconstructing fine-grained details and handling challenging visibility conditions. The present paper is an extended version of the earlier conference paper by Brument et al. (in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024), featuring an accelerated and more robust algorithm as well as a broader empirical evaluation. The code and data relative to this article is available at this https URL.

Hyperspectral imaging (HSI) captures detailed spectral signatures across hundreds of contiguous bands per pixel, being indispensable for remote sensing applications such as land-cover classification, change detection, and environmental monitoring. Due to the high dimensionality of HSI data and the slow rate of data transfer in satellite-based systems, compact and efficient models are required to support onboard processing and minimize the transmission of redundant or low-value data, e.g. cloud-covered areas. To this end, we introduce a novel curriculum multi-task self-supervised learning (CMTSSL) framework designed for lightweight architectures for HSI analysis. CMTSSL integrates masked image modeling with decoupled spatial and spectral jigsaw puzzle solving, guided by a curriculum learning strategy that progressively increases data complexity during self-supervision. This enables the encoder to jointly capture fine-grained spectral continuity, spatial structure, and global semantic features. Unlike prior dual-task SSL methods, CMTSSL simultaneously addresses spatial and spectral reasoning within a unified and computationally efficient design, being particularly suitable for training lightweight models for onboard satellite deployment. We validate our approach on four public benchmark datasets, demonstrating consistent gains in downstream segmentation tasks, using architectures that are over 16,000x lighter than some state-of-the-art models. These results highlight the potential of CMTSSL in generalizable representation learning with lightweight architectures for real-world HSI applications. Our code is publicly available at this https URL.

Recent works indicate that convolutional neural networks (CNN) need large receptive fields (RF) to compete with visual transformers and their attention mechanism. In CNNs, RFs can simply be enlarged by increasing the convolution kernel sizes. Yet the number of trainable parameters, which scales quadratically with the kernel's size in the 2D case, rapidly becomes prohibitive, and the training is notoriously difficult. This paper presents a new method to increase the RF size without increasing the number of parameters. The dilated convolution (DC) has already been proposed for the same purpose. DC can be seen as a convolution with a kernel that contains only a few non-zero elements placed on a regular grid. Here we present a new version of the DC in which the spacings between the non-zero elements, or equivalently their positions, are no longer fixed but learnable via backpropagation thanks to an interpolation technique. We call this method "Dilated Convolution with Learnable Spacings" (DCLS) and generalize it to the n-dimensional convolution case. However, our main focus here will be on the 2D case. We first tried our approach on ResNet50: we drop-in replaced the standard convolutions with DCLS ones, which increased the accuracy of ImageNet1k classification at iso-parameters, but at the expense of the throughput. Next, we used the recent ConvNeXt state-of-the-art convolutional architecture and drop-in replaced the depthwise convolutions with DCLS ones. This not only increased the accuracy of ImageNet1k classification but also of typical downstream and robustness tasks, again at iso-parameters but this time with negligible cost on throughput, as ConvNeXt uses separable convolutions. Conversely, classic DC led to poor performance with both ResNet50 and ConvNeXt. The code of the method is available at: this https URL.

We propose a novel ranking model that combines the Bradley-Terry-Luce probability model with a nonnegative matrix factorization framework to model and uncover the presence of latent variables that influence the performance of top tennis players. We derive an efficient, provably convergent, and numerically stable majorization-minimization-based algorithm to maximize the likelihood of datasets under the proposed statistical model. The model is tested on datasets involving the outcomes of matches between 20 top male and female tennis players over 14 major tournaments for men (including the Grand Slams and the ATP Masters 1000) and 16 major tournaments for women over the past 10 years. Our model automatically infers that the surface of the court (e.g., clay or hard court) is a key determinant of the performances of male players, but less so for females. Top players on various surfaces over this longitudinal period are also identified in an objective manner.

Self-supervised techniques for learning speech representations have been shown to develop linguistic competence from exposure to speech without the need for human labels. In order to fully realize the potential of these approaches and further our understanding of how infants learn language, simulations must closely emulate real-life situations by training on developmentally plausible corpora and benchmarking against appropriate test sets. To this end, we propose a language-acquisition-friendly benchmark to probe spoken language models at the lexical and syntactic levels, both of which are compatible with the vocabulary typical of children's language experiences. This paper introduces the benchmark and summarizes a range of experiments showing its usefulness. In addition, we highlight two exciting challenges that need to be addressed for further progress: bridging the gap between text and speech and between clean speech and in-the-wild speech.

In this paper, we introduce the Extreme Metal Vocals Dataset, which comprises a collection of recordings of extreme vocal techniques performed within the realm of heavy metal music. The dataset consists of 760 audio excerpts of 1 second to 30 seconds long, totaling about 100 min of audio material, roughly composed of 60 minutes of distorted voices and 40 minutes of clear voice recordings. These vocal recordings are from 27 different singers and are provided without accompanying musical instruments or post-processing effects. The distortion taxonomy within this dataset encompasses four distinct distortion techniques and three vocal effects, all performed in different pitch ranges. Performance of a state-of-the-art deep learning model is evaluated for two different classification tasks related to vocal techniques, demonstrating the potential of this resource for the audio processing community.

Despite their impressive performances, Large Language Models (LLMs) remain prone to hallucination, which critically undermines their trustworthiness. While most of the previous work focused on tackling answer and attribution correctness, a recent line of work investigated faithfulness, with a focus on leveraging internal model signals to reflect a model's actual decision-making process while generating the answer. Nevertheless, these methods induce additional latency and have shown limitations in directly aligning token generation with attribution generation. In this paper, we introduce LoDIT, a method that jointly generates and faithfully attributes answers in RAG by leveraging specific token logits during generation. It consists of two steps: (1) marking the documents with specific token identifiers and then leveraging the logits of these tokens to estimate the contribution of each document to the answer during generation, and (2) aggregating these contributions into document attributions. Experiments on a trustworthiness-focused attributed text-generation benchmark, Trust-Align, show that LoDIT significantly outperforms state-of-the-art models on several metrics. Finally, an in-depth analysis of LoDIT shows both its efficiency in terms of latency and its robustness in different settings.