Emanuele Rodolà

Giorgio Strano, Luca Cerovaz, Michele Mancusi et al.

Modern VAEs are rarely trained with the pointwise likelihood implied by the standard $β$-VAE objective. In practice, pointwise reconstruction is often combined with perceptual and adversarial losses, despite a lack of understanding of how this changes the latent dynamics of the model. We show that the choice of reconstruction loss reshapes the rate-distortion problem itself, altering both the information content and the geometry of the learned latent space in ways that may be invisible from reconstructions alone. First, we prove and verify empirically that augmenting pointwise reconstruction with neural terms, such as perceptual and adversarial objectives, reduces the amount of information stored in the latent representations. Second, we show that neural reconstruction losses systematically change the geometry of the latent space: they make representations more isotropic and distribute uncertainty more evenly across latent dimensions, producing different posterior variance profiles. These findings highlight how the rate-distortion tradeoff is not a comprehensive lens to understand the behavior of VAEs, and we propose a more mechanistic approach to investigate how the choice of a distortion metric reshapes the optimization problem.

Luca Moschella, Valentino Maiorca, Marco Fumero et al. · oxford

Neural networks embed the geometric structure of a data manifold lying in a high-dimensional space into latent representations. Ideally, the distribution of the data points in the latent space should depend only on the task, the data, the loss, and other architecture-specific constraints. However, factors such as the random weights initialization, training hyperparameters, or other sources of randomness in the training phase may induce incoherent latent spaces that hinder any form of reuse. Nevertheless, we empirically observe that, under the same data and modeling choices, the angles between the encodings within distinct latent spaces do not change. In this work, we propose the latent similarity between each sample and a fixed set of anchors as an alternative data representation, demonstrating that it can enforce the desired invariances without any additional training. We show how neural architectures can leverage these relative representations to guarantee, in practice, invariance to latent isometries and rescalings, effectively enabling latent space communication: from zero-shot model stitching to latent space comparison between diverse settings. We extensively validate the generalization capability of our approach on different datasets, spanning various modalities (images, text, graphs), tasks (e.g., classification, reconstruction) and architectures (e.g., CNNs, GCNs, transformers).

Abzal Aidakhmetov, Donato Crisostomi, Tommaso Mencattini et al.

Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \emph{stealth data poisoning attack} silently compromises this pipeline. By substituting $4{-}6\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of $20{-}55\%$, $+19\%$ to $+51\%$ over a clean reference. Finally, we find that a refusal-direction orthogonalization defense can recover ${\approx}82\%$ of the ASR gap without harming benign behavior.

Andrea Santilli, Emanuele Rodolà

In recent years Large Language Models (LLMs) have increased the state of the art on several natural language processing tasks. However, their accessibility is often limited to paid API services, posing challenges for researchers in conducting extensive investigations. On the other hand, while some open-source models have been proposed by the community, they are typically English-centric or multilingual without a specific adaptation for the Italian language. In an effort to democratize the available and open resources for the Italian language, in this paper we introduce Camoscio: a language model specifically tuned to follow users' prompts in Italian. Specifically, we finetuned the smallest variant of LLaMA (7b) with LoRA on a corpus of instruction prompts translated to Italian via ChatGPT. Results indicate that the model's zero-shot performance on various downstream tasks in Italian competes favorably with existing models specifically finetuned for those tasks. All the artifacts (code, dataset, model) are released to the community at the following url: https://github.com/teelinsan/camoscio

Andrea Bacciu, Giovanni Trappolini, Andrea Santilli et al.

This paper presents Fauno, the first and largest open-source Italian conversational Large Language Model (LLM). Our goal with Fauno is to democratize the study of LLMs in Italian, demonstrating that obtaining a fine-tuned conversational bot with a single GPU is possible. In addition, we release a collection of datasets for conversational AI in Italian. The datasets on which we fine-tuned Fauno include various topics such as general question answering, computer science, and medical questions. We release our code and datasets on \url{https://github.com/RSTLess-research/Fauno-Italian-LLM}

Marco Fumero, Florian Wenzel, Luca Zancato et al.

Recovering the latent factors of variation of high dimensional data has so far focused on simple synthetic settings. Mostly building on unsupervised and weakly-supervised objectives, prior work missed out on the positive implications for representation learning on real world data. In this work, we propose to leverage knowledge extracted from a diversified set of supervised tasks to learn a common disentangled representation. Assuming each supervised task only depends on an unknown subset of the factors of variation, we disentangle the feature space of a supervised multi-task model, with features activating sparsely across different tasks and information being shared as appropriate. Importantly, we never directly observe the factors of variations but establish that access to multiple tasks is sufficient for identifiability under sufficiency and minimality assumptions. We validate our approach on six real world distribution shift benchmarks, and different data modalities (images, text), demonstrating how disentangled representations can be transferred to real settings.

Irene Cannistraci, Simone Antonelli, Emanuele Palumbo et al.

Foundation models achieve state-of-the-art performance across different tasks, but their size and computational demands raise concerns about accessibility and sustainability. Existing efficiency methods often require additional retraining or finetuning, limiting their practicality. Recent findings suggest that deep neural networks exhibit internal representation similarities. While such similarities across different models have been exploited for enabling techniques such as model stitching and merging, intra-network redundancy remains underexplored as a source for efficiency gains. In this paper, we introduce Transformer Optimization using Adaptive and Simple Transformations (TOAST), a framework that exploits these redundancies to approximate entire transformer blocks with lightweight closed-form mappings, such as linear transformations or even the identity function, without any additional training. Across state-of-the-art pretrained vision models (e.g., ViT, DINOv2, DeiT) and datasets ranging from MNIST to ImageNet-1k, TOAST reduces parameters and computation while preserving, and in some cases improving, downstream performance. These results show that large portions of transformer depth can be replaced by trivial functions, opening a new perspective on efficient foundation models.

Antonio Norelli, Marco Fumero, Valentino Maiorca et al. · oxford

CLIP proved that aligning visual and language spaces is key to solving many vision tasks without explicit training, but required to train image and text encoders from scratch on a huge dataset. LiT improved this by only training the text encoder and using a pre-trained vision network. In this paper, we show that a common space can be created without any training at all, using single-domain encoders (trained with or without supervision) and a much smaller amount of image-text pairs. Furthermore, our model has unique properties. Most notably, deploying a new version with updated training samples can be done in a matter of seconds. Additionally, the representations in the common space are easily interpretable as every dimension corresponds to the similarity of the input to a unique image-text pair in the multimodal dataset. Experiments on standard zero-shot visual benchmarks demonstrate the typical transfer ability of image-text models. Overall, our method represents a simple yet surprisingly strong baseline for foundation multimodal models, raising important questions on their data efficiency and on the role of retrieval in machine learning.

Niloofar Azizi, Horst Possegger, Emanuele Rodolà et al.

3D human pose estimation is fundamental to understanding human behavior. Recently, promising results have been achieved by graph convolutional networks (GCNs), which achieve state-of-the-art performance and provide rather light-weight architectures. However, a major limitation of GCNs is their inability to encode all the transformations between joints explicitly. To address this issue, we propose a novel spectral GCN using the Möbius transformation (MöbiusGCN). In particular, this allows us to directly and explicitly encode the transformation between joints, resulting in a significantly more compact representation. Compared to even the lightest architectures so far, our novel approach requires 90-98% fewer parameters, i.e. our lightest MöbiusGCN uses only 0.042M trainable parameters. Besides the drastic parameter reduction, explicitly encoding the transformation of joints also enables us to achieve state-of-the-art results. We evaluate our approach on the two challenging pose estimation benchmarks, Human3.6M and MPI-INF-3DHP, demonstrating both state-of-the-art results and the generalization capabilities of MöbiusGCN.

Valentino Maiorca, Luca Moschella, Antonio Norelli et al. · oxford

While different neural models often exhibit latent spaces that are alike when exposed to semantically related data, this intrinsic similarity is not always immediately discernible. Towards a better understanding of this phenomenon, our work shows how representations learned from these neural modules can be translated between different pre-trained networks via simpler transformations than previously thought. An advantage of this approach is the ability to estimate these transformations using standard, well-understood algebraic procedures that have closed-form solutions. Our method directly estimates a transformation between two given latent spaces, thereby enabling effective stitching of encoders and decoders without additional training. We extensively validate the adaptability of this translation procedure in different experimental settings: across various trainings, domains, architectures (e.g., ResNet, CNN, ViT), and in multiple downstream tasks (classification, reconstruction). Notably, we show how it is possible to zero-shot stitch text encoders and vision decoders, or vice-versa, yielding surprisingly good classification performance in this multimodal setting.

Irene Cannistraci, Luca Moschella, Marco Fumero et al. · eth-zurich

It has been observed that representations learned by distinct neural networks conceal structural similarities when the models are trained under similar inductive biases. From a geometric perspective, identifying the classes of transformations and the related invariances that connect these representations is fundamental to unlocking applications, such as merging, stitching, and reusing different neural modules. However, estimating task-specific transformations a priori can be challenging and expensive due to several factors (e.g., weights initialization, training hyperparameters, or data modality). To this end, we introduce a versatile method to directly incorporate a set of invariances into the representations, constructing a product space of invariant components on top of the latent representations without requiring prior knowledge about the optimal invariance to infuse. We validate our solution on classification and reconstruction tasks, observing consistent latent similarity and downstream performance improvements in a zero-shot stitching setting. The experimental analysis comprises three modalities (vision, text, and graphs), twelve pretrained foundational models, nine benchmarks, and several architectures trained from scratch.

Irene Cannistraci, Luca Moschella, Valentino Maiorca et al. · eth-zurich, oxford

The use of relative representations for latent embeddings has shown potential in enabling latent space communication and zero-shot model stitching across a wide range of applications. Nevertheless, relative representations rely on a certain amount of parallel anchors to be given as input, which can be impractical to obtain in certain scenarios. To overcome this limitation, we propose an optimization-based method to discover new parallel anchors from a limited known set (seed). Our approach can be used to find semantic correspondence between different domains, align their relative spaces, and achieve competitive results in several tasks.

Jakub Proboszcz, Paweł Kochanski, Karol Korszun et al.

Generative audio models, based on diffusion and autoregressive architectures, have advanced rapidly in both quality and expressiveness. This progress, however, raises pressing copyright concerns, as such models are often trained on vast corpora of artistic and commercial works. A central question is whether one can reliably verify if an artist's material was included in training, thereby providing a means for copyright holders to protect their content. In this work, we investigate the feasibility of such verification through membership inference attacks (MIA) on open-source generative audio models, which attempt to determine whether a specific audio sample was part of the training set. Our empirical results show that membership inference alone is of limited effectiveness at scale, as the per-sample membership signal is weak for models trained on large and diverse datasets. However, artists and media owners typically hold collections of works rather than isolated samples. Building on prior work in text and vision domains, in this work we focus on dataset inference (DI), which aggregates diverse membership evidence across multiple samples. We find that DI is successful in the audio domain, offering a more practical mechanism for assessing whether an artist's works contributed to model training. Our results suggest DI as a promising direction for copyright protection and dataset accountability in the era of large audio generative models.

Donato Crisostomi, Irene Cannistraci, Luca Moschella et al. · eth-zurich

Models trained on semantically related datasets and tasks exhibit comparable inter-sample relations within their latent spaces. We investigate in this study the aggregation of such latent spaces to create a unified space encompassing the combined information. To this end, we introduce Relative Latent Space Aggregation, a two-step approach that first renders the spaces comparable using relative representations, and then aggregates them via a simple mean. We carefully divide a classification problem into a series of learning tasks under three different settings: sharing samples, classes, or neither. We then train a model on each task and aggregate the resulting latent spaces. We compare the aggregated space with that derived from an end-to-end model trained over all tasks and show that the two spaces are similar. We then observe that the aggregated space is better suited for classification, and empirically demonstrate that it is due to the unique imprints left by task-specific embedders within the representations. We finally test our framework in scenarios where no shared region exists and show that it can still be used to merge the spaces, albeit with diminished benefits over naive merging.

Giorgio Mariani, Irene Tallini, Emilian Postolache et al.

In this work, we define a diffusion-based generative model capable of both music synthesis and source separation by learning the score of the joint probability density of sources sharing a context. Alongside the classic total inference tasks (i.e., generating a mixture, separating the sources), we also introduce and experiment on the partial generation task of source imputation, where we generate a subset of the sources given the others (e.g., play a piano track that goes well with the drums). Additionally, we introduce a novel inference method for the separation task based on Dirac likelihood functions. We train our model on Slakh2100, a standard dataset for musical source separation, provide qualitative results in the generation settings, and showcase competitive quantitative results in the source separation setting. Our method is the first example of a single model that can handle both generation and separation tasks, thus representing a step toward general audio models.

Giovanni Trappolini, Valentino Maiorca, Silvio Severino et al.

Graph Neural Networks (GNNs) have proven to be successful in several predictive modeling tasks for graph-structured data. Amongst those tasks, link prediction is one of the fundamental problems for many real-world applications, such as recommender systems. However, GNNs are not immune to adversarial attacks, i.e., carefully crafted malicious examples that are designed to fool the predictive model. In this work, we focus on a specific, white-box attack to GNN-based link prediction models, where a malicious node aims to appear in the list of recommended nodes for a given target victim. To achieve this goal, the attacker node may also count on the cooperation of other existing peers that it directly controls, namely on the ability to inject a number of ``vicious'' nodes in the network. Specifically, all these malicious nodes can add new edges or remove existing ones, thereby perturbing the original graph. Thus, we propose SAVAGE, a novel framework and a method to mount this type of link prediction attacks. SAVAGE formulates the adversary's goal as an optimization task, striking the balance between the effectiveness of the attack and the sparsity of malicious resources required. Extensive experiments conducted on real-world and synthetic datasets demonstrate that adversarial attacks implemented through SAVAGE indeed achieve high attack success rate yet using a small amount of vicious nodes. Finally, despite those attacks require full knowledge of the target model, we show that they are successfully transferable to other black-box methods for link prediction.

Marco Comunità, Riccardo F. Gramaccioni, Emilian Postolache et al.

Sound design involves creatively selecting, recording, and editing sound effects for various media like cinema, video games, and virtual/augmented reality. One of the most time-consuming steps when designing sound is synchronizing audio with video. In some cases, environmental recordings from video shoots are available, which can aid in the process. However, in video games and animations, no reference audio exists, requiring manual annotation of event timings from the video. We propose a system to extract repetitive actions onsets from a video, which are then used - in conjunction with audio or textual embeddings - to condition a diffusion model trained to generate a new synchronized sound effects audio track. In this way, we leave complete creative control to the sound designer while removing the burden of synchronization with video. Furthermore, editing the onset track or changing the conditioning embedding requires much less effort than editing the audio track itself, simplifying the sonification process. We provide sound examples, source code, and pretrained models to faciliate reproducibility

Marco Pegoraro, Sanketh Vedula, Aviv A. Rosenberg et al.

Quantile regression (QR) is a statistical tool for distribution-free estimation of conditional quantiles of a target variable given explanatory features. QR is limited by the assumption that the target distribution is univariate and defined on an Euclidean domain. Although the notion of quantiles was recently extended to multi-variate distributions, QR for multi-variate distributions on manifolds remains underexplored, even though many important applications inherently involve data distributed on, e.g., spheres (climate and geological phenomena), and tori (dihedral angles in proteins). By leveraging optimal transport theory and c-concave functions, we meaningfully define conditional vector quantile functions of high-dimensional variables on manifolds (M-CVQFs). Our approach allows for quantile estimation, regression, and computation of conditional confidence sets and likelihoods. We demonstrate the approach's efficacy and provide insights regarding the meaning of non-Euclidean quantiles through synthetic and real data experiments.

Emanuele Frascaroli, Riccardo Benaglia, Matteo Boschini et al.

While biological intelligence grows organically as new knowledge is gathered throughout life, Artificial Neural Networks forget catastrophically whenever they face a changing training data distribution. Rehearsal-based Continual Learning (CL) approaches have been established as a versatile and reliable solution to overcome this limitation; however, sudden input disruptions and memory constraints are known to alter the consistency of their predictions. We study this phenomenon by investigating the geometric characteristics of the learner's latent space and find that replayed data points of different classes increasingly mix up, interfering with classification. Hence, we propose a geometric regularizer that enforces weak requirements on the Laplacian spectrum of the latent space, promoting a partitioning behavior. Our proposal, called Continual Spectral Regularizer for Incremental Learning (CaSpeR-IL), can be easily combined with any rehearsal-based CL approach and improves the performance of SOTA methods on standard benchmarks.

Daniele Baieri, Stefano Esposito, Filippo Maggioli et al.

Representing 3D surfaces as level sets of continuous functions over $\mathbb{R}^3$ is the common denominator of neural implicit representations, which recently enabled remarkable progress in geometric deep learning and computer vision tasks. In order to represent 3D motion within this framework, it is often assumed (either explicitly or implicitly) that the transformations which a surface may undergo are homeomorphic: this is not necessarily true, for instance, in the case of fluid dynamics. In order to represent more general classes of deformations, we propose to apply this theoretical framework as regularizers for the optimization of simple 4D implicit functions (such as signed distance fields). We show that our representation is capable of capturing both homeomorphic and topology-changing deformations, while also defining correspondences over the continuously-reconstructed surfaces.

Stefano Esposito, Daniele Baieri, Stefan Zellmann et al.

NeRF-based techniques fit wide and deep multi-layer perceptrons (MLPs) to a continuous radiance field that can be rendered from any unseen viewpoint. However, the lack of surface and normals definition and high rendering times limit their usage in typical computer graphics applications. Such limitations have recently been overcome separately, but solving them together remains an open problem. We present KiloNeuS, a neural representation reconstructing an implicit surface represented as a signed distance function (SDF) from multi-view images and enabling real-time rendering by partitioning the space into thousands of tiny MLPs fast to inference. As we learn the implicit surface locally using independent models, resulting in a globally coherent geometry is non-trivial and needs to be addressed during training. We evaluate rendering performance on a GPU-accelerated ray-caster with in-shader neural network inference, resulting in an average of 46 FPS at high resolution, proving a satisfying tradeoff between storage costs and rendering quality. In fact, our evaluation for rendering quality and surface recovery shows that KiloNeuS outperforms its single-MLP counterpart. Finally, to exhibit the versatility of KiloNeuS, we integrate it into an interactive path-tracer taking full advantage of its surface normals. We consider our work a crucial first step toward real-time rendering of implicit neural representations under global illumination.

Michele Miranda, Elena Sofia Ruzzetti, Andrea Santilli et al.

Large Language Models (LLMs) represent a significant advancement in artificial intelligence, finding applications across various domains. However, their reliance on massive internet-sourced datasets for training brings notable privacy issues, which are exacerbated in critical domains (e.g., healthcare). Moreover, certain application-specific scenarios may require fine-tuning these models on private data. This survey critically examines the privacy threats associated with LLMs, emphasizing the potential for these models to memorize and inadvertently reveal sensitive information. We explore current threats by reviewing privacy attacks on LLMs and propose comprehensive solutions for integrating privacy mechanisms throughout the entire learning pipeline. These solutions range from anonymizing training datasets to implementing differential privacy during training or inference and machine unlearning after training. Our comprehensive review of existing literature highlights ongoing challenges, available tools, and future directions for preserving privacy in LLMs. This work aims to guide the development of more secure and trustworthy AI systems by providing a thorough understanding of privacy preservation methods and their effectiveness in mitigating risks.

Marco Pegoraro, Riccardo Marin, Arianna Rampini et al.

In graph learning, maps between graphs and their subgraphs frequently arise. For instance, when coarsening or rewiring operations are present along the pipeline, one needs to keep track of the corresponding nodes between the original and modified graphs. Classically, these maps are represented as binary node-to-node correspondence matrices and used as-is to transfer node-wise features between the graphs. In this paper, we argue that simply changing this map representation can bring notable benefits to graph learning tasks. Drawing inspiration from recent progress in geometry processing, we introduce a spectral representation for maps that is easy to integrate into existing graph learning models. This spectral representation is a compact and straightforward plug-in replacement and is robust to topological changes of the graphs. Remarkably, the representation exhibits structural properties that make it interpretable, drawing an analogy with recent results on smooth manifolds. We demonstrate the benefits of incorporating spectral maps in graph learning pipelines, addressing scenarios where a node-to-node map is not well defined, or in the absence of exact isomorphism. Our approach bears practical benefits in knowledge distillation and hierarchical learning, where we show comparable or improved performance at a fraction of the computational cost.

Emilian Postolache, Giorgio Mariani, Michele Mancusi et al.

Autoregressive models have achieved impressive results over a wide range of domains in terms of generation quality and downstream task performance. In the continuous domain, a key factor behind this success is the usage of quantized latent spaces (e.g., obtained via VQ-VAE autoencoders), which allow for dimensionality reduction and faster inference times. However, using existing pre-trained models to perform new non-trivial tasks is difficult since it requires additional fine-tuning or extensive training to elicit prompting. This paper introduces LASS as a way to perform vector-quantized Latent Autoregressive Source Separation (i.e., de-mixing an input signal into its constituent sources) without requiring additional gradient-based optimization or modifications of existing models. Our separation method relies on the Bayesian formulation in which the autoregressive models are the priors, and a discrete (non-parametric) likelihood function is constructed by performing frequency counts over latent sums of addend tokens. We test our method on images and audio with several sampling strategies (e.g., ancestral, beam search) showing competitive results with existing approaches in terms of separation quality while offering at the same time significant speedups in terms of inference time and scalability to higher dimensional data.

Donato Crisostomi, Simone Antonelli, Valentino Maiorca et al.

Many modern deep-learning techniques do not work without enormous datasets. At the same time, several fields demand methods working in scarcity of data. This problem is even more complex when the samples have varying structures, as in the case of graphs. Graph representation learning techniques have recently proven successful in a variety of domains. Nevertheless, the employed architectures perform miserably when faced with data scarcity. On the other hand, few-shot learning allows employing modern deep learning models in scarce data regimes without waiving their effectiveness. In this work, we tackle the problem of few-shot graph classification, showing that equipping a simple distance metric learning baseline with a state-of-the-art graph embedder allows to obtain competitive results on the task. While the simplicity of the architecture is enough to outperform more complex ones, it also allows straightforward additions. To this end, we show that additional improvements may be obtained by encouraging a task-conditioned embedding space. Finally, we propose a MixUp-based online data augmentation technique acting in the latent space and show its effectiveness on the task.

Akshit Achara, Tatiana Gaintseva, Mateo Mahaut et al.

The Platonic Representation Hypothesis suggests that independently trained neural networks converge to increasingly similar latent spaces. However, current strategies for mapping these representations are inherently pairwise, scaling quadratically with the number of models and failing to yield a consistent global reference. In this paper, we study the alignment of $M \ge 3$ models. We first adapt Generalized Procrustes Analysis (GPA) to construct a shared orthogonal universe that preserves the internal geometry essential for tasks like model stitching. We then show that strict isometric alignment is suboptimal for retrieval, where agreement-maximizing methods like Canonical Correlation Analysis (CCA) typically prevail. To bridge this gap, we finally propose Geometry-Corrected Procrustes Alignment (GCPA), which establishes a robust GPA-based universe followed by a post-hoc correction for directional mismatch. Extensive experiments demonstrate that GCPA consistently improves any-to-any retrieval while retaining a practical shared reference space.

Luca Zhou, Bo Zhao, Rose Yu et al.

Model merging combines knowledge from separately fine-tuned models, yet success factors remain poorly understood. While recent work treats mergeability as an intrinsic property, we show with an architecture-agnostic framework that it fundamentally depends on both the merging method and the partner tasks. Using linear optimization over a set of interpretable pairwise metrics (e.g., gradient L2 distance), we uncover properties correlating with post-merge performance across four merging methods. We find substantial variation in success drivers (46.7% metric overlap; 55.3% sign agreement), revealing method-specific "fingerprints". Crucially, however, subspace overlap and gradient alignment metrics consistently emerge as foundational, method-agnostic prerequisites for compatibility. These findings provide a diagnostic foundation for understanding mergeability and motivate future fine-tuning strategies that explicitly encourage these properties.

Davide Marincione, Donato Crisostomi, Roberto Dessi et al.

Foundation models capable of generalizing across species and tasks represent a promising new frontier in bioacoustics, with NatureLM being one of the most prominent examples. While its domain-specific fine-tuning yields strong performance on bioacoustic benchmarks, we observe that it also introduces trade-offs in instruction-following flexibility. For instance, NatureLM achieves high accuracy when prompted for either the common or scientific name individually, but its accuracy drops significantly when both are requested in a single prompt. We address this by applying a simple model merging strategy that interpolates NatureLM with its base language model, recovering instruction-following capabilities with minimal loss of domain expertise. Finally, we show that the merged model exhibits markedly stronger zero-shot generalization, achieving over a 200% relative improvement and setting a new state-of-the-art in closed-set zero-shot classification of unseen species.

Filippo Rinaldi, Giacomo Capitani, Lorenzo Bonicelli et al.

Foundation models serve as the backbone for numerous specialized models developed through fine-tuning. However, when the underlying pretrained model is updated or retrained (e.g., on larger and more curated datasets), the fine-tuned model becomes obsolete, losing its utility and requiring retraining. This raises the question: is it possible to transfer fine-tuning to a new release of the model? In this work, we investigate how to transfer fine-tuning to a new checkpoint without having to re-train, in a data-free manner. To do so, we draw principles from model re-basin and provide a recipe based on weight permutations to re-base the modifications made to the original base model, often called task vector. In particular, our approach tailors model re-basin for Transformer models, taking into account the challenges of residual connections and multi-head attention layers. Specifically, we propose a two-level method rooted in spectral theory, initially permuting the attention heads and subsequently adjusting parameters within select pairs of heads. Through extensive experiments on visual and textual tasks, we achieve the seamless transfer of fine-tuned knowledge to new pre-trained backbones without relying on a single training step or datapoint. Code is available at https://github.com/aimagelab/TransFusion.

Tommaso Mencattini, Adrian Robert Minut, Donato Crisostomi et al.

Evolutionary model merging enables the creation of high-performing multi-task models but remains computationally prohibitive for consumer hardware. We introduce MERGE$^3$, an efficient framework that makes evolutionary merging feasible on a single GPU by reducing fitness computation costs 50$\times$ while preserving performance. MERGE$^3$ achieves this by Extracting a reduced dataset for evaluation, Estimating model abilities using Item Response Theory (IRT), and Evolving optimal merges via IRT-based performance estimators. Our method enables state-of-the-art multilingual and cross-lingual merging, transferring knowledge across languages with significantly lower computational overhead. We provide theoretical guarantees and an open-source library, democratizing high-quality model merging.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Souhail Hadgi, Luca Moschella, Andrea Santilli et al.

Recent works have shown that, when trained at scale, uni-modal 2D vision and text encoders converge to learned features that share remarkable structural properties, despite arising from different representations. However, the role of 3D encoders with respect to other modalities remains unexplored. Furthermore, existing 3D foundation models that leverage large datasets are typically trained with explicit alignment objectives with respect to frozen encoders from other representations. In this work, we investigate the possibility of a posteriori alignment of representations obtained from uni-modal 3D encoders compared to text-based feature spaces. We show that naive post-training feature alignment of uni-modal text and 3D encoders results in limited performance. We then focus on extracting subspaces of the corresponding feature spaces and discover that by projecting learned representations onto well-chosen lower-dimensional subspaces the quality of alignment becomes significantly higher, leading to improved accuracy on matching and retrieval tasks. Our analysis further sheds light on the nature of these shared subspaces, which roughly separate between semantic and geometric data representations. Overall, ours is the first work that helps to establish a baseline for post-training alignment of 3D uni-modal and text feature spaces, and helps to highlight both the shared and unique properties of 3D data compared to other representations. Our code and weights are available at https://github.com/Souhail-01/3d-text-alignment

Giovanni Trappolini, Andrea Santilli, Emanuele Rodolà et al.

The rise in loosely-structured data available through text, images, and other modalities has called for new ways of querying them. Multimedia Information Retrieval has filled this gap and has witnessed exciting progress in recent years. Tasks such as search and retrieval of extensive multimedia archives have undergone massive performance improvements, driven to a large extent by recent developments in multimodal deep learning. However, methods in this field remain limited in the kinds of queries they support and, in particular, their inability to answer database-like queries. For this reason, inspired by recent work on neural databases, we propose a new framework, which we name Multimodal Neural Databases (MMNDBs). MMNDBs can answer complex database-like queries that involve reasoning over different input modalities, such as text and images, at scale. In this paper, we present the first architecture able to fulfill this set of requirements and test it with several baselines, showing the limitations of currently available models. The results show the potential of these new techniques to process unstructured data coming from different modalities, paving the way for future research in the area. Code to replicate the experiments will be released at https://github.com/GiovanniTRA/MultimodalNeuralDatabases

Emanuele Rossi, Emanuele Rodolà

Natural language is widely used to describe, prompt, and control audio systems, but rarely serves as the representation carrying audio itself. We introduce lexical acoustic coding (LAC), a framework in which pre-trained LLM sender and receiver agents transmit sound through natural language. Under fixed system prompts, the agents write their own analysis and synthesis code, communicating only through a lexical sentence, shared vocabulary, and optional symbolic music structure. The sender analyzes an input waveform into interpretable, non-learned acoustic descriptors, quantizes each with a feature-specific interval vocabulary, and verbalizes the lexical code as English. The receiver parses the sentence back into lexical-acoustic constraints and renders a waveform through closed-loop refinement. The transmitted text serves as both a rich caption and as the transport representation itself. We frame LAC as a finite-rate lossy quantizer, exposing trade-offs between vocabulary size, rate, and fidelity. Experiments on short sounds and symbolic music transfer show that plain text preserves measurable acoustic structure while remaining interpretable, editable, and native to LLM-mediated communication.

Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli et al.

Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to 10% of their original size while retaining 99% of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods.

Davide Marincione, Michele Mancusi, Giorgio Strano et al.

Stem retrieval, the task of matching missing stems to a given audio submix, is a key challenge currently limited by models that discard temporal information. We introduce PHALAR, a contrastive framework achieving a relative accuracy increase of up to $\approx 70\%$ over the state-of-the-art while requiring $<50\%$ of the parameters and a 7$\times$ training speedup. By utilizing a Learned Spectral Pooling layer and a complex-valued head, PHALAR enforces pitch-equivariant and phase-equivariant biases. PHALAR establishes new retrieval state-of-the-art across MoisesDB, Slakh, and ChocoChorales, correlating significantly higher with human coherence judgment than semantic baselines. Finally, zero-shot beat tracking and linear chord probing confirm that PHALAR captures robust musical structures beyond the retrieval task.

Francesco Palandra, Andrea Sanchietti, Daniele Baieri et al.

We present GSEdit, a pipeline for text-guided 3D object editing based on Gaussian Splatting models. Our method enables the editing of the style and appearance of 3D objects without altering their main details, all in a matter of minutes on consumer hardware. We tackle the problem by leveraging Gaussian splatting to represent 3D scenes, and we optimize the model while progressively varying the image supervision by means of a pretrained image-based diffusion model. The input object may be given as a 3D triangular mesh, or directly provided as Gaussians from a generative model such as DreamGaussian. GSEdit ensures consistency across different viewpoints, maintaining the integrity of the original object's information. Compared to previously proposed methods relying on NeRF-like MLP models, GSEdit stands out for its efficiency, making 3D editing tasks much faster. Our editing process is refined via the application of the SDS loss, ensuring that our edits are both precise and accurate. Our comprehensive evaluation demonstrates that GSEdit effectively alters object shape and appearance following the given textual instructions while preserving their coherence and detail.

Ruben Ciranni, Giorgio Mariani, Michele Mancusi et al.

We present COCOLA (Coherence-Oriented Contrastive Learning for Audio), a contrastive learning method for musical audio representations that captures the harmonic and rhythmic coherence between samples. Our method operates at the level of the stems composing music tracks and can input features obtained via Harmonic-Percussive Separation (HPS). COCOLA allows the objective evaluation of generative models for music accompaniment generation, which are difficult to benchmark with established metrics. In this regard, we evaluate recent music accompaniment generation models, demonstrating the effectiveness of the proposed method. We release the model checkpoints trained on public datasets containing separate stems (MUSDB18-HQ, MoisesDB, Slakh2100, and CocoChorales).

Emilian Postolache, Giorgio Mariani, Luca Cosmo et al.

Multi-Source Diffusion Models (MSDM) allow for compositional musical generation tasks: generating a set of coherent sources, creating accompaniments, and performing source separation. Despite their versatility, they require estimating the joint distribution over the sources, necessitating pre-separated musical data, which is rarely available, and fixing the number and type of sources at training time. This paper generalizes MSDM to arbitrary time-domain diffusion models conditioned on text embeddings. These models do not require separated data as they are trained on mixtures, can parameterize an arbitrary number of sources, and allow for rich semantic control. We propose an inference procedure enabling the coherent generation of sources and accompaniments. Additionally, we adapt the Dirac separator of MSDM to perform source separation. We experiment with diffusion models trained on Slakh2100 and MTG-Jamendo, showcasing competitive generation and separation results in a relaxed data setting.

Luca Zhou, Daniele Solombrino, Donato Crisostomi et al.

Model merging has emerged as a cost-efficient approximation to multitask learning. Among merging strategies, task arithmetic is notable for its simplicity and effectiveness. In this work, we provide a theoretical motivation for task vectors by highlighting that, under single-epoch full-batch gradient descent, they are equivalent to multitask gradients. This insight leads us to reinterpret model merging as a single step in an iterative procedure that Alternates between Tuning and Merging (ATM). We propose two applications of ATM: (1) as an alternative to multitask learning in scenarios where data sharing is restricted (e.g., federated settings), and (2) as a lightweight refinement step to improve existing model merging methods using a small validation set. Experiments across diverse vision tasks demonstrate the effectiveness of ATM.

Emilian Postolache, Natalia Polouliakh, Hiroaki Kitano et al.

In this article, we explore the potential of using latent diffusion models, a family of powerful generative models, for the task of reconstructing naturalistic music from electroencephalogram (EEG) recordings. Unlike simpler music with limited timbres, such as MIDI-generated tunes or monophonic pieces, the focus here is on intricate music featuring a diverse array of instruments, voices, and effects, rich in harmonics and timbre. This study represents an initial foray into achieving general music reconstruction of high-quality using non-invasive EEG data, employing an end-to-end training approach directly on raw data without the need for manual pre-processing and channel selection. We train our models on the public NMED-T dataset and perform quantitative evaluation proposing neural embedding-based metrics. Our work contributes to the ongoing research in neural decoding and brain-computer interfaces, offering insights into the feasibility of using EEG data for complex auditory information reconstruction.

Lorenzo Basile, Valentino Maiorca, Luca Bortolussi et al.

When examined through the lens of their residual streams, a puzzling property emerges in transformer networks: residual contributions (e.g., attention heads) sometimes specialize in specific tasks or input attributes. In this paper, we analyze this phenomenon in vision transformers, focusing on the spectral geometry of residuals, and explore its implications for modality alignment in vision-language models. First, we link it to the intrinsically low-dimensional structure of visual head representations, zooming into their principal components and showing that they encode specialized roles across a wide variety of input data distributions. Then, we analyze the effect of head specialization in multimodal models, focusing on how improved alignment between text and specialized heads impacts zero-shot classification performance. This specialization-performance link consistently holds across diverse pre-training data, network sizes, and objectives, demonstrating a powerful new mechanism for boosting zero-shot classification through targeted alignment. Ultimately, we translate these insights into actionable terms by introducing ResiDual, a technique for spectral alignment of the residual stream. Much like panning for gold, it lets the noise from irrelevant unit principal components (i.e., attributes) wash away to amplify task-relevant ones. Remarkably, this dual perspective on modality alignment yields fine-tuning level performance on different data distributions while modelling an extremely interpretable and parameter-efficient transformation, as we extensively show on 70 pre-trained network-dataset combinations (7 models, 10 datasets).

Giorgos Nikolaou, Tommaso Mencattini, Donato Crisostomi et al.

Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the input from a model's representations. In this paper, we challenge this view. First, we prove mathematically that transformer language models mapping discrete input sequences to their corresponding sequence of continuous representations are injective and therefore lossless, a property established at initialization and preserved during training. Second, we confirm this result empirically through billions of collision tests on six state-of-the-art language models, and observe no collisions. Third, we operationalize injectivity: we introduce SipIt, the first algorithm that provably and efficiently reconstructs the exact input text from hidden activations, establishing linear-time guarantees and demonstrating exact invertibility in practice. Overall, our work establishes injectivity as a fundamental and exploitable property of language models, with direct implications for transparency, interpretability, and safe deployment.

Marco Fumero, Luca Moschella, Emanuele Rodolà et al.

Neural networks transform high-dimensional data into compact, structured representations, often modeled as elements of a lower dimensional latent space. In this paper, we present an alternative interpretation of neural models as dynamical systems acting on the latent manifold. Specifically, we show that autoencoder models implicitly define a latent vector field on the manifold, derived by iteratively applying the encoding-decoding map, without any additional training. We observe that standard training procedures introduce inductive biases that lead to the emergence of attractor points within this vector field. Drawing on this insight, we propose to leverage the vector field as a representation for the network, providing a novel tool to analyze the properties of the model and the data. This representation enables to: (i) analyze the generalization and memorization regimes of neural models, even throughout training; (ii) extract prior knowledge encoded in the network's parameters from the attractors, without requiring any input data; (iii) identify out-of-distribution samples from their trajectories in the vector field. We further validate our approach on vision foundation models, showcasing the applicability and effectiveness of our method in real-world scenarios.

Antonio Pio Ricciardi, Valentino Maiorca, Luca Moschella et al.

Visual Reinforcement Learning is a popular and powerful framework that takes full advantage of the Deep Learning breakthrough. It is known that variations in input domains (e.g., different panorama colors due to seasonal changes) or task domains (e.g., altering the target speed of a car) can disrupt agent performance, necessitating new training for each variation. Recent advancements in the field of representation learning have demonstrated the possibility of combining components from different neural networks to create new models in a zero-shot fashion. In this paper, we build upon relative representations, a framework that maps encoder embeddings to a universal space. We adapt this framework to the Visual Reinforcement Learning setting, allowing to combine agents components to create new agents capable of effectively handling novel visual-task pairs not encountered during training. Our findings highlight the potential for model reuse, significantly reducing the need for retraining and, consequently, the time and computational resources required.

Daniele Baieri, Riccardo Cicciarella, Michael Krützen et al.

We address the problem of estimating the metric 3D shape and motion of wild dolphins from monocular video, with the aim of assessing their body condition. While considerable progress has been made in reconstructing 3D models of terrestrial quadrupeds, aquatic animals remain unexplored due to the difficulty of observing them in their natural underwater environment. To address this, we propose a model-based approach that incorporates a transmission model to account for water-induced occlusion. We apply our method to video captured under different sea conditions. We estimate mass and volume, and compare our results to a manual 2D measurements-based method. Additionally, we apply our method to video of captive animals with known ground truth mass. While in our experiments the manual approach is often more accurate, our method demonstrates a distinct advantage when applied to larger specimen. These findings highlight the potential of our method as a scalable and automated alternative for mass and volume estimation of dolphins from monocular video.

Luca Zhou, Daniele Solombrino, Donato Crisostomi et al.

Task arithmetic has emerged as a simple yet powerful technique for model merging, enabling the combination of multiple finetuned models into one. Despite its empirical success, a clear theoretical explanation of why and when it works is lacking. This paper provides a rigorous theoretical foundation for task arithmetic by establishing a connection between task vectors and gradients of the task losses. We show that under standard gradient descent, a task vector generated from one epoch of finetuning is exactly equivalent to the negative gradient of the loss, scaled by the learning rate. For the practical multi-epoch setting, we prove that this equivalence holds approximately, with a second-order error term that we explicitly bound for feed-forward networks. Our empirical analysis across seven vision benchmarks corroborates our theory, demonstrating that the first-epoch gradient dominates the finetuning trajectory in both norm and direction. A key implication is that merging models finetuned for only a single epoch often yields performance comparable to merging fully converged models. These findings reframe task arithmetic as a form of approximate multitask learning, providing a clear rationale for its effectiveness and highlighting the critical role of early training dynamics in model merging.

Francesco Pappone, Donato Crisostomi, Emanuele Rodolà

Recurrent-depth transformers scale test-time compute by iterating latent computations before emitting tokens. We study the geometry of these iterates and argue for a simple, two-scale operational picture: (i) within a looped block, updates act as small-scale refinements; (ii) across consecutive blocks, states undergo a larger-scale drift. Across training, our measurements show that loop steps become smaller and increasingly orthogonal to one another, indicating better local modeling of fine structure rather than merely pushing in a single direction. These dynamics motivate an early-exit mechanism based on the model's second-order difference in step-size, which we show is superior in terms of performance, stability and time-efficiency, when compared to the KL-divergence exit strategy of Geiping et al. and its naive first-order counterpart.

Simone Facchiano, Stefano Saravalle, Matteo Migliarini et al.

Video generative models democratize the creation of visual content through intuitive instruction following, but they also inherit the biases and harmful concepts embedded within their web-scale training data. This inheritance creates a significant risk, as users can readily generate undesirable and even illegal content. This work introduces the first unlearning technique tailored explicitly for video diffusion models to address this critical issue. Our method requires 5 multi-modal prompt pairs only. Each pair contains a "safe" and an "unsafe" example that differ only by the target concept. Averaging their per-layer latent differences produces a "refusal vector", which, once subtracted from the model parameters, neutralizes the unsafe concept. We introduce a novel low-rank factorization approach on the covariance difference of embeddings that yields robust refusal vectors. This isolates the target concept while minimizing collateral unlearning of other semantics, thus preserving the visual quality of the generated video. Our method preserves the model's generation quality while operating without retraining or access to the original training data. By embedding the refusal direction directly into the model's weights, the suppression mechanism becomes inherently more robust against adversarial bypass attempts compared to surface-level input-output filters. In a thorough qualitative and quantitative evaluation, we show that we can neutralize a variety of harmful contents, including explicit nudity, graphic violence, copyrights, and trademarks. Project page: https://www.pinlab.org/video-unlearning.

Antonio D'Orazio, Maria Rosaria Briglia, Donato Crisostomi et al.

CLIP is a discriminative model trained to align images and text in a shared embedding space. Due to its multimodal structure, it serves as the backbone of many generative pipelines, where a decoder is trained to map from the shared space back to images. In this work, we show that image synthesis is nevertheless possible using CLIP alone -- without any decoder, training, or fine-tuning. Our approach optimizes a frequency-aware implicit neural representation that encourages coarse-to-fine generation by stratifying frequencies across network layers. To stabilize this inverse mapping, we introduce adversarially robust initialization, a lightweight Orthogonal Procrustes projection to align local text and image embeddings, and a blending loss that anchors outputs to natural image statistics. Without altering CLIP's weights, this framework unlocks capabilities such as text-to-image generation, style transfer, and image reconstruction. These findings suggest that discriminative models may hold untapped generative potential, hidden in plain sight.