Pubblicazioni scientifiche
Tutte le pubblicazioni in Area Science Park
When Seeing Overrides Knowing: Disentangling Knowledge Conflicts in Vision-Language Models
Abstract Vision-language models (VLMs) increasingly combine visual and textual information to perform complex tasks. However, conflicts between their internal knowledge and external visual input can lead to hallucinations and unreliable predictions. In this work, we investigate the mechanisms that VLMs use to resolve cross-modal conflicts by introducing WHOOPS-AHA!, a dataset of multimodal counterfactual queries that deliberately contradict internal commonsense knowledge. Through logit inspection, we identify a small set of attention heads that mediate this conflict. By intervening in these heads, we can steer the model towards its internal parametric knowledge or the visual information. Our results show that attention patterns on these heads effectively locate image regions that influence visual overrides, providing a more precise attribution compared to gradient-based methods. Authors Francesco Ortu, Zhijing Jin, Diego Doimo, Alberto Cazzaniga Journal Proceedings of the 64th Annual Meeting of the Association of Computational Linguistics Publication Date 01/07/2026 Consult the publication
Scalable, fast and accurate differential gene expression testing from millions of cells of multiple patients
Abstract Since the development of DNA microarrays and later RNA bulk sequencing, testing with statistically independent samples has been the standard method for detecting genes with different transcription patterns. Single-cell assays challenge these assumptions because individual cells are statistically dependent, and all proposed methodologies present mathematical limitations or computational bottlenecks that prevent a seamless integration of data from many cells and patients simultaneously. In this work, we solve this crucial limitation by introducing a Bayesian framework that retrieves the independence structure at the level of individual patients, separating differences across individuals from actual transcriptional differences. Leveraging multi-GPU and variational inference, our approach excels across different experimental designs and scales to analyse over 10 million cells. This framework enables single-cell differential expression analysis that can finally integrate datasets from large clinical cohorts, atlas projects, or drug-response screens with thousands of samples and millions of cells. Authors Giovanni Santacaterina, Niccolò Tosato, Salvatore Milite, Katsiaryna Davydzenka, Edoardo Insaghi, Guido Sanguinetti, Stefano Cozzini, Leonardo Egidi, Giulio Caravagna Journal Nature Communications 2026 Publication Date 19/06/2026 Consult the publication
Work Extraction via Backward Motion in Optimal Closed-Loop Stochastic Control
Abstract We experimentally realize finite-time feedback control in an overdamped colloidal system using real-time optical tweezers with in situ reinforcement learning (RL). By varying the protocol duration tf for displacing the optical trap between prescribed positions, the optimal strategies identified by RL reveal a crossover from deterministic dragging toward the target to feedback-assisted exploitation of thermal fluctuations, reducing and eventually overcoming the energetic cost. The resulting policies agree quantitatively with the exact optimal closed-loop solution. By extending the approach to spatially localized external forcing, we further show that RL can identify optimal feedback strategies in heterogeneous stochastic environments where direct analytical control design is challenging. Authors Luis Frieder Reinalter, Emanuele Panizon, Lokesh Muruga, Clemens Bechinger Journal Preprint Publication Date 16/06/2026 Consult the publication
Tumour evolution as ground truth for cancer whole-genome sequencing
Abstract Cancer genomes are shaped by evolutionary processes that couple mutagenesis, clonal selection, chromosomal instability, spatial growth and treatment response into structured genomic patterns, yet current benchmarking strategies largely ignore this evolutionary dependency. Here, we present SCOUT, a large-scale synthetic whole-genome sequencing resource of over 200 samples, designed for systematic benchmarking of tumour genomic analysis and evolutionary inference under controlled evolutionary ground truth. Unlike conventional task-specific simulations, SCOUT models tumour evolution as a latent generative process that simultaneously shapes mutations, copy-number alterations, variant allele frequencies, mutational signatures and clonal architectures. SCOUT recapitulates key features of solid and haematological malignancies, including driver mutations, chromosomal instability, intratumour heterogeneity, spatial sampling and treatment-associated evolutionary dynamics in tumour and matched-normal longitudinal and multi-region sequencing designs. Using SCOUT, we benchmarked widely used methods for somatic variant detection, copy-number analysis, mutational signature inference and tumour evolutionary reconstruction. Across analytical tasks, performance deteriorated in low-purity, highly subclonal and structurally complex tumours, while spatial sampling bias and hypermutation generated spurious evolutionary signals that confounded tumour interpretation across multiple inference layers. Evolutionary simulations further distinguished lineage-restricted genetic bottlenecks from multi-lineage resistance dynamics associated with tumour plasticity. Tumour purity consistently exerted a stronger effect on inference accuracy than sequencing depth. Together, our results establish evolutionary ground truth as a prerequisite for reproducible benchmarking and biologically interpretable analysis of cancer whole-genome sequencing data. Authors Lucrezia Valeriani, Giorgia Gandolfi, Elena Buscaroli, Katsiaryna Davydzenka, Giovanni Santacatterina, Alice Antonello, Azad Sadr, Virginia Anna Gazziero, Salvatore Milite, Elena Rivaroli, Anna Kabanova, Guido Sanguinetti, Alessio Ansuini, Leonardo Egidi, Stefano Cozzini, Alberto Cazzaniga, Giovanni Tonon, Trevor Graham, Andrea Sottoriva, Riccardo Bergamin, Nicola Calonaci, Alberto Casagrande, Giulio Caravagna Journal Preprint Publication Date 11/06/2026 Consult the publication
Decoding the Grammar of Protein-Protein Interaction Interfaces with Multimodal Representations
Abstract Protein-protein interactions govern essential cellular processes, making the identification of interacting sites a central challenge in structural biology, with important implications for protein engineering and the development of targeted therapeutics. Existing prediction algorithms include sequence-based methods, which lack structural information, or structure-based approaches, which often struggle to effectively integrate evolutionary context. Here, we present ESM3-PPISites, a supervised model for residue-level classification of interfaces, leveraging the multimodal representations of the ESM3 Protein Language Model. To ensure a bias-free evaluation, a stringent redundancy filtering protocol is adopted, systematically eliminating latent homology between the training data and a curated benchmark set in both sequence and structural space. ESM3-PPISites achieves unprecedented accuracy, vastly outperforming current approaches. Our findings demonstrate that while ESM3 largest proprietary version yields the highest predictive power, targeted fine-tuning of its small open-weight counterpart significantly narrows the performance gap. We also show the practical impact of these predictions by integrating them as spatial restraints within the HADDOCK docking platform. When evaluated on an independent subset of 12 complexes from the Docking Benchmark v5, the prediction-guided pipeline strongly enhances the identification of near-native binding poses over blind docking, while reducing computational runtime by an order of magnitude. This framework establishes a scalable paradigm for high-throughput structural characterization of protein–protein interactions. Authors Yuri Gardinazzi, Edith Natalia Villegas Garcia, Sergio Senci, Davide Di Vora, Antonio Feltrin, Francesca Cuturello Journal Preprint Publication Date 09/06/2026 Consult the publication
Visual Instruction Tuning Aligns Modalities through Abstraction
Abstract Visual instruction tuning effectively adapts a pre-trained Large Language Model (LLM) to process image information alongside text. Yet, it remains unclear how visual features are embedded into the layer-wise hierarchy of abstractions of the LLM backbone. Across a diverse set of vision-language architectures, we show that instruction tuning primarily serves as a bridge, embedding visual features directly into the intermediate semantic layers of the LLM, bypassing the early layers devoted to unimodal processing. With probing analyses and causal interventions, we show that these intermediate layers are the semantic core of vision-language processing and play a critical role in the performance on a broad set of multimodal benchmarks. In addition, by comparing the geometry of semantically equivalent visual and textual representations, we find that fine-tuning extends and strengthens the existing abstraction phase, aligning visual features with pre-existing textual ones. Finally, we confirm the functional role of this localized alignment by restricting fine-tuning to intermediate layers alone: this strategy preserves the performance of full fine-tuning on vision-centric benchmarks while reducing training time. Our results suggest that multimodal integration is a localized phenomenon driven by the repurposing of the internal abstraction engine of the LLM. Authors Luis Palacios, Lorenzo Basile, Diego Doimo, Alberto Cazzaniga Journal Preprint Publication Date 02/06/2026
Zigzag Persistence of Neural Responses to Time-Varying Stimuli
Abstract We use topological data analysis to study neural population activity in the Sensorium 2023 dataset, which records responses from thousands of mouse visual cortex neurons to diverse video stimuli. For each video, we build frame-by-frame cubical complexes from neuronal activity and apply zigzag persistent homology to capture how topological structure evolves over time. These dynamics are summarized with persistence landscapes, providing a compact vectorized representation of temporal features. We focus on one-dimensional topological features—loops in the data—that reflect coordinated, cyclical patterns of neural co-activation. To test their informativeness, we compare repeated trials of different videos by clustering their resulting topological neural representations. Our results show that these topological descriptors reliably distinguish neural responses to distinct stimuli. This work highlights a connection between evolving neuronal activity and interpretable topological signatures, advancing the use of topological data analysis for uncovering neural coding in complex dynamical systems. Authors Yuri Gardinazzi, Alessio Ansuini, Eugenio Piasini, Fabio Anselmi, Matteo Biagetti Journal Proceedings of Machine Learning Research (PMLR) 2026 Publication Date 01/06/2026 Consult the publication
Zigzag Persistence of Large Language Models Representations
Abstract We analyze internal representations of large language models with zigzag persistent homology, treating depth as a discrete time axis for point clouds of last-token embeddings. At each layer we build a k-nearest-neighbors clique complex, connect adjacent layers via intersections, and summarize the resulting diagrams with effective persistence images. From these we derive two descriptors: Births’ Relative Frequency (at what rate new p-dimensional features appear) and Inter-Layer Persistence (how long they survive across depth). On the SST movie reviews dataset and three open-source models (Llama-3.1, OSS-20B, Phi-4), we consistently observe three evolving phases: early rapid changes, a middle regime of stable organization, and a final reorganization before output. Using the stability signal (inter-layer persistence) to guide where to remove contiguous blocks of layers, we find that pruning within high-persistence regions maintains 5-shot MMLU performance (with the same trend visible even for the more pruning-sensitive OSS-20B). This suggests that zigzag-based summaries capture meaningful, system-level dynamics and can inform lightweight pruning. Authors Yuri Gardinazzi, Karthik Viswanathan, Giada Panerai, Alessio Ansuini, Alberto Cazzaniga, Matteo Biagetti Journal Proceedings of Machine Learning Research (PMLR) 2025 Publication Date 01/06/2026 Consult the publication
Perceptual misalignment of texture representations in convolutional neural networks
Abstract Mathematical modeling of visual textures traces back to Julesz’s intuition that texture perception in humans is based on local correlations between image features. An influential approach for texture analysis and generation generalizes this notion to linear correlations between the nonlinear features computed by convolutional neural networks (CNNs), compiled into Gram matrices. Given that CNNs are often used as models for the visual system, it is natural to ask whether such “texture representations” spontaneously align with the textures’ perceptual content, and in particular whether those CNNs that are regarded as better models for the visual system also possess more human-like texture representations. Here we quantify the perceptual content captured by feature correlations computed for a diverse pool of CNNs, and we compare it to the models’ perceptual alignment with the mammalian visual system as measured by Brain-Score. Surprisingly, we find that there is no connection between conventional measures of CNN quality as a model of the visual system and its alignment with human texture perception. We conclude that texture perception involves mechanisms that are distinct from those that are commonly modeled using approaches based on CNNs trained on object recognition, possibly depending on the integration of contextual information. Authors Ludovica de Paolis, Fabio Anselmi, Alessio Ansuini, Eugenio Piasini Journal Preprint Publication Date 18/05/2026 Consult the publication
TopoFisher: Learning Topological Summary Statistics by Maximizing Fisher Information
Abstract Persistence diagrams provide stable, interpretable summaries of geometric and topological structure and are useful for simulation-based inference when low-order statistics miss key information. Yet persistence-based pipelines require hand-chosen filtrations, vectorizations, and compressors, typically without an objective tied to parameter uncertainty. We introduce \textbf{TopoFisher}, a differentiable persistent-homology pipeline that learns topological summaries by maximizing local Gaussian Fisher information. Using simulations near a fiducial parameter, TopoFisher optimizes trainable filtrations, diagram vectorizations, and compressors without posterior samples or supervised regression targets, while retaining stable topological inductive bias. We also give sufficient regularity conditions for the log-determinant Fisher loss to be locally Lipschitz in trainable parameters. Controlled experiments on noisy spirals and Gaussian random fields, where total Fisher information is known, show that TopoFisher recovers much of the available information and outperforms fixed topological vectorizations. Our main results are on weak gravitational lensing, a high-dimensional non-Gaussian cosmological field-inference problem. Learned topological summaries reach higher Fisher information than state-of-the-art cosmological summaries and approach an unconstrained Information Maximising Neural Network baseline with up to ∼80× fewer parameters. The learned filtrations also generalize better: under simulator shift from lognormal to LPT-based maps it retains most Fisher information, while the neural baseline drops, and in neural posterior estimation they give tighter constraints than the neural baseline, and of state-of-the-art cosmological summaries. These results support Fisher-based topological optimization as a robust, parameter-efficient front end for simulation-based inference. Authors Matteo Biagetti, Mathieu Carrière, Francesco Conti, Enrico Maria Ferrari, Sven Heydenreich, Karthik Viswanathan Journal Preprint Publication Date 08/05/2026 Consult the publication
Balancing Compute in LLM Inference: Model Selection, Quantization, and Test-Time Scaling
Abstract Deploying large language models efficiently requires balancing accuracy, throughput, energy consumption, and memory usage. We present an empirical study of three complementary optimization dimensions for LLM inference: model size selection, posttraining quantization, and test-time compute scaling. Using the Qwen3 model family (0.6B-32B parameters) on NVIDIA L40S GPUs, we evaluate six quantization methods across a combination of six benchmarks and measure energy consumption per output token. We then apply budget-forced reasoning to quantized and unquan-tized models on the GSM8K benchmark. Our results show that model size selection provides up to 8× variation in energy per token, while quantization method choice accounts for up to 1.7×. Combining quantization with test-time scaling, a 4B model with a 2000-token reasoning budget achieves 90% accuracy on GSM8K, outperforming a 14B model (82%) while using 71% less memory and 2.7× less energy. We further demonstrate an adaptive thinking approach that allocates inference compute proportionally to problem difficulty, reducing average energy consumption while maintaining accuracy. These results provide practical guidance for configuring LLM inference systems that balance model performance and resource efficiency. Authors Tomislav Subic, Stefano Cozzini Journal Proceedings of the Sixth European Workshop on Machine Learning and Systems (EuroMLSys) 2026 Publication Date 28/04/2026 Consult the publication
Optimal Control of a Mesoscopic Information Engine
Abstract We analytically solve the finite-time control problem of driving an overdamped particle via an optical trap under costly measurement. By formulating this mesoscopic information engine within the Partially Observable Markov Decision Process (POMDP) framework, we demonstrate that the underlying Linear-Quadratic-Gaussian (LQG) dynamics decouple the optimal measurement and driving protocols. We derive the optimal feedback control law for the trap placement, which recovers the discontinuous Schmiedl-Seifert driving protocol in the open-loop limit and extends it to any measurement scheduling. For a costly, binary (on/off) sensor, we evaluate the optimal measurement protocol and derive physical bounds on the maximum net gain that can be extracted from thermal fluctuations. We show the emergence of deadline-induced blindness, a phenomenon where all measurements cease as the deadline approaches regardless of their cost. Taking the infinite-horizon limit, we find the exact periodic measurement schedules for the steady state as a function of the measurement cost C and derive the macroscopic velocity envelopes beyond which viscous drag forces the engine into a net-dissipative regime. Finally, we generalize the results to a variable-precision sensor. Authors Emanuele Panizon Journal Accepted for publication in Physical Reviews E Publication Date 09/04/2026 Consult the publication