Publications

Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood. In this work, we study massive activations: a small subset of hidden-state channels whose responses are consistently much larger than the rest. We show that, despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses. First, they are functionally critical: a controlled disruption probe that zeroes the massive channels causes a sharp collapse in generation quality, while disrupting an equally-sized set of low-statistic channels has marginal effect. Second, they are spatially organized: restricting image-stream tokens to massive channels and clustering them yields coherent partitions that closely align with the main subject and salient regions, exposing a structured spatial code hidden inside an apparently outlier-like subspace. Third, they are transferable: transporting massive activations from one prompt-conditioned trajectory into another, shifts the final image toward the source prompt while preserving substantial content from the target, producing localized semantic interpolation rather than unstructured pixel blending. We exploit this property in two use cases: text-conditioned and image-conditioned semantic transport, where massive activations transport enables prompt interpolation and subject-driven generation without any additional training. Together, these results recast massive activations not as activation anomalies, but as a sparse promptconditioned carrier subspace that organizes and controls semantic information in modern DiT models.

Tool learning with foundation models aims to endow AI systems with the ability to invoke external resources — such as APIs, computational utilities, and specialized models — to solve complex tasks beyond the reach of standalone language generation. While recent advances in Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) have expanded their reasoning and perception capabilities, existing tool-use methods are predominantly limited to text-only inputs and closed-world settings. Consequently, they struggle to interpret multimodal user instructions and cannot generalize to tools unseen during training. In this work, we introduce RaTA-Tool, a novel framework for open-world multimodal tool selection. Rather than learning direct mappings from user queries to fixed tool identifiers, our approach enables an MLLM to convert a multimodal query into a structured task description and subsequently retrieve the most appropriate tool by matching this representation against semantically rich, machine-readable tool descriptions. This retrieval-based formulation naturally supports extensibility to new tools without retraining. To further improve alignment between task descriptions and tool selection, we incorporate a preference-based optimization stage using Direct Preference Optimization (DPO). To support research in this setting, we also introduce the first dataset for open-world multimodal tool use, featuring standardized tool descriptions derived from Hugging Face model cards. Extensive experiments demonstrate that our approach significantly improves tool-selection performance, particularly in open-world, multimodal scenarios.

Inference-time scaling has emerged as an effective way to improve generative models at test time by using a verifier to score and select candidate outputs. A common choice is to employ Multimodal Large Language Models (MLLMs) as verifiers, which can improve performance but introduce substantial inference-time cost. Indeed, diffusion pipelines operate in an autoencoder latent space to reduce computation, yet MLLM verifiers still require decoding candidates to pixel space and re-encoding them into the visual embedding space, leading to redundant and costly operations. In this work, we propose Verifier on Hidden States (VHS), a verifier that operates directly on intermediate hidden representations of Diffusion Transformer (DiT) single-step generators. VHS analyzes generator features without decoding to pixel space, thereby reducing the per-candidate verification cost while improving or matching the performance of MLLM-based competitors. We show that, under tiny inference budgets with only a small number of candidates per prompt, VHS enables more efficient inference-time scaling reducing joint generation-and-verification time by 63.3%, compute FLOPs by 51% and VRAM usage by 14.5% with respect to a standard MLLM verifier, achieving a +2.7% improvement on GenEval at the same inference-time budget.

Classifying artworks presents a significant challenge due to the complex interplay of fine-grained details and abstract features that condition the style or genre of an artwork. This paper presents a systematic investigation of the effectiveness of supervised and self-supervised backbones as feature extractors for both artwork classification and retrieval, with a particular focus on paintings. We conduct an extensive experimental evaluation using the DINO family and CLIP models, assessing multiple classification strategies and feature representations. Our results demonstrate that employing a self-supervised backbone leads to consistent improvements in artwork classification performance. Moreover, our work provides insights into the applicability of classification and retrieval modules in real-world applications, such as virtual reality (VR) applications that support museum navigation.

The rapid growth of the solar photovoltaic industry underlines the importance of effective operation and maintenance strategies, particularly for large-scale systems. Aerial infrared thermography has become an essential tool for detecting anomalies in photovoltaic modules due to its cost-effectiveness and scalability. Continuous monitoring through advanced fault detection and classification methods can maintain optimal system performance and extend the life of PV modules. This study investigates the application of advanced artificial intelligence methods for fault detection and classification comparing the performance of GPT-4o, a multimodal large language model, and ResNet, a convolutional neural network renowned for image classification tasks. Our research evaluates the effectiveness of both models using infrared images, focusing on binary defect detection and multiclass classification. ResNet demonstrated advantages in terms of computational efficiency and ease of implementation. Conversely, GPT-4o offered superior adaptability and interpretability, effectively analysing multimodal data to identify and explain subtle anomalies in thermal imagery. However, its higher computational requirements limit its feasibility in resource-limited settings. The results highlight the complementary strengths of these models and provide valuable insights into their role in advancing automated fault diagnosis in photovoltaic systems.

Formula 1 is a highly competitive and ever-evolving sport, with teams constantly searching for ways to gain an edge over the competition. In order to meet this challenge, we propose a custom Genetic Algorithm that can simulate a race strategy given data from free practices and compute an optimal strategy for a specific circuit. The algorithm takes into account a variety of factors that can affect race performance, including weather conditions as well as tire choice, pit-stops, fuel weight, and tire wear. By simulating and computing multiple race strategies, the algorithm provides valuable insights and can help make informed strategic decisions, in order to optimize the performance on the track. The algorithm has been evaluated on both a video-game simulation and with real data on tire consumption provided by the tire manufacturer Pirelli. With the help of the race strategy engineers from Pirelli, we have been able to prove the real applicability of the proposed algorithm.