Elisabetta Rocchetti

Alfio Ferrara, Sergio Picascia, Elisabetta Rocchetti

Text-to-image diffusion models have demonstrated remarkable capabilities in generating artistic content by learning from billions of images, including popular artworks. However, the fundamental question of how these models internally represent concepts, such as content and style in paintings, remains unexplored. Traditional computer vision assumes content and style are orthogonal, but diffusion models receive no explicit guidance about this distinction during training. In this work, we investigate how transformer-based text-to-image diffusion models encode content and style concepts when generating artworks. We leverage cross-attention heatmaps to attribute pixels in generated images to specific prompt tokens, enabling us to isolate image regions influenced by content-describing versus style-describing tokens. Our findings reveal that diffusion models demonstrate varying degrees of content-style separation depending on the specific artistic prompt and style requested. In many cases, content tokens primarily influence object-related regions while style tokens affect background and texture areas, suggesting an emergent understanding of the content-style distinction. These insights contribute to our understanding of how large-scale generative models internally represent complex artistic concepts without explicit supervision. We share the code and dataset, together with an exploratory tool for visualizing attention maps at https://github.com/umilISLab/artistic-prompt-interpretation.

Elisabetta Rocchetti, Alfio Ferrara

Instruction tuning is commonly assumed to endow language models with a domain-general ability to follow instructions, yet the underlying mechanism remains poorly understood. Does instruction-following rely on a universal mechanism or compositional skill deployment? We investigate this through diagnostic probing across nine diverse tasks in three instruction-tuned models. Our analysis provides converging evidence against a universal mechanism. First, general probes trained across all tasks consistently underperform task-specific specialists, indicating limited representational sharing. Second, cross-task transfer is weak and clustered by skill similarity. Third, causal ablation reveals sparse asymmetric dependencies rather than shared representations. Tasks also stratify by complexity across layers, with structural constraints emerging early and semantic tasks emerging late. Finally, temporal analysis shows constraint satisfaction operates as dynamic monitoring during generation rather than pre-generation planning. These findings indicate that instruction-following is better characterized as skillful coordination of diverse linguistic capabilities rather than deployment of a single abstract constraint-checking process.

Elisabetta Rocchetti, Alfio Ferrara

Adhering to explicit length constraints, such as generating text with a precise word count, remains a significant challenge for Large Language Models (LLMs). This study aims at investigating the differences between foundation models and their instruction-tuned counterparts, on length-controlled text generation in English and Italian. We analyze both performance and internal component contributions using Cumulative Weighted Attribution, a metric derived from Direct Logit Attribution. Our findings reveal that instruction-tuning substantially improves length control, primarily by specializing components in deeper model layers. Specifically, attention heads in later layers of IT models show increasingly positive contributions, particularly in English. In Italian, while attention contributions are more attenuated, final-layer MLPs exhibit a stronger positive role, suggesting a compensatory mechanism. These results indicate that instruction-tuning reconfigures later layers for task adherence, with component-level strategies potentially adapting to linguistic context.

Elisabetta Rocchetti

The process by which Large Language Models (LLMs) acquire complex capabilities during training remains a key open question in mechanistic interpretability. This project investigates whether these learning dynamics can be characterized through the lens of Complex Network Theory (CNT). I introduce a novel methodology to represent a Transformer-based LLM as a directed, weighted graph where nodes are the model's computational components (attention heads and MLPs) and edges represent causal influence, measured via an intervention-based ablation technique. By tracking the evolution of this component-graph across 143 training checkpoints of the Pythia-14M model on a canonical induction task, I analyze a suite of graph-theoretic metrics. The results reveal that the network's structure evolves through distinct phases of exploration, consolidation, and refinement. Specifically, I identify the emergence of a stable hierarchy of information spreader components and a dynamic set of information gatherer components, whose roles reconfigure at key learning junctures. This work demonstrates that a component-level network perspective offers a powerful macroscopic lens for visualizing and understanding the self-organizing principles that drive the formation of functional circuits in LLMs.

Alfio Ferrara, Sergio Picascia, Laura Pinnavaia et al.

Proprietary Large Language Models (LLMs) have shown tendencies toward politeness, formality, and implicit content moderation. While previous research has primarily focused on explicitly training models to moderate and detoxify sensitive content, there has been limited exploration of whether LLMs implicitly sanitize language without explicit instructions. This study empirically analyzes the implicit moderation behavior of GPT-4o-mini when paraphrasing sensitive content and evaluates the extent of sensitivity shifts. Our experiments indicate that GPT-4o-mini systematically moderates content toward less sensitive classes, with substantial reductions in derogatory and taboo language. Also, we evaluate the zero-shot capabilities of LLMs in classifying sentence sensitivity, comparing their performances against traditional methods.