Elisabetta Rocchetti

Reverse-engineering internal circuits and representations in language models. Developed a diagnostic framework (probing, INLP, causal ablation, PWCCA) to show that instruction-following in LLMs operates via compositional skill coordination rather than a universal mechanism. Investigated how LLMs internally track and enforce word-count constraints using Direct Logit Attribution and model diffing. Reproduced and extended research on how refusal is mediated by a single direction in activation space, comparing difference-in-means and INLP-based methods for extracting and steering it.

Developing a novel methodology to represent Transformer-based LLMs as directed, weighted graphs (nodes = attention heads and MLPs, edges = causal influence via intervention-based ablation). Tracked evolution of graph-theoretic metrics across 143 training checkpoints of Pythia-14M on induction tasks. Explored equivalence classes in Transformer input spaces using Riemannian geometry and pullback metrics.

Active research on safety-relevant properties of LLMs: refusal mechanisms, behavioral gaps between base and instruction-tuned models, and implicit content moderation. Co-founding collaborator in the creation of the Italian AI Safety community. Completed the BlueDot Impact Technical AI Safety course and currently pursuing their Technical AI Safety Project Sprint.

Investigating how text-to-image diffusion models internally represent artistic concepts (content vs. style) using cross-attention heatmaps. Published at IEEE MLSP 2025.