Alessandro Suglia

State-of-the-art VLA models exhibit significant behavioural divergence under perturbation, even when Task Success Rate remains unchanged, revealing a critical blind spot in current robustness evaluations.

Some dialogue games leverage knowledge transfer more effectively than finetuning, with visuospatial tasks showing the highest transfer potential.

Pixel-based language models can now handle multiple languages and scripts, achieving strong multilingual performance and robustness without tokenization.

Hybrid Transformer-SSM architectures can match or exceed pure Transformers in data-efficient in-context retrieval, but only if the task doesn't require precise positional reasoning.