3 accepted papers
Mechanistic Interpretability (MI) aims to reverse-engineer model behaviors by identifying functional sub-networks. Yet, the scientific validity of these findings depends on their stability. In this work, we argue that circuit discovery is not a standalone task but a statistical estimation problem bu…
As AI systems are increasingly deployed in high-stakes applications, ensuring their interpretability is essential. Mechanistic Interpretability (MI) aims to reverse-engineer neural networks by extracting human-understandable algorithms embedded within their structures to explain their behavior. This…
The paper investigates methods of compressing part-of-speech models that are developed for an on-device grapheme-to-phoneme conversion module. The performance of part-of-speech models is analyzed under different compression regimes. The evaluation is done with respect to French, German and Italian d…