Oscar Stanchi
Oscar Stanchi, MSc
Computer Engineer and PhD Candidate
I'm currently pursuing a PhD in Computer Science with a focus on Interpretability of DNNs, under joint supervision between Universitat Rovira i Virgili and Universidad Nacional de La Plata, with a scholarship from CONICET. I primarily work with PyTorch and Captum.
I teach Software Engineering I and II at the Facultad de Informática (UNLP). I also cover topics in Deep Learning, emphasizing its ethical and social implications in the Master's program in Data Mining & Knowledge Discovery, which is offered at the Universidad de Buenos Aires (UBA).
URV UNLP CONICET CONICET
Latest Publications
AI Debaters are More Persuasive when Arguing in Alignment with Their Own Beliefs

AI Debaters are More Persuasive when Arguing in Alignment with Their Own Beliefs
Conference paper

Poster @ NeurIPS 2025

This paper explores how LLMs perform in AI debates when their arguments align or conflict with their prior beliefs. By measuring models’ priors and comparing debate protocols, we find that models are more persuasive when defending positions consistent with their beliefs, offering insights into persuasion dynamics and bias in scalable oversight.

SignAttention: On the Interpretability of Transformer Models for Sign Language Translation

SignAttention: On the Interpretability of Transformer Models for Sign Language Translation
Conference paper

Spotlight @ NeurIPS 2024

This paper studies a Transformer-based model for translating Greek Sign Language to glosses and text. It finds that the model focuses on clusters of frames, with alignment weakening as glosses increase, and shifts from video frames to predicted tokens during translation. This work enhances understanding of SLT models and promotes more transparent translation systems.

VISTA: Vision Improvement via Split and Reconstruct Deep Neural Network for Fundus Image Quality Assessment

VISTA: Vision Improvement via Split and Reconstruct Deep Neural Network for Fundus Image Quality Assessment
Journal paper

This paper introduces a deep learning model for assessing fundus image quality, essential for diagnosing eye conditions like cataracts and diabetic retinopathy. The model preserves high resolution and includes an autoencoder for reconstruction and classification. Results on the EyeQ dataset show 90.66% accuracy, 88.43% precision, 89.05% recall, and an F1-score of 88.68%.

© 2025 Oscar Stanchi