Publications | Samuele Punzo

Is my model perplexed for the right reason? Contrasting LLMs' Benchmark Behavior with Token-Level Perplexity

Tue, 31 Mar 2026 00:00:00 +0000

This paper introduces a simple, principled interpretability framework based on token-level perplexity to test whether LLMs rely on linguistically relevant cues when solving benchmarks.

Key contributions:

A hypothesis-driven analysis method using perplexity distributions over minimal sentence pairs
No reliance on unstable feature-attribution techniques
Experiments on controlled linguistic benchmarks with several open-weight LLMs
Evidence that linguistically important tokens influence but never fully explain model behavior, revealing reliance on non-linguistic heuristics

Machine Learning for analysis of Multiple Sclerosis cross-tissue bulk and single-cell transcriptomics data

Thu, 05 Mar 2026 00:00:00 +0000

This paper presents an integrative approach combining multi-omics data (Microarray and scRNA-seq) with explainable machine learning methods for the discovery of pathway-level signatures in Multiple Sclerosis.

Key contributions:

Cross-tissue integration of bulk and single-cell transcriptomics data
Explainable ML pipeline for pathway-level biomarker identification
Comparison of Microarray integration methods (ComBat, XPN)
Biological enrichment analysis for pathway discovery

A Machine Learning Pipeline for Multiple Sclerosis Biomarker Discovery: Comparing explainable AI and Traditional Statistical Approaches

Fri, 26 Sep 2025 00:00:00 +0000

This paper presents a comprehensive machine learning pipeline for biomarker discovery in Multiple Sclerosis, comparing explainable AI methods (SHAP, LIME) with traditional statistical approaches. The pipeline was applied to Microarray and scRNA-seq transcriptomics data, evaluating different feature selection and interpretation strategies.

Key contributions:

Systematic comparison of XAI-based vs. statistical biomarker identification
Integration of multiple data modalities (Microarray, scRNA-seq)
Biological validation through enrichment analysis tools (DAVID, StringDB, Cytoscape)