Papers by Fabrizio Silvestri

A Survey on Multimodal Disinformation Detection (2022.coling-1)

Copied to clipboard

Firoj Alam, Stefano Cresci, Tanmoy Chakraborty, Fabrizio Silvestri, Dimiter Dimitrov, Giovanni Da San Martino, Shaden Shaar, Hamed Firooz, Preslav Nakov

Multimodality

Challenge:	Recent years have witnessed the proliferation of offensive content online such as fake news, propaganda, misinformation, and disinformation.
Approach:	They propose to tackle online multimodal offensive content using different modalities and combinations thereof.
Outcome:	The proposed approach combines factuality and harmfulness in a framework that can be used for multiple modalities and combinations of modality.

Do RAG Systems Really Suffer From Positional Bias? (2025.emnlp-main)

Copied to clipboard

Florin Cuconasu, Simone Filice, Guy Horowitz, Yoelle Maarek, Fabrizio Silvestri

Responsible & Trustworthy NLP

Challenge:	Retrieval Augmented Generation (RAG) improves the factual accuracy of LLMs on knowledgeintensive tasks by including in the prompt passages retrieved from an external corpus.
Approach:	They propose to use a retrieval algorithm to add passages from an external corpus to the LLM prompt to improve the factual accuracy of LLMs.
Outcome:	The proposed approach improves the factual accuracy of LLMs on knowledgeintensive tasks by including in the prompt passages retrieved from an external corpus.

How Decoding Strategies Affect the Verifiability of Generated Text (2020.findings-emnlp)

Copied to clipboard

Luca Massarelli, Fabio Petroni, Aleksandra Piktus, Myle Ott, Tim Rocktäschel, Vassilis Plachouras, Fabrizio Silvestri, Sebastian Riedel

Language Models

Challenge:	Recent advances in pre-trained language models have generated text of an increasingly high quality.
Approach:	They propose a decoding strategy that produces less repetitive and more verifiable text.
Outcome:	The proposed method produces less repetitive and more verifiable text than previously used decoding strategies.

Redefining Retrieval Evaluation in the Era of LLMs (2026.eacl-long)

Copied to clipboard

Giovanni Trappolini, Florin Cuconasu, Simone Filice, Yoelle Maarek, Fabrizio Silvestri

Information Retrieval

Challenge:	Traditional IR metrics assume that humans examine documents sequentially with diminishing attention to lower ranks.
Approach:	They propose a utility-based annotation schema that quantifies positive contribution of relevant passages and negative impact of distracting ones.
Outcome:	The proposed metric improves correlation with the end-to-end answer accuracy by up to 36% compared to traditional metrics.

Detecting Propaganda Techniques in Memes (2021.acl-long)

Copied to clipboard

Dimitar Dimitrov, Bishr Bin Ali, Shaden Shaar, Firoj Alam, Fabrizio Silvestri, Hamed Firooz, Preslav Nakov, Giovanni Da San Martino

Multimodality

Challenge:	Propaganda can be defined as a form of communication that aims to influence opinions or the actions of people towards a specific goal.
Approach:	They propose to detect the type of propaganda techniques used in memes by annotating them with 22 techniques.
Outcome:	The proposed model identifies 22 propaganda techniques in memes, which can appear in text, image or both .

Statistical Foundations of DIME: Risk Estimation for Practical Index Selection (2026.eacl-short)

Copied to clipboard

Giulio D'Erasmo, Cesare Campagnano, Antonio Mallia, Pierpaolo Brutti, Nicola Tonellotto, Fabrizio Silvestri

Semantic Text Processing

Challenge:	High-dimensional dense embeddings are noisy or redundant, causing performance degradation and causing errors.
Approach:	They propose a method that scores each dimension by fusing the embeddings into a query-dependent matrix.
Outcome:	The proposed method improves retrieval effectiveness and reduces embedding size by an average 50% of across different models and datasets at inference time.

Beyond Position: the emergence of wavelet-like properties in Transformers (2025.acl-long)

Copied to clipboard

Valeria Ruscio, Umberto Nanni, Fabrizio Silvestri

Language Models

Challenge:	Despite its widespread adoption, theoretical limitations in positional encodings are resolved by developing emergent, wavelet-like processing strategies.
Approach:	They propose to use Rotary Position Embeddings to develop emergent, wavelet-like properties that compensate for the positional encoding’s theoretical limitations.
Outcome:	The attention heads evolve to implement multi-resolution processing analogous to wavelet transforms.

Database reasoning over text (2021.acl-long)

Copied to clipboard

James Thorne, Majid Yazdani, Marzieh Saeidi, Fabrizio Silvestri, Sebastian Riedel, Alon Halevy

Language Models

Challenge:	Existing models cannot handle database queries such as “List/Count all female athletes who were born in 20th century”.
Approach:	They propose a modular architecture to answer database-style queries over multiple spans from text and aggregate them at scale.
Outcome:	The proposed architecture scales to databases containing thousands of facts whereas current models are limited by how many facts can be encoded.

Attention Sinks in Diffusion Language Models (2026.findings-acl)

Copied to clipboard

Maximo Eduardo Rulli, Simone Petruzzi, Edoardo Michielon, Fabrizio Silvestri, Simone Scardapane, Alessio Devoto

Language Models

Challenge:	Masked Diffusion Language Models (DLMs) employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance.
Approach:	They conduct an empirical analysis of DLM attention patterns focusing on the attention sinking phenomenon . they find that DLMs also exhibit attention sinks, but with distinct characteristics .
Outcome:	The proposed models employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance.

DanteLLM: Let’s Push Italian LLM Research Forward! (2024.lrec-main)

Copied to clipboard

Andrea Bacciu, Cesare Campagnano, Giovanni Trappolini, Fabrizio Silvestri

Challenge:	Existing models for large language processing in the English language are limited in resources and evaluation tools for non-English languages.
Approach:	They propose a benchmark and an open LLM Leaderboard to evaluate LLMs’ performance in Italian and propose 'DanteLLM' it is the most performant LLM in the world, with improvements of up to 6 points .
Outcome:	The proposed model outperforms existing models in Italian and offers a blueprint for the development and evaluation of LLMs in other languages.

Misspelling Oblivious Word Embeddings (N19-1)

Copied to clipboard

Aleksandra Piktus, Necati Bora Edizel, Piotr Bojanowski, Edouard Grave, Rui Ferreira, Fabrizio Silvestri

Representation Learning

Challenge:	Existing word embeddings have limited applicability to malformed texts . misspellings are frequent and embeddable for words that have not been observed at training time .
Approach:	They propose a method to learn word embeddings that are resilient to misspellings . they use FastText with subwords to train embeddables on a new dataset .
Outcome:	The proposed method is tested on a publicly available dataset.

The Mechanics of Interference: Defusing Distractors in RAG via Sparse Autoencoder Interventions (2026.findings-acl)

Copied to clipboard

Christian Giannetti, Giovanni Trappolini, Nicola Tonellotto, Fabrizio Silvestri, Pietro Lio

Language Models

Challenge:	Large language models exhibit a critical vulnerability to distractor interference when processing retrieval-augmented contexts.
Approach:	They propose a mechanistic framework that corrects this failure mode through targeted interventions in the model’s latent space.
Outcome:	The proposed framework achieves recovery rates of up to 94% on distractor-vulnerable samples on Gemma-2 and Llama-3 model families across three QA benchmarks.