arXiv subjects

Gianluca Bonifazi

Publications and source records attributed to Gianluca Bonifazi.

8 recordsLinked to original sources

arXiv2026

A Matryoshka Hierarchical RAG for Efficient Multi-Hop Question Answering

Retrieval-Augmented Generation (RAG) systems for multi-hop Question Answering (QA) must balance retrieval quality with computational cost. This cost is incurred during indexing time, through the use of expensive Knowledge Graphs (KGs) or Large Language Models (LLMs) to generate summaries, or during querying, through iterative LLM-driven retrieval. To reduce it while maintaining retrieval quality, we present MatRAG, a hierarchical framework that combines RAG systems with Matryoshka Representation Learning (MRL). MatRAG addresses both kinds of cost by aligning the semantic hierarchy of a clustering structure with the nested structure of MRL. Specifically, it organizes the corpus of documents into a Directed Acyclic Graph (DAG) of clusters with progressively coarser granularity. Each level is indexed by a lower Matryoshka dimension. MatRAG pairs an iterative, top-down traversal of the DAG with an entity-driven mechanism that controls the hop budget and re-ranks candidates. We evaluated MatRAG on three standard multi-hop QA benchmarks against seven representative baselines. MatRAG outperforms its strongest competitors in terms of retrieval quality; furthermore, it reduces indexing costs by avoiding KG construction and LLM-based summarization, and lowers query-time costs through dimension-aware similarity.

cs.CL↗

arXiv2026

Latent Trajectory Discrimination for AI-Generated Text Detection

Most existing approaches to AI-Generated Text Detection (AIGTD) treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. However, this perspective overlooks the inherently dynamic nature of autoregressive generation, where content evolves progressively through the latent space. In this paper, we reformulate AIGTD as the problem of distinguishing between latent generation trajectories. Instead of relying on static representations, we model how textual representations evolve across the sequence. To this end, we propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation. GTCL then applies contrastive learning to these trajectories to learn geometric regularities associated with the autoregressive generation. Evaluations performed on three different benchmarks and several approaches show that GTCL outperforms detection baselines consistently, which implies that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains. These results suggest that modeling trajectory differences could improve detection and open up a dynamic direction that has been underexplored in previous AIGTD literature.

cs.CL↗

arXiv2026

Efficient Retrieval-Augmented Generation via Token Co-occurrence Graphs

Retrieval-Augmented Generation (RAG) mitigates hallucinations in Large Language Models (LLMs) by grounding the generation process on external knowledge. However, standard RAG approaches struggle with multi-hop reasoning. While recent graph-based RAG methods improve the retrieval of interconnected chunks, they often rely on computationally expensive and error-prone LLM-based extraction pipelines. To address these issues, we propose TIGRAG (Token-Induced GraphRAG), an efficient graph-augmented RAG framework based on a token co-occurrence Knowledge Graph. TIGRAG directly models topological relationships between tokens using sliding-window co-occurrence statistics, thus enabling scalable graph construction. During inference, it combines graph-based semantic expansion and neural reranking to retrieve interconnected evidence for multi-hop reasoning. Specifically, it introduces an iterative entity-driven retrieval strategy that progressively expands the query using bridging entities extracted from previously retrieved contexts. We evaluated TIGRAG on three widely adopted multi-hop Question Answering (QA) benchmarks. Experimental results demonstrated that our framework consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks, while substantially reducing indexing time, inference latency, and prompt footprint.

cs.CL↗

arXiv2024

A statistical model to identify excess mortality in Italy in the period 2011-2022

Excess mortality is defined as an increase in the number of deaths above what is expected based on historical trends, hereafter called baseline. In a previous paper, we introduced a statistical method that allows an unbiased and robust determination of the baseline to be used for the computation of excesses. A good determination of the baseline allows us to efficiently evaluate the excess of casualties that occurred in Italy in the last 12 years and in particular in the last 3 years due to the Coronavirus Disease 2019 (COVID-19) epidemic. To this extent, we have analyzed the data on mortality in Italy in the period January 1st 2011 to December 31th 2022, provided by the Italian National Institute of Statistics (ISTAT). The dataset contains information on deaths for all possible causes, without specific reference to any particular one. The data exhibit strikingly evident periodicity in the number of deaths with pronounced maxima in the winter and minima in the summer, repeating itself in amplitude along the whole twelve-year sample. Superimposed on this wave-like structure are often present excesses of casualties, most likely due to occasional causes of death such as the flu epidemics (in winter) and heat waves (in summer). The very accurate periodicity along the seasons (the "baseline"), allows us to determine with great accuracy and confidence the number of expected deaths for each day of the year in the absence of occasional contributions. Each of the latter can be modeled with an additional function that parameterizes the deviation from the baseline.

physics.med-ph↗

arXiv2021

Study on the effects of the restrictive measures for containment of the COVID-19 pandemic on the reproduction number $R_t$ in Italian regions

Since November 6$^{\mathrm{th}}$, 2020, Italian regions have been classified according to four levels, corresponding to specific risk scenarios, for which specific restrictive measures have been foreseen. By analyzing the time evolution of the reproduction number $R_t$, we estimate how much different restrictive measures affect $R_t$, and we quantify the combined effect of the diffusion of virus variants and the beginning of the vaccination campaign upon the $R_t$ trend. We also compute the time delay between implementation of restrictive measures and the resulting effects. Three different models to describe the effects of restrictive measures are discussed and the results are cross-checked with two different algorithms for the computation of $R_t$.

physics.soc-ph↗

arXiv2021

A simplified estimate of the Effective Reproduction Number $R_t$ using its relation with the doubling time and application to Italian COVID-19 data

A simplified method to compute $R_t$, the Effective Reproduction Number, is presented. The method relates the value of $R_t$ to the estimation of the doubling time performed with a local exponential fit. The condition $R_t = 1$ corresponds to a growth rate equal to zero or equivalently an infinite doubling time. Different assumptions on the probability distribution of the generation time are considered. A simple analytical solution is presented in case the generation time follows a gamma distribution.

physics.soc-ph↗

arXiv2021

A statistical analysis of death rates in Italy for the years 2015-2020 and a comparison with the casualties reported for the COVID-19 pandemic

We analyze the data about casualties in Italy in the period 01/01/2015 to 30/09/2020 released by the Italian National Institute of Statistics (ISTAT). The data exhibit a clear sinusoidal behavior, whose fit allows for a robust subtraction of the baseline trend of casualties in Italy, with a surplus of mortality in correspondence to the flu epidemics in winter and to the hottest periods in summer. While these peaks are symmetric in shape, the peak in coincidence with the COVID-19 pandemics is asymmetric and more pronounced. We fit the former with a Gaussian function and the latter with a Gompertz function, in order to quantify number of casualties, the duration and the position of all causes of excess deaths. The overall quality of the fit to the data turns out to be very good. We discuss the trend of casualties in Italy by different classes of ages and for the different genders. We finally compare the data-subtracted casualties as reported by ISTAT with those reported by the Italian Department for Civil Protection (DPC) relative to the deaths directly attributed to COVID-19, and we discuss the differences.

physics.soc-ph↗

arXiv2021

A study on the possible merits of using symptomatic cases to trace the development of the COVID-19 pandemic

In a recent work we introduced a novel method to compute the effective reproduction number $R_t$ and we applied it to describe the development of the COVID-19 outbreak in Italy. The study is based on the number of daily positive swabs as reported by the Italian Dipartimento di Protezione Civile. Recently, the Italian Istituto Superiore di Sanità made available the data relative of the symptomatic cases, where the reporting date is the date of beginning of symptoms instead of the date of the reporting of the positive swab. In this paper we will discuss merits and drawbacks of this data, quantitatively comparing the quality of the pandemic indicators computed with the two samples.

physics.soc-ph↗