Federico Cinus

Marco Minici, Cristian Consonni, Federico Cinus et al.

Large Language Models (LLMs) increasingly act as gateways to web content, shaping how millions of users encounter online information. Unlike traditional search engines, whose retrieval and ranking mechanisms are well studied, the selection processes of web-connected LLMs add layers of opacity to how answers are generated. By determining which news outlets users see, these systems can influence public opinion, reinforce echo chambers, and pose risks to civic discourse and public trust. This work extends two decades of research in algorithmic auditing to examine how LLMs function as news engines. We present the first audit comparing three leading agents, GPT-4o-Mini, Claude-3.7-Sonnet, and Gemini-2.0-Flash, against Google News, asking: \textit{How do LLMs differ from traditional aggregators in the diversity, ideology, and reliability of the media they expose to users?} Across 24 global topics, we find that, compared to Google News, LLMs surface significantly fewer unique outlets and allocate attention more unevenly. In the same way, GPT-4o-Mini emphasizes more factual and right-leaning sources; Claude-3.7-Sonnet favors institutional and civil-society domains and slightly amplifies right-leaning exposure; and Gemini-2.0-Flash exhibits a modest left-leaning tilt without significant changes in factuality. These patterns remain robust under prompt variations and alternative reliability benchmarks. Together, our findings show that LLMs already enact \textit{agentic editorial policies}, curating information in ways that diverge from conventional aggregators. Understanding and governing their emerging editorial power will be critical for ensuring transparency, pluralism, and trust in digital information ecosystems.

Marco Minici, Federico Cinus, Corrado Monti et al.

Despite echo chambers in social media have been under considerable scrutiny, general models for their detection and analysis are missing. In this work, we aim to fill this gap by proposing a probabilistic generative model that explains social media footprints -- i.e., social network structure and propagations of information -- through a set of latent communities, characterized by a degree of echo-chamber behavior and by an opinion polarity. Specifically, echo chambers are modeled as communities that are permeable to pieces of information with similar ideological polarity, and impermeable to information of opposed leaning: this allows discriminating echo chambers from communities that lack a clear ideological alignment. To learn the model parameters we propose a scalable, stochastic adaptation of the Generalized Expectation Maximization algorithm, that optimizes the joint likelihood of observing social connections and information propagation. Experiments on synthetic data show that our algorithm is able to correctly reconstruct ground-truth latent communities with their degree of echo-chamber behavior and opinion polarity. Experiments on real-world data about polarized social and political debates, such as the Brexit referendum or the COVID-19 vaccine campaign, confirm the effectiveness of our proposal in detecting echo chambers. Finally, we show how our model can improve accuracy in auxiliary predictive tasks, such as stance detection and prediction of future propagations.

Federico Cinus, Aristides Gionis, Francesco Bonchi

Social media have great potential for enabling public discourse on important societal issues. However, adverse effects, such as polarization and echo chambers, greatly impact the benefits of social media and call for algorithms that mitigate these effects. In this paper, we propose a novel problem formulation aimed at slightly nudging users' social feeds in order to strike a balance between relevance and diversity, thus mitigating the emergence of polarization, without lowering the quality of the feed. Our approach is based on re-weighting the relative importance of the accounts that a user follows, so as to calibrate the frequency with which the content produced by various accounts is shown to the user. We analyze the convexity properties of the problem, demonstrating the non-matrix convexity of the objective function and the convexity of the feasible set. To efficiently address the problem, we develop a scalable algorithm based on projected gradient descent. We also prove that our problem statement is a proper generalization of the undirected-case problem so that our method can also be adopted for undirected social networks. As a baseline for comparison in the undirected case, we develop a semidefinite programming approach, which provides the optimal solution. Through extensive experiments on synthetic and real-world datasets, we validate the effectiveness of our approach, which outperforms non-trivial baselines, underscoring its ability to foster healthier and more cohesive online communities.

Erica Coppolillo, Federico Cinus, Marco Minici et al.

Large Language Models (LLMs) demonstrate significant persuasive capabilities in one-on-one interactions, but their influence within social networks, where interconnected users and complex opinion dynamics pose unique challenges, remains underexplored. This paper addresses the research question: \emph{Can LLMs generate meaningful content that maximizes user engagement on social networks?} To answer this, we propose a pipeline using reinforcement learning with simulated feedback, where the network's response to LLM-generated content (i.e., the reward) is simulated through a formal engagement model. This approach bypasses the temporal cost and complexity of live experiments, enabling an efficient feedback loop between the LLM and the network under study. It also allows to control over endogenous factors such as the LLM's position within the social network and the distribution of opinions on a given topic. Our approach is adaptive to the opinion distribution of the underlying network and agnostic to the specifics of the engagement model, which is embedded as a plug-and-play component. Such flexibility makes it suitable for more complex engagement tasks and interventions in computational social science. Using our framework, we analyze the performance of LLMs in generating social engagement under different conditions, showcasing their full potential in this task. The experimental code is publicly available at https://github.com/mminici/Engagement-Driven-Content-Generation.

Marco Minici, Federico Cinus, Francesco Bonchi et al.

Signed Graph Neural Networks (SGNNs) have recently gained attention as an effective tool for several learning tasks on signed networks, i.e., graphs where edges have an associated polarity. One of these tasks is to predict the polarity of the links for which this information is missing, starting from the network structure and the other available polarities. However, when the available polarities are few and potentially noisy, such a task becomes challenging. In this work, we devise a semi-supervised learning framework that builds around the novel concept of \emph{multiscale social balance} to improve the prediction of link polarities in settings characterized by limited data quantity and quality. Our model-agnostic approach can seamlessly integrate with any SGNN architecture, dynamically reweighting the importance of each data sample while making strategic use of the structural information from unlabeled edges combined with social balance theory. Empirical validation demonstrates that our approach outperforms established baseline models, effectively addressing the limitations imposed by noisy and sparse data. This result underlines the benefits of incorporating multiscale social balance into SGNNs, opening new avenues for robust and accurate predictions in signed network analysis.

Jehyeok Yeon, Federico Cinus, Yifan Wu et al.

Large language models (LLMs) face critical safety challenges, as they can be manipulated to generate harmful content through adversarial prompts and jailbreak attacks. Many defenses are typically either black-box guardrails that filter outputs, or internals-based methods that steer hidden activations by operationalizing safety as a single latent feature or dimension. While effective for simple concepts, this assumption is limiting, as recent evidence shows that abstract concepts such as refusal and temporality are distributed across multiple features rather than isolated in one. To address this limitation, we introduce Graph-Regularized Sparse Autoencoders (GSAEs), which extends SAEs with a Laplacian smoothness penalty on the neuron co-activation graph. Unlike standard SAEs that assign each concept to a single latent feature, GSAEs recover smooth, distributed safety representations as coherent patterns spanning multiple features. We empirically demonstrate that GSAE enables effective runtime safety steering, assembling features into a weighted set of safety-relevant directions and controlling them with a two-stage gating mechanism that activates interventions only when harmful prompts or continuations are detected during generation. This approach enforces refusals adaptively while preserving utility on benign queries. Across safety and QA benchmarks, GSAE steering achieves an average 82% selective refusal rate, substantially outperforming standard SAE steering (42%), while maintaining strong task accuracy (70% on TriviaQA, 65% on TruthfulQA, 74% on GSM8K). Robustness experiments further show generalization across LLaMA-3, Mistral, Qwen, and Phi families and resilience against jailbreak attacks (GCG, AutoDAN), consistently maintaining >= 90% refusal of harmful content.

Federico Cinus, Yuko Kuroki, Atsushi Miyauchi et al.

We study the problem of minimizing polarization and disagreement in the Friedkin-Johnsen opinion dynamics model under incomplete information. Unlike prior work that assumes a static setting with full knowledge of users' innate opinions, we address the more realistic online setting where innate opinions are unknown and must be learned through sequential observations. This novel setting, which naturally mirrors periodic interventions on social media platforms, is formulated as a regret minimization problem, establishing a key connection between algorithmic interventions on social media platforms and theory of multi-armed bandits. In our formulation, a learner observes only a scalar feedback of the overall polarization and disagreement after an intervention. For this novel bandit problem, we propose a two-stage algorithm based on low-rank matrix bandits. The algorithm first performs subspace estimation to identify an underlying low-dimensional structure, and then employs a linear bandit algorithm within the compact dimensional representation derived from the estimated subspace. We prove that our algorithm achieves an $ \widetilde{O}(\sqrt{T}) $ cumulative regret over any time horizon $T$. Empirical results validate that our algorithm significantly outperforms a linear bandit baseline in terms of both cumulative regret and running time.