My research investigates artificial intelligence through the conceptual and methodological perspective of complex systems science. I am interested in understanding how information, organization, hierarchy, dynamics, and emergent structures arise in artificial learning systems, and how these properties can help explain their behaviour, internal representations, and computational capabilities.
This research program develops from my background in Computational Intelligence, Pattern Recognition, machine learning, and complex systems, and extends these foundations toward natural language, Transformer architectures, Large Language Models, generative AI, and emerging agentic systems.
A central question guides this trajectory: how can increasingly complex artificial systems be understood through the organization of their representations, interactions, information flows, and dynamics across multiple scales?
My work combines theoretical investigation, computational modelling, and empirical analysis, drawing on information theory, nonlinear dynamics, statistical and multifractal methods, representation learning, explainable artificial intelligence, and complex systems modelling.
The long-term objective is to contribute to a science of artificial intelligence capable of relating observable capabilities to the structures and processes from which they emerge.
The program is articulated across five interconnected research directions, spanning the study of artificial representations and generative models, explainability and internal organization, adaptive and agentic systems, complex engineering applications, and the relationship between AI complexity, efficiency, and sustainability.
Language provides a particularly rich domain for investigating the organization of complex artificial systems. My research studies natural and machine-generated language as structured phenomena characterized by dependencies, regularities, and patterns distributed across multiple scales.
A substantial part of this work concerns Transformer-based architectures and Large Language Models. I investigate the representations they learn, the structures they generate, and the relationships between statistical learning, abstraction, hierarchical organization, and semantic behaviour.
Previous work has addressed the multifractal and structural properties of linguistic sequences and the differences between human- and machine-generated texts. Current research extends this perspective toward the internal organization of generative models and toward architectures in which representations may progressively move from token-level regularities to more abstract semantic and conceptual structures.
This direction also includes emerging paradigms such as Large Concept Models and related approaches to representation and abstraction.
Understanding a complex artificial system requires descriptions that connect its observable behaviour with the organization of the computational processes that produce it.
My research investigates how concepts from information theory, complex systems science, structural analysis, and representation learning can contribute to the interpretability of modern AI architectures. Particular attention is devoted to internal representations, information flow, hierarchical organization, correlations across scales, and dynamical regularities.
This perspective intersects with Explainable Artificial Intelligence and, increasingly, with mechanistic interpretability. I am interested in identifying intermediate descriptions between low-level model parameters and high-level behavioural explanations, with particular attention to structures that may reveal how information is progressively organized, transformed, and recombined within a model.
The broader aim is to develop forms of explanation that remain compatible with the distributed, high-dimensional, and strongly interconnected nature of contemporary artificial intelligence.
Artificial intelligence is progressively moving from isolated predictive models toward systems composed of interacting agents, models, memories, tools, external information sources, and computational environments.
I study these architectures through the framework of complex adaptive systems. Relevant questions concern coordination, interaction networks, distributed intelligence, collective dynamics, adaptation, and the emergence of capabilities at the level of the whole system.
Multi-agent and agentic AI provide a particularly interesting setting for studying relationships between local interactions and global behaviour. They also offer an opportunity to revisit classical concepts from complexity science, including emergence, self-organization, adaptation, distributed computation, and collective behaviour, within the context of contemporary artificial architectures.
A central objective is to understand when system-level capabilities can be traced to individual components and when they depend essentially on patterns of interaction, organization, and feedback among those components.
A long-standing part of my research concerns the development and application of Computational Intelligence and machine learning methods to complex engineering systems.
My work in this area includes Smart Grids, Renewable Energy Communities, energy management systems, forecasting, battery modelling, fault diagnosis, predictive maintenance, and intelligent decision support.
These domains provide technologically relevant applications while also offering natural experimental settings for studying uncertainty, adaptation, distributed decision-making, interpretability, and interactions between data-driven models and physical systems.
The underlying methodological perspective remains closely connected with the broader research program. Complex engineering infrastructures are characterized by multiple interacting components, heterogeneous sources of information, nonlinear behaviour, changing operating conditions, and decisions distributed across different temporal and spatial scales. They therefore provide a valuable context in which methods developed within Computational Intelligence and complex systems research can be tested against real-world constraints.
The increasing scale of contemporary artificial intelligence raises fundamental questions about the relationship between model complexity, computational resources, information processing, performance, and energy consumption.
Parameter count and computational cost provide useful engineering measures, yet they offer only a partial description of the effective organization of a model. My research interests include the development of richer measures capable of characterizing structural, informational, and dynamical aspects of artificial systems.
This perspective creates a direct connection between complexity science and the study of efficient artificial intelligence. Information-theoretic quantities, multiscale organization, representation structure, and dynamical properties may provide useful tools for investigating how computational resources are transformed into model capabilities.
A related objective concerns the sustainability of generative AI and the development of methodologies capable of jointly considering representational efficiency, computational complexity, energy consumption, and performance. This direction also connects my work on artificial intelligence with my research experience in energy systems and energy management.
Part of my research activity extends beyond academic investigation toward technology transfer and the development of AI-based solutions for real-world systems. I am co-founder of TensorLoops, a Sapienza University of Rome innovative startup focused on artificial intelligence, complex systems modelling, explainable AI, and advanced data-driven solutions, and of TensorLoops Energy, which develops AI and machine-learning technologies for forecasting, optimization, energy management, Smart Grids, and Renewable Energy Communities.
These initiatives provide an additional environment in which methodological research can interact with industrial requirements, operational constraints, and the technological challenges associated with the deployment of artificial intelligence in complex systems.
Human–AI Interaction, Cognition and Epistemology
The increasing linguistic, interactive, and agentic capabilities of artificial systems also raise questions that extend beyond their computational performance.
I investigate how humans interpret artificial behaviour and how interactions with generative systems affect the attribution of agency, intentionality, mind, meaning, and knowledge. This line of research lies at the intersection of artificial intelligence, cognitive science, epistemology, philosophy of technology, and semiotics.
A particular interest concerns the phenomenology of human–AI interaction. Generative systems produce signs whose linguistic organization can be experienced as evidence of an underlying cognitive presence, even when the ontological status of such a presence remains unresolved.
Within this framework, I introduced the concept of noosemia, defined as the experience of the sign as a sign of mind. The concept is intended to describe the interpretative phenomenon through which an artificial sign is experienced as carrying traces of mindedness, agency, or intentional organization.
Noosemia provides a conceptual starting point for investigating the changing relationship between artificial language, human interpretation, perceived agency, and the epistemic conditions under which humans increasingly interact with generative and agentic systems.
The broader ambition of this research program is to develop a unified perspective in which artificial intelligence can be studied as a family of complex informational systems.
Such a perspective requires connections between different levels of description, ranging from computational mechanisms and internal representations to system-level behaviour, interaction, emergence, and human interpretation.
Complexity science offers a conceptual framework through which these levels can be related. My research aims to explore this possibility by combining mathematical and computational analysis with machine learning, engineering applications, and theoretical investigation of the cognitive and epistemic consequences of increasingly sophisticated artificial systems.
Manifesto on Artificial Intelligence — Sapienza University of Rome (2026)
I co-authored the Manifesto sull’Intelligenza Artificiale (Manifesto on Artificial Intelligence) of the Department of Information Engineering, Electronics and Telecommunications (DIET), Sapienza University of Rome, together with Antonello Rizzi, Sergio Barbarossa, Fabrizio Frezza, Mauro Olivieri, and Massimo Panella.
Officially approved by the Department Council on October 21, 2026, and issued on November 19, 2026, the Manifesto provides a scientific and institutional framework for the responsible integration of emerging AI technologies into university education and research.
The document addresses the transformation of scientific practices and knowledge production, the evolution of teaching and assessment, intellectual autonomy, and the epistemological and societal implications of increasingly capable artificial systems.
Official publication — https://diet.web.uniroma1.it/it/manifesto-ai
Full document (PDF) — https://diet.web.uniroma1.it/sites/default/files/allegati/2026-11/Manifesto_AI_DIET.pdf
An English-language version is forthcoming.
De Santis, E., & Rizzi, A. (in press). Emergent multiscale structure in Transformer representation spaces during training. In G. Minati & D. Rollo (Eds.), Emergent systemic properties (Contemporary Systems Thinking). Springer. https://doi.org/10.1007/978-3-032-41552-3_8 (in press)
De Santis, E., Martino, A., Cervellini, M., & Rizzi, A. (2026). Diffusion-limited aggregation signatures for structural text analysis, a case study across human, LLM, code, and controlled synthetic texts. In 2026 IEEE Workshop on Complexity in Engineering (COMPENG). Institutional record
De Santis, E., Martino, A., & Rizzi, A. (2026). Beyond perplexity: A multi-faceted analysis of a novel densely connected transformer. Applied Sciences, 16(6), Article 2721. https://doi.org/10.3390/app16062721
De Santis, E., & Rizzi, A. (2025). Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human–generative AI interaction [Preprint]. arXiv. https://arxiv.org/abs/2508.02622
De Santis, E., Martino, A., Bruno, E., & Rizzi, A. (2025). 2025: A GPT odyssey. Deconstructing intelligence by gradual dissolution of a transformer. In 2025 International Joint Conference on Neural Networks (IJCNN). IEEE. https://doi.org/10.1109/IJCNN64981.2025.11227797
De Santis, E., Martino, A., Ronci, F., & Rizzi, A. (2025b). LSTM in recursive feedback loops: A study on textual evolution and complexity. In 2025 International Joint Conference on Neural Networks (IJCNN). IEEE. https://doi.org/10.1109/IJCNN64981.2025.11227357
De Santis, E., Martino, A., Ronci, F., & Rizzi, A. (2025a). From bag-of-words to transformers: A comparative study for text classification in healthcare discussions in social media. IEEE Transactions on Emerging Topics in Computational Intelligence, 9(1), 1063–1077. https://doi.org/10.1109/TETCI.2024.3423444
De Santis, E., Martino, A., & Rizzi, A. (2024). Human versus machine intelligence: Assessing natural language generation models through complex systems theory. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7), 4812–4829. https://doi.org/10.1109/TPAMI.2024.3358168
De Santis, E., De Santis, G., & Rizzi, A. (2023). Multifractal characterization of texts for pattern recognition: On the complexity of morphological structures in modern and ancient languages. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8), 10143–10160. https://doi.org/10.1109/TPAMI.2023.3245886
De Santis, E., & Rizzi, A. (2023). Prototype theory meets word embedding: A novel approach for text categorization via granular computing. Cognitive Computation, 15(3), 976–997. https://doi.org/10.1007/s12559-023-10132-9
De Santis, E., Martino, A., Ronci, F., & Rizzi, A. (2023a). A comparison of neural word embedding language models for classifying social media users in the healthcare context. In 2023 International Joint Conference on Neural Networks (IJCNN) (pp. 1–9). IEEE. https://doi.org/10.1109/IJCNN54540.2023.10191583
De Santis, E., Martino, A., Ronci, F., & Rizzi, A. (2023b). An unsupervised graph-based approach for detecting relevant topics: A case study on the Italian Twitter cohort during the Russia–Ukraine conflict. Information, 14(6), Article 330. https://doi.org/10.3390/info14060330
De Santis, E., Capillo, A., Ferrandino, E., Frattale Mascioli, F. M., & Rizzi, A. (2023). An information granulation approach through m-grams for text classification. In J. Garibaldi, C. Wagner, T. Bäck, H.-K. Lam, M. Cottrell, K. Madani, & K. Warwick (Eds.), Computational intelligence (Studies in Computational Intelligence, Vol. 1119, pp. 73–89). Springer. https://doi.org/10.1007/978-3-031-46221-4_4
Martino, A., De Santis, E., & Rizzi, A. (2020). An ecology-based index for text embedding and classification. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1–8). IEEE. https://doi.org/10.1109/IJCNN48605.2020.9207299
De Santis, E., Martino, A., & Rizzi, A. (2020). An infoveillance system for detecting and tracking relevant topics from Italian tweets during the COVID-19 event. IEEE Access, 8, 132527–132538. https://doi.org/10.1109/ACCESS.2020.3010033
Capillo, A., De Santis, E., Frattale Mascioli, F. M., & Rizzi, A. (2020). Mining m-grams by a granular computing approach for text classification. In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI) (Vol. 1, pp. 350–360). SCITEPRESS. https://doi.org/10.5220/0010109803500360
Pelagalli, F., Greco, F., & De Santis, E. (2017). Social emotional data analysis. The map of Europe. In A. Petrucci & R. Verde (Eds.), SIS 2017: Statistics and data science: New challenges, new generations (pp. 779–784). Firenze University Press. Full text
De Santis, E., De Santis, G., & Rizzi, A. (2025). Noosemia. Prometeo, (172), 16–24. https://iris.uniroma1.it/handle/11573/1757711
De Santis, E. (2023). Apocalissi digitali e alchimie artificiali. Il linguaggio nell'epoca della sua riproducibilità tecnica. Prometeo, (163), 32–41. https://iris.uniroma1.it/handle/11573/1690704
De Santis, E. (2021). Umanità, complessità e intelligenza artificiale. Un connubio perfetto. Aracne.