Research proposals 42th cycle

Local energy markets in citizen-centered energy communities

The diffusion of distributed (renewable) energy sources poses new challenges in the underlying energy infrastructure, e.g., distribution and transmission networks and/or within micro (private) electric grids. The optimal, efficient and safe management and dispatch of electricity flows among different actors (i.e., prosumers) is key to supporting the diffusion of the distributed energy sources paradigm. The goal of the project is to explore different corporate structures, billing and sharing mechanisms inside energy communities. For instance, the use of smart energy contracts based on Distributed Ledger Technology (blockchain) for energy management in local energy communities will be studied. A testbed comprising of physical hardware (e.g., smart meters) connected in the loop with a simulated energy community environment (e.g., a building or a cluster of buildings) exploiting different Renewable Energy Sources (RES) and energy storage technology will be developed and tested during the three-year program. Hence, the research will focus on the development of agents capable of describing:- the final customer/prosumer beliefs desires and intentions and opinions.- the local energy market where prosumers can trade their energy and or flexibility- the local system operator that has to provide the grid reliability

All the software entities will be coupled with external simulators of the grid and energy sources in a plug-and-play fashion. Hence, the overall framework has to be able to work in a co-simulation environment with the possibility of performing hardware in the loop. The final outcomes of this research will be an agent-based modelling tool that can be exploited for:- Planning the evolution of future smart multi-energy systems by taking into account the operational phase- Evaluating the effect of different policies and related customer satisfaction- Evaluating the diffusion of technologies and/or energy policies under different regulatory scenarios- Evaluating new business models for energy communities and aggregators

During the 1st year, the candidate will study state-of-the-art solutions of existing agent-based modelling tools in order to identify the best available solution for large-scale smart energy system simulation in distributed environments. Furthermore, the candidate will review the state of the art in prosumers/aggregators/market modelling in order to identify the challenges and identify possible innovations. Moreover, the candidate will focus on the review of possible corporate structures, billing and sharing mechanisms of energy communities. Finally, it will start the design of the overall platform starting with the requirements identification and definition.

During the 2nd year, the candidate will terminate the design phase and will start the implementation of the agent intelligence. Furthermore, it will start to integrate agents and simulators together in order to create the first beta version of the tool.

During 3rd year, the candidate will ultimate the overall platform and test it in different case studies and scenarios in order to show all the effects of the different corporate structures, billing and sharing mechanisms in energy communities.

Possible international scientific journals and conferences:

- IEEE Transaction Smart Grid,

- IEEE Transactions on Evolutionary Computation,

- IEEE Transactions on Control of Network Systems,

- Environmental modelling and Software,

- JASSS,- ACM e-Energy,

- IEEE EEEIC internatational conference

- IEEE SEST internatational conference

- IEEE Compsac internatational conference

Simulation and Modelling of V2X connectivity with traffic simulation

This research aims at developing novel simulation tools for smart cities/smart grid scenarios that exploit the Agent-Based Modelling (ABM) approach to evaluate novel strategies to manage the V2X connectivity with traffic simulation. The candidate will develop an ABM simulator that will provide a realistic and virtual city where different scenarios will be executed. The ABM should be based on real data, demand profiles and traffic patterns. Furthermore, the simulation framework should be flexible and extendable so that i) It can be improved with new data from the field; ii) it can be interfaced with other simulation layers (i.e. physical grid simulators, communication simulators); iii) It can interact with external tools executing real policies (such as energy aggregation). This simulator will be a useful tool to analyse how V2X connectivity and the associated services impact both social behaviours and traffic. It will also help the understanding of the impact of new actors and companies (e.g., sharing companies) in both the marketplace and the society, again by analysing the social behaviours and the traffic conditions. In a nutshell, ABM simulator will simulate both traffic variation and the possible advantages of V2X connectivity strategies in a smart grid context. This ABM simulator will be designed and developed to span different spatial-temporal resolutions. All the software entities will be coupled with external simulators of grid and energy sources in a plug-and-play fashion to be ready for being integrated with external simulators and platforms. This will enhance the resulting AMB framework unlocking also hardware in the loop features.

The outcomes of this research will be an agent-based modelling tool that can be exploited for:- Simulating V2X connectivity considering traffic conditions- Evaluating the effect of different policies and related customer satisfaction- Evaluating the diffusion and acceptance of demand flexibility strategies- Evaluating the new business model for future companies and services

During the 1st year, the candidate will study the state-of-the-art solution of existing agent-based modelling tools to identify the best available solution for large-scale traffic simulation in distributed environments. Furthermore, the candidate will review the state of the art of V2X connectivity to identify the challenges and identify possible innovations. Moreover, the candidate will focus on the review of Artificial Intelligence algorithms for simulating traffic conditions and variation for estimating EV flexibility and users' preferences. Finally, he/she will start the design of the overall ABM framework and algorithms starting with the requirements identification and definition.

During the 2nd year, the candidate will terminate the design phase and will start the implementation of the agents' intelligence and test the first version of the proposed solution.

During the 3rd year, the candidate will ultimate the overall ABM framework and AI algorithms and test it in different case studies and scenarios to assess the impact of V2X connection strategies and novel business models.

Possible international scientific journals and conferences:- IEEE Transaction Smart Grid,- IEEE Transactions on Evolutionary Computation,- IEEE Transactions on Control of Network Systems,- Environmental modelling and Software,- JASSS,- ACM e-Energy,- IEEE EEEIC international conference- IEEE SEST international conference- IEEE Compsac international conference

Robust AI systems for data-limited applications

https://eda.polito.it/, 

https://www.linkedin.com/company/edagroup-polito/

Artificial Intelligence applications for advanced manufacturing systems

https://eda.polito.it/, 

https://www.linkedin.com/company/edagroup-polito/

The main goal of this PhD program is the investigation, design and deployment of state-of-the-art Artificial Intelligence approaches in the context of the smart factory, with special regards with new generation manufacturing systems. These tasks include:
- quality assurance and inspection of manufactured product via heterogeneous sensors data (e.g., images from visible range or IR cameras, time-series, etc.)
- process monitoring and forecasting
- anomaly detection
- failure prediction and maintenance planning support
While the Artificial Intelligence technologies able to address such tasks may already exist and be successfully consolidated in other real-world applications, the specific domain of manufacturing systems poses severe challenges to the effective deployment of these techniques. Among the others:
-     the immaturity of the involved technologies
-     the complexity of the underlying physical/chemical processes
-     the lack of effective infrastructures for data collection, integration, and annotation
-     the necessity to handle heterogeneous and noisy data from different types of sensors/machines
-     the lack of annotated datasets for training supervised models
-     the lack of standardized quality measures and benchmarks

This PhD program seeks solutions to these challenges, with specific focus on new generation manufacturing systems involving complex processes. For example: Additive Manufacturing (AM) and semiconductor manufacturing (SM).
-     AM includes many innovative 3D printing processes, which are rapidly revolutionizing manufacturing in the direction of higher digitalization of the process and higher flexibility of production. AM involves a fully digitalized process from design to product finishing, and hence it is a perfect candidate for the deployment of Artificial Intelligence. Nonetheless, it is a very complex and still immature technology, with tremendous room for improvement in terms of production time and product defectiveness. Specific use-cases in this regard will stem from the supervisors' collaborations with the Inter-departmental Center for Additive Manufacturing in Politecnico di Torino, as well as with several major industrial partners such as Prima Additive, Stellantis, Avio Aero, etc.
-     SM is another highly complex process, entailing a wide array of subprocesses and diverse equipment. Driven by the Industry 4.0 revolution and European Chips Act, the semiconductor industry is investing heavily in the digitalization of its production chain. As a result of these investments, the chip production process has been equipped with multiple sensors that constantly monitor the evolution of each manufacturing phase, from oxidation to testing and packaging, thus collecting a tremendous amount of heterogeneous data. To fully unveil the potential and hidden knowledge of such data, Artificial Intelligence is widely acknowledged to have a fundamental role. Use-cases in this regard will stem from the supervisors' collaborations with important industrial players in this sector, such as STMicroelectronics.

The outline of the PhD program can be divided into 3 consecutive phases, one per each year of the program.
-     In the first year, the candidate will acquire the necessary background by attending PhD courses and surveying the relevant literature and will start experimenting state-of-the-art techniques on the available datasets, either from public sources or from past projects of the supervisors. A seminal conference publication is expected at the end of the year.
-     In the second year, the candidate will select and address some relevant use-cases, with real data from the industrial partners, and will seek solutions to the technological and computational challenges posed by the specific industrial application.
At the end of the second year, the candidate is expected to target at least a second conference paper in a well-reputed industry-oriented conference (e.g. ETFA), and possibly another publication in a Q1 journal of the Computer Science sector (e.g. IEEE Transactions on Industrial Informatics, Expert Systems with Applications, etc.).
-     In the third year, the candidate will consolidate the models and approaches that were investigated in the second year, and possibly integrate them into a standalone framework. The candidate will also finalize this work into at least another major journal publication, as well as into a PhD thesis to defend at the end of the program.

Secure TinyML on RISC-V Architectures 

Smart digital technologies for advanced and personalised health monitoring

Introduction: In an increasingly ageing society, the prevalence of health-related problems (e.g. cardiovascular complications) is rapidly increasing, as is the incidence of chronic and neurodegenerative disorders. In recent decades, the development of new digital technologies promises to revolutionise the management of diseases and health-related issues. In particular, digital solutions are being designed and optimised to facilitate early diagnosis, unconstrained continuous monitoring and evaluation of disease progression and treatment effectiveness. These new technologies include wearable sensors that can be worn on the body for long periods of time and enable minimally invasive, unconstrained, continuous and accurate real-time health monitoring. In addition, devices such as RGB cameras, infrared sensors and radar systems offer the advantage of non-contact monitoring of vital signs and human movement. Non-invasive data collection procedures ensure the ecological validity of the recorded data, promoting patient comfort and compliance and thus enabling long-term monitoring. At the same time, the huge amount of multimodal data generated by different sensors can pose challenges for data analysis, fusion, processing, synthesis, and interpretation. To overcome these challenges, it is necessary to implement advanced signal processing and machine learning algorithms tailored to detect imperceptible changes in health parameters, aid early diagnosis and provide robust predictions.

Research objectives: This project aims to provide a broad set of computer-based methods capable of handling large amounts of multimodal health-related data. Algorithms include tailor-made signal processing methods to improve signal quality and fuse data from different domains, as well as feature- and data-driven machine learning models that process data to provide a robust and comprehensive health assessment. The design, implementation and optimisation of prediction algorithms must consider the trade-off between performance, interpretability, and computational complexity. Indeed, algorithms capable of running in real time on autonomous, low-resource portable devices will be developed.

Outline of the research work plan: The research is inherently multidisciplinary and involves the application of computer methods to medical data. Data will be recorded in clinical settings under the supervision of medical personnel, as well as collected continuously in remote settings (e.g. at home). Both young, healthy and elderly subjects as well as subjects with chronic and neurological diseases will be involved in the data acquisition process. Various commercial sensors and prototypes providing effective and non-invasive health monitoring will be exploited. Simple to complex algorithms will be developed and validated, both on large public data sets and on proprietary self-collected data. General models will be implemented to aid early diagnosis of chronic diseases and related symptoms. Subject-specific processing pipelines will be designed to match the characteristics of each patient, thus providing patient-centered solutions. Finally, secure and privacy-preserving methods, such as data anonymisation and federated learning, will be exploited to ensure adequate training of models while reducing the risk associated with the transfer of sensitive medical data.

List of possible venues for publications: The research results will be presented at recognised international scientific conferences (e.g., IEEE International Conference on Biomedical and Health Informatics, IEEE Symposium on Computer-Based Medical Systems, ACM International Symposium on Wearable Computers, IEEE International Conference on Digital Health) and top-tier international journals (e.g., Nature Digital Medicine, Artificial Intelligence in Medicine, Computers in Biology and Medicine, IEEE Journal of Biomedical and Health Informatics) at the intersection of computer science, engineering and medicine.

Cooperations: Research studies will be conducted and published in close cooperation with clinical professionals (e.g. neurologists, cardiologists, diabetologists, family doctors) and patient associations, to ensure that the methods developed meet current clinical gaps and the real needs of patients. In addition, national and international cooperation with academic research units (e.g. electronics, neuroscience, computer science) and specialised laboratories will provide a comprehensive set of knowledge and tools that will facilitate and promote high-quality research. Finally, collaborations with industrial partners will facilitate the design of effective, ready-to-use solutions for use in the real world.

• Multi-disciplinary experience at the intersection of medicine and computer science
• Strong motivation, ambition, curiosity, and creativity
• Programming in Python and/or Matlab
• Fundamentals of signal and image processing
• Fundamentals of machine learning

Secure and Green Digital Networks

Performance Optimization of ML-Based Compressed Video Communications on Packet Networks

In recent years, machine learning (ML) has been successfully employed also to improve video compression performance. Recent deep-learning based video compression algorithms achieve performance which is comparable if not better than state-of-the-art traditional standards such as, for instance, the MPEG family. However, algorithms are still relatively new and not adapted to specific communications scenarios where several impairments may happen, notably missing data due to packet losses.

We aim to improve the situation by investigating the baseline robustness of decoding algorithms, followed by proposals of efficient techniques to include coded data in packets in order to minimize quality impairments due to packet losses. Starting from Internet Media Group's traditional background on optimizing multimedia communications over packet networks, as well as ongoing work on media quality evaluation experiments, the candidate will first measure the quality that is obtained by running decoding algorithms, then using such knowledge to design schemes to make communications more robust to losses.

The workplan of the activities is detailed in the following. In the first year the PhD candidate will first familiarize with machine learning and AI-based video compression algorithms, using the available open source software which is released by researchers working in the field. In parallel, a framework will be created to efficiently conduct simulations and experiments where simulated video communications is subject to packet losses. The framework may include some modifications and adaptation to existing video decoders to make them work correctly in the simulations, since they might not have been originally designed to handle these conditions. These initial investigation and activities are expected to lead to conference publications.

In the second year, building on the developed simulators the theoretical knowledge already present in the research group, new efficient packetization strategies will be developed, simulated, and tested to demonstrate their performance and in particular their ability to make the communication as much robust as possible to packet losses. In this context, potential shortcomings of such algorithms will be systematically identified and the resulting performance in terms of video quality will be carefully measured by means of widespread objective video quality measures suitable for such purpose. These results are expected to yield one or more journal publications.
In the third year the activity will then be expanded to investigate if the encoders can be slightly modified to improve the resilience of the compressed data to missing information in the decoding phase. Such proposal will target journal publications.

Possible targets for research publications, well known to the proposer, include IEEE Transactions on Multimedia, Elsevier Signal Processing: Image Communication, ACM Transactions on Multimedia Computing Communications and Applications, Elsevier Multimedia Tools and Applications, various IEEE/ACM international conferences (IEEE ICME, IEEE MMSP, QoMEX, ACM MM, ACM MMSys).

The proposer is actively collaborating with the Video Quality Experts Group (VQEG), an international group of experts from academia and industry that aims to develop new standards in the context of video quality. In particular the tutor is co-chair of the JEG-Hybrid project which is very interested in the activity previously described.

AI for early and differential diagnosis of dementia, using facial expressions and voice.

• Multi-disciplinary experience at the intersection of medicine and computer science
• Strong motivation, ambition, curiosity, and creativity
• Programming in Python and/or Matlab
• Fundamentals of signal and image processing
• Fundamentals of machine learning

Exploiting Vision and Multimodal Foundation Models for Neurodegenerative Disease Assessment

• Strong background in machine learning and computer vision
• Programming skills in Python (PyTorch, TensorFlow or similar)
• Fundamentals of signal and image processing
• Interest in biomedical applications and interdisciplinary research
• Motivation to work at the intersection of AI and healthcare

Designing delay-trained spiking neural networks on heterogeneous cloud to edge neuromorphic systems

Self-Adaptive Security for Resilient Networks

Next-generation networks, such as virtualized networks and edge-to-cloud infrastructures, are reshaping the nature of digital environments. Their distributed, programmable, and dynamic nature opens new opportunities but also increases the surface and complexity of cyberattacks. As a consequence, traditional manual security (re)configuration strategies are no longer feasible, as they require excessive time to mitigate detected attacks and are prone to misconfigurations. In such highly dynamic environments, security is no longer only a matter of preventing attacks, but also of ensuring that the network can continue operating, adapt under stress, and recover quickly from disruptions. 

This PhD research proposal aims to study and design self-adaptive security mechanisms for resilient networks, so as to strengthen the ability of networked systems to withstand, absorb, and recover automatically from cyberattacks. The objective is to move beyond static defense configurations toward resilient security architectures that dynamically detect threats, assess their impact, and autonomously enforce appropriate countermeasures while preserving service continuity. Although several security automation tools exist today, they still operate mostly through predefined rules or they have limited reaction capabilities, leaving part of the decision-making process to human operators. As a result, they struggle to respond effectively to complex, multi-stage attacks that evolve over time and exploit the dynamic nature of programmable, distributed infrastructures, often compromising the system's overall resilience.

To address these challenges, the candidate will investigate how resilient networked infrastructures can be designed to adapt their security configuration in response to evolving threats and changing operational conditions. The candidate will explore techniques of various nature, including Artificial Intelligence-based algorithms or provably correct strategies rooted in constraint programming, to design approaches that can enable the continuous assessment of risk, the identification of exposed network components, and the selection of adaptive mitigation strategies. Rather than focusing only on stopping ongoing attacks, the proposed work will consider resilience as the ability of the network to sustain essential functions, degrade gracefully under stress, and recover efficiently after security incidents. In this perspective, the designed self-adaptive security mechanisms should continuously observe the system, reason on its current state, and modify protections accordingly. The ultimate goal is to define methodologies and mechanisms that allow next-generation networks not only to defend themselves more effectively, but also to preserve availability, maintain acceptable service levels, and progressively strengthen their security configuration over time.

The research activity will be organized in three phases:

Phase 1 (1st year): The candidate will study the literature on network security configuration automation, and will critically identify its main issues and limitations. Subsequently, with the tutor's guidance, the candidate will start identifying and defining new self-adaptive security mechanisms for resilient computer networks. Some preliminary results are expected to be published at this phase's end. During the first year, the candidate will also acquire the background necessary for the research. This will be done by attending courses and by personal study.

Phase 2 (2nd year): The candidate will consolidate the proposed approaches, fully implement them, and conduct experiments with them, e.g., to study their correctness, generality, and performance. In this year, particular emphasis will be given to the identified use cases, properly tuning the developed solutions to real scenarios. The results of this consolidated work will also be submitted for publication, aiming at least at a journal publication.

Phase 3 (3rd year): based on the results achieved in the previous phase, the proposed approach will be further refined to improve its scalability, performance, and applicability (e.g., different security properties and strategies will be considered), and the related dissemination activity will be completed.

The contributions produced by the proposed research can be published in conferences and journals belonging to the areas of cybersecurity (e.g. IEEE S&P, ACM CCS, NDSS, ESORICS, IFIP SEC, DSN, ACM Transactions on Information and System Security, or IEEE Transactions on Secure and Dependable Computing), and applications (e.g. IEEE Transactions on Industrial Informatics or IEEE Transactions on Vehicular Technology).

Secure Agent: Security and privacy in Agentic AI

Scenario and motivations: 
The rapid emergence of Agentic AI systems -- AI systems capable of autonomous planning, tool use, and interaction with external services -- creates a new and still poorly understood security and privacy landscape. These systems rely on emerging interaction protocols (e.g., MCP, A2A, OpenAI tool calling) and complex orchestration frameworks that significantly enlarge the attack surface compared to traditional AI systems. Potential threats include prompt injection, schema poisoning, context manipulation, tool capability escalation, data exfiltration, spoofing, and adversarial coordination between agents.

Research objectives:
The main objective of this PhD research is to study the security implications of Agentic AI ecosystems and develop a secure-by-design framework for the development and deployment of agent-based AI systems. Rather than retrofitting protections after deployment, the research will aim to embed security and privacy guarantees directly into the architecture of agentic systems. Specific goals include:- Systematically analysing the threat landscape of Agentic AI systems and emerging agent communication protocols.- Identifying and categorising vulnerabilities across the agent lifecycle (planning, tool selection, execution, memory, and inter-agent communication).- Designing security models, architectural patterns, and defensive mechanisms for agent orchestration platforms.- Developing automated techniques to detect and mitigate agent-specific attacks such as prompt injection, tool misuse, and context poisoning.- Proposing guidelines, frameworks, and evaluation methodologies for secure development of agentic AI infrastructures.

Research work plan: The research will be organised into several phases.- Phase 1 ? Ecosystem and threat analysis. The first phase will focus on analysing the architecture of existing Agentic AI frameworks, orchestration platforms, and communication protocols. A comprehensive threat model will be developed, identifying attack vectors, trust boundaries, and possible security failures in agent-tool and agent-agent interactions.- Phase 2 ? Vulnerability discovery and experimental evaluation. This phase will involve the design of experimental environments to systematically study vulnerabilities in agentic architectures. The candidate will analyse how different orchestration patterns and protocol designs affect the security posture of the system. Empirical studies will be conducted to evaluate the effectiveness of known and novel attacks.- Phase 3 ? Secure-by-design architectures. Based on the identified vulnerabilities, the research will propose architectural principles and mechanisms to secure agentic AI systems. Possible directions include capability-based access control for tools, protocol-level safeguards, secure context management, isolation mechanisms for agents, and verification of tool schemas and interaction policies.- Phase 4 ? Detection and mitigation mechanisms. The candidate will design automated techniques for the detection and mitigation of attacks in agent ecosystems. These may include runtime monitoring, anomaly detection for agent behaviour, defensive prompting strategies, secure tool wrappers, and validation layers for agent interactions.- Phase 5 ? Evaluation and guidelines. The proposed solutions will be evaluated through experimental implementations and benchmarking scenarios. The outcome of the research will include best practices, reference architectures, and security guidelines for developers and organisations building agentic AI systems.

The PhD candidate will be hosted for 18 months at EURECOM, working with Professor Antonioli, and 18 months at Polito, working with Professor Mellia and his group. Whatever the current location, the remote party will hold periodic (e.g., weekly) meetings with the hosting professor and the PhD student, so that the latter will effectively be co-supervised for the whole period of their scholarship.

References:- Mohamed Amine Ferrag, Norbert Tihanyi, Djallel Hamouda, Leandros Maglaras, Abderrahmane Lakas, Merouane Debbah, From prompt injections to protocol exploits: Threats in LLM-powered AI agents workflows, ICT Express, Volume 12, Issue 2, 2026,- Datta, S., Nahin, S. K., Chhabra, A., & Mohapatra, P. (2025). Agentic AI security: Threats, defences, evaluation, and open challenges. arXiv preprint arXiv:2510.23883.- Zhan, Q., Liang, Z., Ying, Z., & Kang, D. (2024, August). Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 10471-10506).

List of possible venues for publications:- Security venues: IEEE Security & Privacy, USENIX Security, ACM Computer and Communications Security (CCS), IEEE Transactions on Information Forensics and Security, ACM Transactions on Privacy and Security.- AI / Data Science venues: NeurIPS, ICLR, ICML, AAAI Conference, ACM KDD, ECML/PKDD- Systems / Web venues: ACM IMC, WWW/TheWebConf

• Strong programming skills in Python
• Knowledge of Web and communication protocols
• Knowledge of machine learning models and techniques
• Knowledge of cybersecurity
• Knowledge of AI and LLM

Physics-Informed Model-Based Reinforcement Learning for Multi-Scale Energy Systems

einforcement Learning (RL) is a promising approach for optimizing complex energy systems, but standard model-free methods often suffer from limited sample efficiency, poor generalization, and lack of physical consistency. This research focuses on model-based and physics-informed RL for Multi-Energy Systems (MES), where learning is guided by system dynamics and domain knowledge.

Modern MES integrate renewable energy sources, distributed generation, storage systems, and flexible loads such as HVAC units, resulting in highly dynamic, stochastic, and multi-scale environments. Traditional control approaches (e.g., PID) are often inadequate in handling these complexities, while purely data-driven RL may lead to unsafe or physically inconsistent solutions. This project addresses these limitations by combining model-based RL, Physics-Informed Neural Networks (PINNs), and hybrid control strategies.

Research Objectives: - Develop model-based RL algorithms that exploit both learned and physics-based models to improve data efficiency and control stability. - Integrate PINNs to encode physical laws (e.g., thermal dynamics, power flow constraints) into learning and system identification. - Design hybrid control architectures combining RL with classical approaches (e.g., MPC, PID) to ensure safety and robustness. - Extend the framework to Multi-Agent RL (MARL) for coordinated control in energy communities and district-level systems. - Assess performance in terms of energy efficiency, operational costs, robustness, and constraint satisfaction.

Research Plan:
Year 1: - Review literature on model-based RL, PINNs, and advanced control for energy systems. - Develop physics-informed models of key components (HVAC systems, batteries, flexible loads). - Design initial model-based RL algorithms for component-level optimisation. - Investigate safe RL techniques and constraint handling through physics-based regularizsation.

Year 2: - Extend the framework to household-level energy management, including demand response and flexibility modelling. - Develop multi-agent and hierarchical RL architectures for coordinated control.- Integrate PINNs into both system modelling and policy learning.- Explore hybrid RL-MPC/PID strategies to improve reliability and interpretability.

Year 3:- Validate the proposed methods at the district scale using co-simulation platforms. - Analyze robustness under uncertainty (renewable variability, demand fluctuations, market signals). - Collaborate with industrial partners for validation in realistic scenarios. - Benchmark against model-free RL and traditional control approaches.

Starting Point and Collaborations: 
The research will leverage existing co-simulation platforms of the Energy Center Lab at Politecnico di Torino, which support integrated modeling of buildings, storage systems, and distribution networks. Ongoing collaborations with academic and industrial partners will provide data, use cases, and validation opportunities.

Expected Outcomes: - Novel model-based and physics-informed RL methods tailored to MES. - Improved reliability, interpretability, and constraint compliance compared to purely data-driven approaches. - Scalable solutions for multi-scale and multi-agent energy management. - Contributions to hybrid AI-control methodologies for energy systems.

Publication Venues: - Journals: IEEE Transactions on Smart Grid; IEEE Transactions on Sustainable Computing; Energy and AI; Engineering Applications of Artificial Intelligence. - ?Conferences: NeurIPS; ICML; ACM e-Energy; IEEE INDIN; IEEE SEST.

OS-Driven Management of Aging Errors for Sustainable Embedded Chips

This proposal aims to transform hardware aging from a reliability threat into a system?level opportunity for sustainable computing in embedded platforms. Starting from climate and sustainability goals, it focuses on extending the useful lifetime of integrated circuits by reusing aging or discarded chips, thereby reducing the environmental costs of semiconductor manufacturing. The work targets systems based on SRAM and modern embedded processors (e.g., RISC?V), and integrates hardware characterization with firmware and operating?system techniques.?

Research objectives
The main objective is to develop methodologies for modeling, predicting, and managing hardware errors caused by chip aging or reuse in embedded systems. Specifically, the PhD will:?- Characterize complex spatial and temporal error patterns in SRAM and microarchitectural state under accelerated aging and stress conditions.?- Develop positional error models and predictive frameworks capturing the evolution of aging?related faults.?- Design OS? and firmware?level mechanisms that transparently: (i) compensate performance degradation (e.g., frequency drops) and (ii) exploit acceptable errors for approximate workloads, without changing application code.?- Validate these strategies on two application domains: error?resilient machine?learning workloads and highly reliable satellite control systems.?

Outline of the research work plan

Phase 1 - Experimental characterization (Months 1?12): The candidate will study the effects of aging on embedded platforms, focusing on RISC?V?based systems and SRAM memories. Accelerated-aging campaigns (thermal cycling, voltage variations, high?frequency operation) will be used to induce degradation and to collect functional and low?level data. This phase will produce datasets and a reusable experimental framework describing bit?flip rates, positional and temporal error patterns, and their evolution.

Phase 2 - Modeling and mitigation prototypes (Months 13?24): building on Phase 1, the candidate will derive mathematical and simulation models of aging?induced errors and their impact on timing, reliability, and energy. These models will guide the design of OS and firmware mechanisms such as frequency? and voltage?aware schedulers, adaptive parameter tuning, error?aware memory management, and approximate computing policies that classify and route tolerant workloads to degraded resources. Prototype implementations will be integrated into an embedded OS stack to assess feasibility and overhead.

Phase 3 - Validation and integration (Months 25?36): The models and OS?level strategies will be validated in realistic scenarios. On the ML side, representative inference workloads will be used to quantify trade?offs between error tolerance, energy savings, and performance. On the satellite?control side, representative control and monitoring tasks will evaluate how parameter tuning and selective approximation can preserve safety and functional reliability as the satellite ages. The final outcome will be an integrated hardware?firmware?software framework for aging?aware approximate computing, together with quantitative evaluations in terms of energy, Energy?Delay Product, Quality of Result, and system lifetime.

Industrial collaborations, projects, and context

The proposal contributes to the European agenda on sustainable digital systems (e.g., ?Beyond Green ICT: building a truly sustainable digital future?). The candidate will benefit from active industrial collaborations in aerospace and automotive, as well as with semiconductor companies, which offer realistic use cases, platforms, and potential datasets. These collaborations support validation on industrial?grade systems and improve the PhD graduate's employability in both academia and industry.

Possible venues for publications
The expected results can be published in leading conferences and journals in computer engineering, including: IEEE/ACM DATE, DAC, IEEE ETS, IEEE ITC, IEEE IOLTS, IEEE DFT, and DDECS for test, reliability, and approximate computing; embedded and systems venues such as IEEE ISVLSI and ACM TECS; and journals like IEEE Transactions on Computers, IEEE Transactions on Emerging Topics in Computing, Journal of Electronic Testing, and Proceedings of the IEEE for broader impact on dependable and sustainable computing.

Agentic AI for vRAN in Cooperative, Connected and Automated Mobility

Future Connected, Cooperative and Automated Mobility (CCAM) systems will rely on communication infrastructures that are not only high performing, but also adaptive, context-aware, and capable of autonomous decision making. In this context, virtualized Radio Access Networks (vRANs) offer a flexible platform where networking functions can be dynamically deployed, scaled, and optimized at the edge or in the cloud. This PhD proposal investigates how agentic AI can be embedded into vRAN architectures to support CCAM services, with particular attention to automated mobility, teleoperated driving, cooperative perception, maneuver coordination, and other latency- and reliability-sensitive applications.

The core idea is to move beyond static or purely reactive network control and design AI agents that can perceive network and service conditions, reason about system objectives, plan actions, and coordinate with other agents across vehicles, roadside infrastructure, and network nodes. Rather than treating the communication system as a passive transport layer, the research will explore how the vRAN can become an active participant in the orchestration of cooperative mobility services. In particular, the PhD will study how agentic AI can support selective information exchange, fusion and prioritization of data, adaptive allocation of compute resources, and coordinated decisions across distributed entities.

Research objectives- define a reference architecture for agentic AI in vRAN-enabled CCAM systems, identifying the role of distributed agents, their observation space, decision variables, coordination mechanisms, and interaction with RAN control and orchestration functions.- develop AI-driven strategies for selective and goal-oriented communication, so that only the information that is relevant for a given cooperative task, such as perception sharing or maneuver negotiation, is exchanged with the required timeliness and reliability.- design resource management and orchestration mechanisms that jointly optimize compute, and service-level performance in dynamic mobility scenarios.- evaluate the impact of agentic AI on key CCAM metrics, including latency, reliability, scalability, network efficiency, and quality of service for cooperative applications.

Outline of the research work plan
The research will start with a state-of-the-art analysis on vRAN, AI-native RAN control, multi-agent systems, and communication support for CCAM use cases. This phase will identify the main gaps in current approaches, especially in terms of distributed autonomy, explainability of decisions, and task-aware communication.

A second phase will focus on system modeling and architecture design. This will include the definition of representative use cases, such as cooperative perception and coordinated maneuvers, and the mapping of service requirements onto vRAN functions and agent behaviors.

A third phase will develop agentic AI mechanisms for distributed decision making in the vRAN. Candidate approaches may include multi-agent reinforcement learning, planning-based agents, hierarchical control, or hybrid AI methods combining learning and rule-based policies. Special attention will be given to selective task exchange, prioritization of control-relevant information, and interaction between network intelligence and application-level objectives.

A fourth phase will address implementation and performance evaluation through simulation and, where possible, emulation or small-scale prototyping. The research may leverage digital-twin environments and integrated mobility-network simulators to assess the proposed framework under realistic traffic, wireless, and service conditions.

The final phase will consolidate the results, compare the proposed solutions against conventional vRAN and non-agentic baselines, and derive design guidelines for future AI-native CCAM infrastructures.

List of possible venues for publications
Possible publication venues include leading journals such as IEEE Transactions on Wireless Communications, IEEE Transactions on Vehicular Technology, IEEE Transactions on Mobile Computing, IEEE Open Journal of Vehicular Technology, IEEE Network, and IEEE Vehicular Technology Magazine. Relevant conferences include IEEE VTC, IEEE ICC, IEEE Globecom, IEEE INFOCOM, ACM MobiCom, IEEE VNC, IEEE MeditCom, IEEE WCNC, and selected workshops on 6G, AI-native networks, and CCAM.

Digital Trust Data Infrastructures for Certified AI: blockchain-based architectures and governance

https://www.dauin.polito.it/

https://informatica.unito.it/do/home.pl

Evolutionary Artificial Intelligence

Evolutionary Computation (EC), a subfield of artificial intelligence, draws its inspiration from biological evolution. Since the 1990s, EC techniques have been applied in various practical domains, although sometimes under different names; in the past decades, their usage has been gradually reduced by the newer machine learning and deep learning techniques. However, existing sub-symbolic approaches have been proven inherently incapable of solving tasks demanding highly structured solutions, such as complex program synthesis problems or the ARC-AGI benchmarks. Evolutionary algorithms, with their unique characteristics, are experiencing a resurgence of interest (e.g., see "AlphaEvolve" by DeepMind). The objective of the proposal is to leverage EC to address the fundamental limitations of current AI approaches in challenging domains.

The long-term objective of this research is to overcome many limitations of sub-symbolic AI by developing novel frameworks based on Evolutionary Computation (EC) frameworks. In more details, the research objectives include:

The study of more powerful EC tools using alternative encoding for EC; more specifically, graph-based representations have led to novel applications of EC in circuit design, cryptography, image analysis, and other fields. The analysis of the "divergence of character", the most impairing single problem in the field of EC; the research activity would tackle "diversity promotion", that is either "increasing" or "preserving" diversity in an EC population, both from a practical and theoretical point of view. It will also include the related problems of defining and measuring diversity.

The task of automatically generating executable programs from high-level specifications, such as logical constraints, input-output examples, or natural language descriptions. The candidate will work on an open-source Python project, currently under active development.

The study of Core Knowledge theory, aiming at defining the "building blocks" to be used by the evolutionary algorithm to create programs, and bridge the gap between sub-symbolic and symbolic AI.

Target Publications

Journals with impact factors

- ASOC - Applied Soft Computing
- ECJ - Evolutionary Computation Journal
- GPem - Genetic Programming and Evolvable Machines
- Informatics and Computer Science Intelligent Systems Applications
- IS - Information Sciences
- NC - Natural Computing
- TCIAIG - IEEE Transactions on Computational Intelligence and AI in Games
- TEC - IEEE Transactions on Evolutionary Computation

Top conferences

- ACM GECCO - Genetic and Evolutionary Computation Conference
- IEEE CEC/WCCI - World Congress on Computational Intelligence
- PPSN - Parallel Problem Solving From Nature

Notes

The tutors regularly present tutorials on Diversity Preservation at top conferences in the field, such as GECCO, PPSN, and CEC. Additionally, they are involved in the organization of a workshops focused on graph-based representation for EA. Moreover, the research group is in contact with industries that actively consider exploiting evolutionary machine-learning for enhancing their biological models, for instance, KRD (Czech Republic), Teregroup (Italy), and BioVal Process (France).

The research group has also a long record of successful applications of evolutionary algorithms in several different domains. For instance, the on-going collaboration with STMicroelectronics on test and validation of programmable devices, does exploit evolutionary algorithms and would benefit from the research.

Orchestrating a Dynamically Federated Edge-to-Cloud Continuum

Research Objectives
==================================

Cloud-native technologies are increasingly deployed at the edge of the network, usually through tiny datacenters made by a few servers that maintain the main characteristics (powerful CPUs, high-speed network) of the well-known cloud datacenters. This project proposes to aggregate the above available hardware into a gigantic, distributed datacenter, potentially controlled by multiple administrative entities. This edge-to-cloud continuum would be populated by a set of highly dynamic, yet still secure and robust, ?private virtual spaces? spanning across multiple physical infrastructures, under the control of multiple federated actors.

The key objectives of the present research are the following:
?       Architectural Paradigms and Resource Management: Develop new paradigms, algorithms, and protocols for advertising, negotiating, and acquiring resources across different administrative domains.
?       Orchestration and Lifecycle Management: Create scalable algorithms and protocols for orchestrating the computing continuum, optimizing the placement of software services, and managing the lifecycle of virtual clusters.
?       Resilience and Evolution: Ensure the software infrastructure and virtual clusters can survive and evolve during network outages or planned disconnections.
?       Data Exchange Optimization: Optimize data exchange between virtual services when clusters span multiple sites.



Outline of the research work plan
==================================

Year 1: Foundation and Initial Development
-       Q1-Q2: Literature Review and Requirement Analysis
o  Conduct a comprehensive literature review on cloud-native technologies, edge computing, and distributed datacenters.
o  Identify key challenges and gaps in current technologies.
o  Define detailed requirements for the project.
-       Q3-Q4: Architectural Design and Initial Prototyping
-       Develop novel architectural paradigms for resource management across multiple administrative domains.
-       Design initial scalable algorithms and protocols for resource advertisement, negotiation, and acquisition.
-       Create a prototype for testing basic functionalities.

Year 2: Development and Testing
-       Q1-Q2: Advanced Algorithm Development
o  Develop scalable algorithms and protocols for orchestrating the computing continuum.
o  Focus on optimized placement for software services and applications.
o  Implement lifecycle management for virtual clusters.
-       Q3-Q4: Resilience and Data Exchange Optimization
o  Design and implement algorithms for infrastructure resilience during network outages or planned disconnections.
o  Develop scalable algorithms for optimizing data exchange between virtual services across multiple sites.
o  Conduct extensive testing and validation of the developed algorithms and protocols.

Year 3: Integration and Evaluation
-       Q1-Q2: Integration and System Testing
o  Integrate all developed components into a cohesive system.
o  Perform comprehensive system testing to ensure functionality and performance.
o  Optimize the system based on testing results.
-       Q3-Q4: Evaluation and Dissemination
o  Evaluate the system in real-world scenarios to assess its effectiveness and robustness.
o  Document findings and prepare research papers for publication.
o  Present results at conferences and workshops.
o  Plan for potential exploitation or further research extensions.


List of possible venues for publications
==================================

Top conferences:
- USENIX Symposium on Operating Systems Design and Implementation (OSDI)
- USENIX Symposium on Networked Systems Design and Implementation (NSDI)
- International Conference on Computer Communications (INFOCOM)
- ACM European Conference on Computer Systems (EuroSys)
- ACM Symposium on Principles of Distributed Computing (PODC)
- ACM Symposium on Operating Systems Principles (SOSP)

Journals:
- IEEE/ACM Transactions on Networking
- IEEE Transactions on Computers
- ACM Transactions on Computer Systems (TOCS)
- IEEE Transactions on Cloud Computing

Magazines:
- IEEE Computer
- IEEE Network

Industrial collaborations and involved projects
==================================

The research activity will be based on outputs of the IPCEI-CIS Project initiative (https://www.bmwk.de/Redaktion/EN/Artikel/Industry/ipcei-cis.html), in particular in collaboration with TIM S.p.A, and on the EURO-3C HE project, beginning on July 1st, 2026 (https://digital-strategy.ec.europa.eu/en/news/commission-announces-eu75-million-euro-3c-project-build-federated-telco-edge-cloud-infrastructure).

AI for Digital Design Automation: LLM-Based SoC and IP Integration and Optimization

Context and Motivation.

Modern System-on-Chip (SoC) platforms, particularly in automotive and industrial domains, are developed through complex integration flows based on standardized descriptions (e.g., IP-XACT) and custom toolchains. While these flows enable automatic HDL code generation, a significant part of the integration process remains manual and knowledge-intensive. Designers must interpret specifications, define connectivity among heterogeneous IPs, introduce glue logic, and manage a large number of device variants to reach an optimal implementation. Recent advances in Generative AI and Large Language Models (LLMs) have demonstrated strong capabilities in code generation and structured information extraction. However, their application to digital design automation, particularly to industrial SoC design flows, remains largely unexplored. This project investigates LLM-based methodologies to support and automate SoC assembly and optimization, leveraging both specification-driven generation with knowledge extracted from existing designs.
 

Research goals. 

- Specification-to-structure translation: Develop LLM-based techniques to derive structured design artifacts (e.g., IP-XACT descriptions) from natural language or semi-formal specifications.
- Integration Support and Automation: Design methods to assist SoC integration, including connectivity definition, inconsistency detection, and enforcement of integration rules (e.g., bus protocols, NoC configurations, glue logic patterns).
- Knowledge Extraction and Reuse: Leverage existing SoC designs and IP configurations to learn recurring integration patterns, enable their reuse, and support design recommendations.
- Design Space Exploration: Investigate techniques for rapid generation and evaluation of SoC configuration variants under high-level constraints such as power, performance, and area (PPA).
- AI-assisted IP Optimization: Investigate the integration of Generative AI techniques within High-Level Synthesis (HLS) flows to support architecture exploration and optimization of hardware IPs.

Outline of research work plan.

Year 1 ? Foundations and Benchmark Definition. Review the state of the art on LLMs for HDL code generation and design automation. Analyze industrial SoC integration flows, focusing on IP-XACT methodologies, RTL generation, and scripting environments. Define representative use cases and datasets to ground the research. Develop and evaluate initial prototypes for specification translation and integration template generation.

Year 2 ? Methodology Development and Integration. Develop structured LLM-based methodologies for integration support, including connectivity generation, rule checking, consistency validation, and optimization. Integrate knowledge extraction and reuse from existing designs. Progressively integrate the proposed techniques into industrial workflows and validate them on representative case studies.

Year 3 ? Extension and Validation. 
Extend the methodologies toward design space exploration, enabling efficient generation and evaluation of alternative SoC configurations under PPA constraints. Improve robustness, scalability, and usability of the developed tools. Explore extensions toward AI-assisted IP optimization via HLS. Perform comprehensive validation on industrial case studies to assess improvements in productivity, integration quality, and turnaround time.

Possible publication venues.

- IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- IEEE Transactions on Computers
- ACM Transactions on Design Automation of Electronic Systems
- ACM Transactions on Embedded Computing Systems
- Design Automation Conference (DAC)
- Design, Automation and Test in Europe (DATE)
- International Conference on Computer-Aided Design (ICCAD)
- Annual Conference on Machine Learning and Systems (MLSys)
- Annual Conference on Neural Information Processing Systems (NeurIPS)
- IEEE International Conference on LLM-Aided Design (LAD)
- ACM/IEEE International Symposium on Machine Learning for CAD (MLCAD)

Industrial and Project Context

The activity will be carried out in collaboration with STMicroelectronics, which will provide industrial use cases, including automotive SoCs, IP libraries, and IP-XACT and HLS-based design flows. The candidate will be co-supervised by an industrial tutor and will work on realistic integration scenarios, with the possibility of on-site periods at STMicroelectronics.

The research is framed within ongoing and upcoming industrial-oriented projects, including AIDA - AI-assisted Design Automation for Heterogeneous Electronic Systems (FESR 2026-2028) and ODE4EC - Open Design Environment for European Chips (Chips JU 2026-2029), addressing advanced methodologies for electronic design automation and next-generation semiconductor design platforms.

RISC-V Based Trusted Execution Environments for Software Defined Vehicles

The evolution of Software-Defined Vehicles (SDVs) is transforming automotive systems into highly connected, programmable platforms where critical functionalities?ranging from infotainment to advanced driver assistance?are deployed as software services. While this shift enables unprecedented flexibility and innovation, it also significantly enlarges the attack surface, raising critical concerns about security, safety, and trustworthiness.

In this context, Trusted Execution Environments (TEEs) are a key technological enabler for ensuring the secure, isolated execution of sensitive workloads. However, current TEE solutions are often proprietary, limited in flexibility, and not fully aligned with the stringent requirements of automotive systems. At the same time, the emergence of the RISC-V open instruction set architecture offers a unique opportunity to rethink hardware-based security from the ground up, enabling transparent, customizable, and verifiable trust anchors.

This research aims to bridge these gaps by designing and developing a novel hardware-based TEE tailored for SDVs, leveraging RISC-V's openness and extensibility. The central vision is to create a secure execution platform that guarantees integrity, isolation, and attestation of onboard services while remaining adaptable to evolving automotive requirements.

Research objectives

The main objective is to design a RISC-V-based TEE architecture that enhances security and trust in SDVs. This includes: (i) defining hardware mechanisms for strong isolation between trusted and untrusted domains, (ii) enabling secure boot and cryptographic attestation, (iii) supporting the trustworthy deployment of safety-critical and AI-driven services, and (iv) ensuring compatibility with automotive constraints such as real-time performance and functional safety.

Research work plan

The research will begin with an in-depth analysis of the state of the art in automotive cybersecurity and TEEs, alongside the definition of a comprehensive threat model tailored to SDVs. Building on this foundation, a novel RISC-V-based TEE architecture will be designed, exploiting hardware extensions to enforce isolation and security guarantees.

A prototype implementation will be developed using FPGA platforms or existing RISC-V cores, enabling practical validation of the proposed concepts. Core security features?such as secure boot, remote attestation, and runtime monitoring?will be implemented and integrated into the system.

The proposed solution will then be evaluated for performance, scalability, and resistance to realistic attack scenarios. Finally, the research will demonstrate the integration of the TEE within representative SDV use cases, such as over-the-air updates and secure execution of ADAS or AI services, highlighting its practical impact.

Possible venues for publications

The outcomes of this research are expected to be disseminated in leading venues in hardware security, embedded systems, and automotive technologies, including conferences such as DAC, DATE, HOST, CCS, USENIX Security, and NDSS, as well as journals like IEEE TDSC, IEEE TCAD, ACM TECS, and IEEE Security & Privacy. Additional dissemination may target automotive-focused venues such as IEEE IV and ITSC.

Neuromorphic Hardware Development

The growing demand for energy-efficient artificial intelligence is pushing beyond conventional von Neumann architectures toward brain-inspired computing paradigms. Neuromorphic systems, which emulate neural dynamics and event-driven processing, offer a promising path to drastically reduce power consumption while maintaining adaptive capabilities. However, current neuromorphic hardware often relies on simplified neuron models and largely synchronous designs, limiting their efficiency and biological realism.

This research proposes to explore a new generation of neuromorphic accelerators that combine asynchronous design principles with advanced neural models and adaptive mechanisms inspired by biological systems. The central hypothesis is that integrating event-driven computation with richer neural dynamics can unlock significant gains in both energy efficiency and computational expressiveness.

A key focus will be on asynchronous circuit implementations, which naturally align with the sparse and event-driven nature of neural activity. By eliminating global clocks and enabling computation only when needed, these architectures can significantly reduce energy consumption and improve scalability.

In parallel, the research will investigate liquid neural network architectures, which offer continuous-time dynamics and greater adaptability than traditional spiking or layered models. These architectures are particularly suitable for processing temporal and uncertain data, making them relevant for real-world applications.

Another innovative aspect of this work is the integration of homeostatic mechanisms directly into hardware. Inspired by biological neural systems, homeostasis allows networks to self-regulate their activity, improving stability, robustness, and long-term operation without external intervention.

Research objectives
- Design asynchronous neuromorphic accelerator architectures- Develop hardware support for liquid neural networks- Integrate homeostatic mechanisms for self-regulation- Optimize energy efficiency and scalability- Validate designs on FPGA/ASIC prototypes

Research work plan
- State of the art & modeling- Analyze neuromorphic hardware, asynchronous design techniques, and emerging neural models.- Architecture design- Define an event-driven, asynchronous accelerator architecture.- Neural model integration- Implement liquid neural networks and evaluate their hardware mapping.- Homeostasis mechanisms- Design and integrate adaptive regulation circuits.- Prototype implementation- Develop FPGA/ASIC prototypes.- Evaluation- Assess energy efficiency, latency, and robustness.

Possible venues for publications
- Conferences: ISCAS, DATE, DAC, ISSCC, NeurIPS, ICLR- Journals: IEEE TCAS, IEEE TCAD, Nature Electronics, Frontiers in Neuroscience
The research of this proposal is part of the WISE4.0 Horizon Europe project.

Neuromorphic Training & Continuous Learning

Despite significant advances in neuromorphic computing, most existing systems still rely on offline training and simplified learning schemes, limiting their adaptability and biological realism. In contrast, natural intelligence emerges from continuous, local, and diverse learning processes occurring directly within neural circuits.

This research aims to bring learning closer to the hardware by enabling fully embedded training and continuous adaptation in neuromorphic systems. The core vision is to design architectures that support local, online learning rules and multiple forms of synaptic plasticity, allowing systems to learn autonomously from streaming data without external supervision.

A central aspect of this work is the implementation of local training mechanisms directly in hardware, eliminating the need for global error propagation and reducing energy and communication overhead. This includes exploring spike-based and event-driven learning rules that can be efficiently mapped onto neuromorphic substrates.

In addition, the research will investigate the coexistence of multiple plasticity mechanisms?such as short-term, long-term, and structural plasticity?within the same hardware system. This multi-scale adaptability is key to achieving robust and flexible learning behaviors.

Moving beyond traditional layer-based architectures, the project will explore more general network topologies inspired by biological neural systems, with dynamic, recurrent, and heterogeneous connectivity. Such approaches can enable richer dynamics and improved generalization.

Research objectives
- Develop hardware-supported local learning rules- Implement multiple synaptic plasticity mechanisms- Enable continuous and online learning- Explore non-layered and recurrent architectures- Evaluate efficiency and adaptability

Research work plan
- State of the art & theory- Study neuromorphic learning, plasticity models, and hardware constraints.- Learning rule design- Define local, hardware-friendly training algorithms.- Hardware architecture- Design systems supporting in-memory and event-driven learning.- Plasticity integration- Implement multi-timescale plasticity mechanisms.- Prototype development- Validate on FPGA or neuromorphic platforms.- Evaluation- Assess learning performance, efficiency, and adaptability.

Possible venues for publications
- Conferences: NeurIPS, ICLR, ISCAS, DATE, DAC, ICONS- Journals: IEEE TNNLS, Nature Machine Intelligence, Frontiers in Neuroscience
The research of this proposal is part of the WISE4.0 Horizon Europe project.

Trustworthy and Safe AI under Hardware Faults

As artificial intelligence systems are increasingly deployed in safety-critical and harsh environments?such as automotive, aerospace, and edge devices?their reliability under hardware faults becomes a fundamental concern. Phenomena such as radiation-induced soft errors, aging, and transient faults can alter computations at the hardware level, potentially leading to silent and unpredictable failures in machine learning models.

Despite the growing importance of trustworthy AI, the interaction between hardware faults and model behavior remains poorly understood, particularly for large-scale neural networks. Most existing approaches focus either on hardware-level mitigation or on software robustness in isolation, leaving a critical gap in cross-layer understanding.

This research aims to bridge this gap by systematically studying how machine learning models respond to hardware-induced faults and by developing new methodologies to enhance their resilience. The central idea is to combine large-scale fault emulation with biologically inspired recovery mechanisms, drawing inspiration from the human brain's ability to tolerate damage and adapt through selective relearning.

A first key objective is to design scalable techniques to emulate hardware faults in large models, including bit flips, timing errors, and memory corruptions. These methods will enable controlled experimentation and detailed analysis of how faults propagate through neural computations and affect model outputs.

Building on this, the research will investigate the sensitivity of different model architectures, layers, and parameters to faults, identifying critical components and failure modes. This analysis will provide insights into the relationship between model structure and fault vulnerability.

The second major direction focuses on resilience. Inspired by neurobiological processes, the project will explore selective retraining strategies that adapt only the affected regions of the model after faults occur, reducing computational cost while restoring performance. Additional mechanisms for error tolerance, redundancy, and adaptive reconfiguration will also be considered.

Ultimately, the goal is to develop a new generation of AI systems that are inherently robust, capable of maintaining reliable operation even in the presence of hardware imperfections.

Research objectives
- Develop scalable fault emulation techniques for large ML models- Analyze fault propagation and impact on model behavior- Identify vulnerability patterns across architectures and tasks- Design resilience strategies based on selective retraining- Explore brain-inspired error tolerance and recovery mechanisms

Research work plan
- State of the art & fault modeling- Review hardware fault mechanisms and existing robustness approaches.- Fault emulation framework- Develop tools to inject faults in large-scale models (training/inference).- Sensitivity analysis- Study the impact of faults across layers, parameters, and architectures.- Resilience strategies- Design selective retraining and adaptive recovery methods.- Brain-inspired mechanisms- Explore redundancy, plasticity, and self-repair principles.- Evaluation- Validate on representative tasks and deployment scenarios.

Possible venues for publications
- Conferences: NeurIPS, ICML, ICLR, DSN, DATE, DAC- Journals: IEEE TDSC, IEEE TCAD, Nature Machine Intelligence, IEEE TPAMI

AI-driven Smart Systems for Sustainable Precision Agriculture

The research aims to design and develop intelligent agricultural systems that optimize resource use and support data-driven decision-making, thereby improving the productivity, sustainability, and resilience of farming systems. By integrating IoT sensing, computer vision, and advanced machine learning, the project targets three main objectives:
1. Accurate prediction of soil and crop conditions for efficient irrigation and input management;
2. Robust assessment of plant health using multispectral and imaging data;
3. Automation of field operations such as weed detection and selective intervention.

The project combines heterogeneous data sources, including in-field sensors (soil moisture, temperature, humidity), RGB and NIR imagery, and multispectral/hyperspectral data. Machine learning models will be developed along three main directions:
- Time-series modeling: Recurrent architectures (RNN, GRU, LSTM) will be designed to forecast soil moisture and crop water needs. Models will incorporate exogenous variables (weather, soil properties) and will be evaluated for generalization and robustness.
- Multispectral analysis: Feature selection techniques and statistical models will be used to identify informative wavelengths, reducing dimensionality while preserving classification accuracy for plant health monitoring.
- Computer vision for precision farming: Deep learning models (e.g., CNNs, transformers) will be trained on RGB?NIR data for weed detection, crop classification, and potentially fruit detection/harvesting support. Emphasis will be placed on lightweight architectures suitable for edge deployment.

The main phases of the work plan are:
1. Data Acquisition and System Design.
Design the sensing infrastructure and collect multimodal datasets from experimental fields or controlled environments. Activities include sensor calibration, image acquisition protocols, and dataset annotation. Initial exploratory analysis and preprocessing pipelines will be established.
2. Modeling and Algorithm Development
Develop predictive models for soil moisture using time-series techniques. Implement multispectral analysis pipelines with wavelength selection and classification models. Design and train deep learning models for weed detection and crop monitoring. Perform systematic validation (cross-validation, learning curves, permutation tests) to ensure robustness and avoid overfitting.
3. Integration and System Optimization
Integrate models into a unified decision support system. Optimize computational efficiency for real-time or near-real-time applications, including edge/embedded deployment. Evaluate trade-offs between model complexity and performance.
4. Field Validation and Impact Assessment
Validate the system in real or simulated agricultural scenarios. Assess improvements in water use efficiency, reduction of chemical inputs, and operational benefits. Compare against baseline practices.

Additional activities include:
- Development of explainable AI methods to improve model interpretability for agronomic experts;
- Exploration of data fusion strategies combining temporal, spectral, and spatial information;
- Collaboration with agronomists and stakeholders for domain validation;
- Dissemination through scientific publications and open datasets/tools.

The project will deliver scalable AI-based tools for precision agriculture, including predictive irrigation models, efficient spectral analysis methods, and vision-based systems for targeted interventions. These outcomes will contribute to more sustainable and resource-efficient farming practices.
Expected venues for publications are:
- Transaction on AgriFood Electronics, IEEE
- Computers and Electronics in Agriculture, Elsevier
- Smart Agricultural Technology, Elsevier
- Information Processing in Agriculture, China Agricultural University
- Precision Agriculture, Springer

The activities will be carried out in collaboration with PIC4SeR, PoliTo Interdepartmental Centre for Service Robotics.

Integrating Certification-Oriented Security Assurance into RTL Verification of RISC-V CPUs

On 8 December 2025, the European Common Criteria (EUCC) certification scheme under the EU Cybersecurity Act was amended. The update revised the state-of-the-art evaluation documents to reflect recent developments in hardware security, evolving threats, and industry practices for integrated circuits (IC) such as microcontrollers.

Under this certification scheme, ICs may obtain certification at either the ?substantial? or the ?high? assurance level, depending on the required resistance to attack. At the ?substantial? assurance level, the EUCC scheme requires vulnerability analysis and penetration testing activities. In current industrial practices, a vulnerability analysis for certification is often run after architectural stabilization or near tape-out. Differently, physical penetration testing is typically performed post-silicon; it intensifies when silicon or FPGA prototypes are available.

How can vulnerability analyses and penetration-test methodologies be shifted earlier and integrated into RTL verification flows for RISC-V CPUs? This represents a major challenge in the security field.

The PhD thesis addresses these challenges by proposing a certification-oriented security framework embedded in the RTL verification phase of RISC-V IPs. The goal is to remove exploitable design weaknesses before the product enters formal certification, by leveraging Large Language Models (LLMs) to translate vulnerability hypotheses into RTL-level verification scenarios and derive formal security invariants.

The PhD thesis is structured as follows: - During the first year of the PhD, the project will develop a RISC-V?specific threat model and vulnerability taxonomy. The analysis will consider privilege modes (Machine, Supervisor, User), CSR access control, PMP and MMU configuration, trap and interrupt handling, and debug interface behavior. Assets and trust boundaries will be formally defined. The result will be a structured mapping between potential state-of-the-art attack classes required by the updated EU Cybersecurity ACT and CPU architectural elements.- During the second year of the PhD, we will translate vulnerability hypotheses into verification tests. We will investigate LLM-based techniques to automatically translate structured vulnerability hypotheses into RTL-level verification stimuli, reducing manual effort and improving coverage of security-relevant corner cases. A dedicated library will be developed to emulate realistic attack attempts at the RTL level. Examples include illegal CSR access, privilege escalation attempts, and improper debug entry. While physical fault injection attacks such as clock glitches and voltage and electromagnetic glitches are typically evaluated post-silicon, their logical effects can be approximated at RTL. In addition, security-oriented RTL mutation techniques will be explored during the qualification phase to evaluate the robustness of the verification framework and the completeness of security checks.- During the third year of the PhD, the project will define security invariants for RISC-V CPUs. These invariants will be encoded as formally verified properties and integrated into the RTL verification environment. LLMs will be explored to assist in the derivation of formal security invariants from high-level specifications. In conventional verification flows, assertions are primarily derived from ISA compliance and functional correctness (they ensure that the processor behaves as specified under normal operating conditions). In this project, security invariants are instead derived from the threat model. They are designed to prevent misuse, privilege escalation, and isolation bypass, not only functional errors. For example, for out-of-order CPUs, security invariants can also be formulated as software-hardware contracts for secure speculation. 
The project contributes to hardware security verification by introducing security-oriented libraries tailored to RISC-V CPU architectures, and explores the use of LLMs in order to reduce the effort for the verification and optimization of the implemented security mechanisms.

This PhD thesis is conducted in industrial collaboration with STMicroelectronics.

Generative AI for Intelligent Biofabrication in Osteochondral Regeneration

Osteochondral regeneration requires the coordination of multiple biological, mechanical, and process-related factors, making the design and optimisation of biofabrication protocols particularly challenging. Experimental data are often heterogeneous, limited in size, and difficult to integrate with computational models. This PhD project aims to develop interpretable generative AI methods to support the analysis and optimisation of biofabrication processes for osteochondral applications, with a specific focus on multimodal data integration and biologically grounded data augmentation.

The research objectives are: 
(1) to design generative models able to integrate heterogeneous in vitro and in silico data relevant to osteochondral biofabrication, while incorporating biological and biomechanical constraints; 
(2) to investigate the use of synthetic data generation and latent-space modelling to improve the robustness of predictive analyses and support exploration of selected process variables; and 
(3) to assess the usefulness of these models in guiding the optimisation of specific biofabrication conditions, with attention to interpretability, reliability, and consistency with experimental knowledge.

The work plan is structured in four phases. In Phase 1, the candidate will analyse the available datasets and relevant domain knowledge, identify the most informative modalities and variables, and define a coherent computational representation of the biofabrication process. In Phase 2, the candidate will develop and compare generative models for multimodal data fusion and synthetic data generation, evaluating biological plausibility and utility for downstream tasks. In Phase 3, the research will focus on interpretable modelling strategies to relate latent representations and generated samples to relevant process and outcome variables. In Phase 4, the candidate will validate the proposed methods on selected case studies in osteochondral biofabrication, assessing whether generative modelling can support more efficient design-space exploration and data-driven optimisation of experimental settings.

The expected outcome is a computational framework for interpretable generative modelling in osteochondral biofabrication, capable of improving the use of limited multimodal datasets and supporting data-driven protocol refinement. The project will provide methodological advances in AI for biofabrication while remaining closely connected to experimentally meaningful questions.

Possible publication venues include Biofabrication, Acta Biomaterialia, Computer Methods and Programs in Biomedicine, IEEE Journal of Biomedical and Health Informatics, Frontiers in Bioengineering and Biotechnology, and conferences such as EMBC, MICCAI workshops, ML4H, and AI for Science workshops.

Multi-modal Data Intelligence and Digital Twins for Liver In Vitro Screenings

High-throughput liver in vitro screenings generate heterogeneous datasets spanning imaging, omics, simulation outputs, and process measurements, yet these resources are often fragmented, difficult to reuse, and poorly connected to predictive modelling pipelines. This PhD will develop multi-modal data intelligence methodologies and digital-twin-oriented infrastructures to integrate, analyse, and govern such datasets, enabling semantic interoperability, transparent reuse, and predictive knowledge discovery.

The research objectives are: 
(1) to design scalable data flows and integration strategies for imaging, omics, simulation, and process data derived from healthy and cancerous liver models; 
(2) to apply interpretable artificial intelligence and statistical learning methods to identify biomarkers, latent patterns, and mechanistic signatures that can improve digital twin calibration and predictive performance; and 
(3) to implement FAIR-compliant standards, metadata models, ontologies, and visual analytics tools that guarantee data traceability, interoperability, and reproducibility throughout the digital twin lifecycle.

The work plan is structured in four phases. 
In Phase 1, the candidate will map the data ecosystem of liver in vitro screenings, define metadata requirements, and design interoperable schemas linking experimental, computational, and process-level information. 
In Phase 2, the candidate will implement multi-modal integration pipelines for imaging, omics, simulation, and process data, with a focus on scalable ingestion, harmonisation, and semantic enrichment. 
In Phase 3, the project will develop interpretable AI and statistical learning approaches for biomarker discovery, pattern extraction, and calibration support, identifying the data features that most strongly drive digital twin behaviour. 
In Phase 4, the candidate will implement FAIR-compliant governance and visualisation tools, validate the full pipeline on use cases involving healthy and cancerous liver models, and demonstrate how the resulting platform supports transparent, reusable, and predictive digital twin workflows.

Expected outcomes include an open multi-modal analytics platform for liver in vitro screenings, a knowledge layer linking datasets, models, metadata, and experimental outcomes, and a set of methods for trustworthy calibration and reuse of liver digital twins. The project will contribute both methodological advances in multi-modal biomedical data intelligence and practical tools for interoperable digital twin creation.

Possible publication venues include Bioinformatics, GigaScience, Scientific Data, Journal of Biomedical Informatics, Patterns, Computer Methods and Programs in Biomedicine, Frontiers in Bioengineering and Biotechnology, and conferences such as ISMB, IEEE BIBM, MICCAI, and KDD workshops on health and biomedicine.

Data-driven and Surrogate Modelling of In Vitro Tumor Vascularisation

Tumors are characterised by complex vascular and microenvironmental dynamics that are difficult to capture with mechanistic models alone, especially when multiscale simulations become too computationally demanding for extensive exploration, calibration, or integration with experimental pipelines. This PhD will develop explainable surrogate models and interoperable data-driven environments to compress complex tumor vascularisation dynamics into efficient predictive tools, accelerating simulation while preserving biological relevance and interpretability.

The research objectives are: 
(1) to perform network inference from transcriptomic and experimental datasets generated from vascularised tumor models, identifying regulatory and phenotypic patterns linked to vascular development, tumour progression, and microenvironmental adaptation; 
(2) to design hybrid surrogate models that combine mechanistic knowledge with machine learning, including physics-informed, systems-biology-informed, and explainable AI approaches, to emulate key components of multiscale simulations; and 
(3) to implement scalable model reduction and uncertainty quantification strategies so that surrogate predictions remain computationally efficient, robust, and trustworthy for downstream scientific use.

The work plan is organised in four phases. 
In Phase 1, the candidate will study available mechanistic models, curate transcriptomic and experimental datasets, and define interoperable representations of biological entities, states, and outputs. 
In Phase 2, the candidate will perform network inference and feature extraction to identify the dominant pathways, state transitions, and observables that should guide surrogate construction. 
In Phase 3, the project will focus on developing white-box/black-box hybrid surrogates, combining mechanistic constraints, machine learning, and explainability tools to produce interpretable predictive approximations of complex simulations. 
In Phase 4, the candidate will integrate uncertainty quantification and scalable reductions, validate the models against experimental and simulated data, and deploy the resulting tools in interoperable environments that connect in vitro tumor assays with multiscale in silico workflows.

Expected outcomes include accelerated and interpretable simulation pipelines for tumor vascularisation, interoperable environments for integrating experimental and computational data, and hybrid surrogate models that support hypothesis generation, simulation-based analysis, and digital experimentation across related applications.

Possible publication venues include PLOS Computational Biology, Bioinformatics, Briefings in Bioinformatics, npj Systems Biology and Applications, Computer Methods and Programs in Biomedicine, Journal of Theoretical Biology, Cancers, and conferences such as ISMB, RECOMB, MICCAI, and IEEE BIBM.

Interaction Design for Everyday Augmented Reality

World Grounding for Virtual Humans in Extended Reality

AI-Assisted XR Production Pipelines: Authoring Tools and Workflow Support for Technical Artists

Knowledge-grounded data generation for reliable and scalable AI agents

Research Objectives

The goal of this research is to investigate the ability of Large Language Models (LLMs) to handle domain-specific constraints on their outputs. This is particularly important when using LLMs to generate synthetic data, especially in scientific and industrial domains where real-world data is scarce, costly, or constrained by privacy and regulatory limitations.

A central premise of this work is that many real-world domains (e.g., engineering systems, legal reasoning, industrial logs) are strongly governed by structured knowledge, domain constraints, and theoretical principles. Traditional machine learning approaches often struggle in these contexts due to limited data availability and the difficulty of encoding prior knowledge directly into models. At the same time, recent advances in LLMs suggest that properly guiding these models can generate rich, structured, and semantically coherent data that may serve as a viable complement to real datasets.

Within this context, the research aims to explore whether and how synthetic data, generated within carefully designed pipelines and aligned with domain knowledge, can improve the performance, robustness, and generalization capabilities of deep learning models and agentic systems. Particular emphasis is placed on understanding the conditions under which such improvements occur, as well as the limitations and risks associated with the use of synthetic data, such as the crucial model collapse issue.

Main research objectives.

Methodology Development: developing methodologies for generating synthetic datasets and digital twins of real-world environments that align explicitly with domain-specific knowledge and constraints, incorporating structured representations like ontologies, rules, physical laws, or policy constraints.

Benchmark Design: designing controlled benchmarks based on synthetic data, aimed at systematically evaluating the capabilities of models and agents. These benchmarks will focus on key aspects such as reasoning, robustness to perturbation, and generalization across scenarios.

Agentic Systems Support: exploring how synthetic data can support models that explicitly incorporate domain knowledge, as well as multi-agent systems navigating constrained environments. Synthetic data will simulate complex interactions, such as agent dialogues, sequences of tool use, and structured reasoning processes.

Model Distillation: investigating the feasibility of using large LLMs as synthetic data generators to support smaller, more efficient models, via in-context learning and dedicated fine-tuning stages, while maintaining strict constraint compliance.

Outline

1st Year: the candidate will investigate the state of the art in LLMs, synthetic data generation, and constraint-aware AI, focusing on methods for incorporating domain knowledge (e.g., ontologies, rules, physical laws) into generative processes. The study will analyze controllable text generation, structured data synthesis, and digital twin construction, alongside techniques for evaluating factuality and adherence to constraints in LLM outputs. In parallel, relevant application domains, suitable datasets, and simulation environments will be identified.

2nd Year: the candidate will develop novel methodologies for constraint-aware synthetic data generation using LLMs, ensuring alignment with domain knowledge and the controllability of generated outputs. This includes designing hybrid pipelines combining LLMs with symbolic or rule-based components, as well as automatic validation and filtering stages to enforce constraints. The research will explore generating complex synthetic scenarios (e.g., agent interactions, tool usage sequences). Additionally, the candidate will investigate using large LLMs as teachers to support smaller models, analyzing performance, robustness, and constraint adherence.

3rd Year: the proposed methodologies will be extended to more complex, large-scale, multi-domain scenarios, and industrial use cases. The focus will be on optimizing synthetic data generation pipelines for scalability, efficiency, and reliability, and improving their integration with downstream models and agentic architectures. Extensive experimental evaluation will identify conditions where synthetic data provides measurable benefits, alongside its limitations and potential risks (e.g., bias propagation or constraint violations). Strategies for transferring knowledge from large LLMs to smaller, deployable models will be refined.

Target Publications
Conferences: ACL, KDD, NeurIPS, ICLR, AAAI, ECML-PKDD
Journals: IEEE TKDE, ACM TKDD, ACM TIST, JMLR, Information Sciences (Elsevier), Engineering Applications of Artificial Intelligence (Elsevier)

Knowledge-Informed Machine Learning for Data Science and Scientific AI

Research Objectives

The research aims to define new methodologies for integrating scientific and domain knowledge within advanced data science models and agentic AI architectures, with a focus on advancing Scientific AI. 
The goal is to propose innovative algorithms, agent designs, and hybrid model structures, explore their applications in various scientific and engineering domains, investigate the limitations of current approaches, and advance solutions based on the integration of data-driven methods and theory-informed reasoning.
The ultimate goal is to contribute to improving the capability of AI systems to understand and model complex systems by deeply integrating existing domain knowledge (e.g., physical laws, causal relationships, theoretical principles) with data-driven insights. This will drive towards more robust, interpretable, and generalizable models and intelligent agents capable of sophisticated reasoning and discovery, opening new perspectives for accelerating scientific breakthroughs and designing more efficient and reliable engineering solutions.

To this end, the main research objectives include:
?    Developing new strategies to embed scientific knowledge (e.g., physical laws, causal relationships) and structured domain knowledge into advanced data science models and agentic AI architectures, enabling them to reason from first principles.
?    Designing intelligent agents and multi-agent systems capable of representing, reasoning about, and acting upon complex entities and relationships within scientific domains, incorporating physical constraints and theoretical knowledge into their behavior and learning processes.
?    Addressing the challenges in current AI models, including agentic systems, to effectively manage and integrate complex, multi-scale scientific knowledge, thereby enhancing their reasoning capabilities and the expressive power of the derived solutions.

Outline
?    1st year. The candidate will explore the state-of-the-art techniques in Knowledge-Informed Machine Learning, Scientific AI, Theory-Guided Data Science, Agentic AI, and advanced data modeling. This will include a focus on how scientific knowledge can inform the design and learning processes of intelligent agents and sophisticated analytical models. Applications to physical phenomena, engineering simulations, social data, and other scientific domains will be analyzed. Specific datasets, simulation environments, and problems for experimentation will be identified.
?    2nd year. The candidate will define innovative solutions based on agentic principles and advanced data science models to overcome the limitations described in the research objectives, experimenting with the proposed techniques on the identified real-world or simulated problems. The development and experimental phase will be conducted on public, synthetic, and possibly real-world datasets/environments, with particular attention to validation against known scientific principles and theoretical understanding. New challenges and limitations are expected to be identified in this phase.
?    3rd year. During the third year, the candidate will extend the research by broadening the experimental evaluation to more complex phenomena capable of better leveraging the domain knowledge provided by the developed Agentic AI systems and advanced data science models. The candidate will perform optimizations on the designed algorithms and agent architectures, establishing the limitations of the developed solutions and possible improvements in new application fields, with a focus on generating scientifically valid, interpretable, and actionable results.

Target publications

IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TOIS (Trans. on Information Systems)
ACM TIST (Trans. on Intelligent Systems and Technology)
IEEE TPAMI (Trans. on Pattern Analysis and Machine Intelligence)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
ACM Transactions on Spatial Algorithms and Systems (TSAS)

Agentic AI for the Cloud Continuum

Few-shot imitation learning for real world manipulation 

Deploying anthropomorphic robotic manipulators in real-world environments still requires learning methods that can generalize beyond a single robot, setup, or task. In this project, the candidate will study advanced imitation learning approaches that acquire manipulation policies directly from visual demonstrations, using 2D or 2.5D observations and, when beneficial, structured intermediate representations such as scene graphs or other graph-based encodings. The objective is to reduce the amount of supervision and robot interaction needed to learn complex behaviours, while improving robustness to changes in viewpoint, lighting, background, object arrangement, task configuration and possibly also robot embodiment.

Research objectives:- design visual imitation learning pipelines that learn manipulation policies from a limited number of demonstrations.- investigate how 2D, 2.5D, and graph-based representations can improve policy learning, temporal reasoning, and action generalization.- develop neural architectures that transfer across robot morphologies (i.e from human examples or from other robots), simulators, and manipulation tasks with little fine-tuning data.- study few-shot adaptation strategies for rapidly specializing a pretrained policy to new domains, embodiments, or task variants.- evaluate precision, robustness, and sample efficiency on manipulation tasks involving anthropomorphic robotic arms.

The research work plan will be articulated in four main phases. 
First, the candidate will review the literature on imitation learning, behavior cloning, inverse models, video-based policy learning, visuomotor transformers, and graph neural networks for structured scene understanding. In parallel, benchmark tasks, datasets, and experimental protocols will be defined for manipulation from visual demonstrations.
Second, the candidate will develop perception modules able to extract informative representations from demonstrations, including RGB, depth or pseudo-depth cues, keypoints, object-centric features, and graph-based relational descriptions of the scene. Particular attention will be given to temporal consistency and to representations that remain stable under changes in camera conditions and scene appearance.
Third, the candidate will design and train policy architectures for learning manipulation skills from few demonstrations. These models may include attention-based visuomotor backbones, latent action representations, graph-conditioned policy networks, and adaptation modules for transfer across robots and tasks. The work will also investigate few-shot fine-tuning, domain adaptation, and sim-to-real or robot-to-robot transfer.
Fourth, the developed methods will be validated in simulation and, on anthropomorphic robotic manipulators in laboratory settings. Evaluation will consider task success, motion precision, generalization to unseen objects or layouts, adaptation speed, and the number of demonstrations required to reach target performance. The expected outcome is a transferable and data-efficient framework for visual imitation learning that lowers training costs and accelerates deployment in new applications.

We plan to publish outcomes of this research to premium conferences like CoRL, RSS, ICRA, IROS, CVPR, ICCV, and ECCV, as well as journals such as IEEE Robotics and Automation Letters, IEEE Transactions on Robotics, The International Journal of Robotics Research, IEEE Transactions on Pattern Analysis and Machine Intelligence, and the International Journal of Computer Vision.

Measuring Digital Wellbeing Beyond Screen Time

The growing integration of digital technologies into everyday life has raised fundamental concerns about their impact on users' psychological wellbeing. Social media platforms, streaming services, and mobile applications increasingly rely on Attention-Capture Damaging Patterns (ACDPs)?such as infinite scroll, content autoplay, and algorithmic recommendations?designed to maximize engagement metrics rather than support healthy digital experiences. These design strategies compromise users' Sense of Agency (SoA), i.e., their perceived control over their own actions and outcomes during interaction.

Despite growing research on digital wellbeing, the HCI community still lacks robust methods for measuring the real-time impact of these patterns on users' cognitive and affective states. Current approaches rely on two paradigms: (1) screen-time metrics, which reduce digital experience to a single quantitative proxy, and (2) self-report questionnaires, which capture only post-hoc reflective judgments and are vulnerable to recall bias. Neither can adequately capture the in-the-moment experiential dimension of digital wellbeing?how users actually feel and behave while interacting with an interface.

Building on empirical evidence showing that small feed-level design changes can measurably influence both reflective and experiential dimensions of agency on platforms like TikTok [1], this PhD proposal aims to develop and validate new methods for measuring digital wellbeing beyond screen time and self-reports, and to leverage these insights to design effective digital interventions. The PhD student will design, develop, and evaluate novel measurement frameworks, experimental paradigms, and intervention strategies that capture and address the impact of interface design patterns on users' sense of agency, attention, and wellbeing. Possible areas of investigation are:
- Developing a multimodal measurement framework that integrates behavioral signals (e.g., interaction rhythm, scrolling patterns), physiological indicators (e.g., eye tracking), and lightweight self-report instruments to capture the experiential dimension of users' cognitive and affective states during social media use.- Creating validated experimental paradigms for isolating the effects of specific ACDPs on users' cognitive and affective states. This includes developing reusable research tools (e.g., browser extensions) that allow researchers to systematically manipulate design patterns and measure their impact on constructs such as normative dissociation, temporal distortion, and perceived control.- Exploring lightweight approaches for identifying and characterizing ACDPs in user interfaces to help researchers and designers assess the presence and severity of attention-capture mechanisms in existing platforms.- Designing and evaluating digital interventions that restore users' positive cognitive and affective states in the presence of ACDPs. This includes developing and testing concrete design alternatives (e.g., feed-level modifications, reflective nudges, adaptive friction) that promote awareness and control without excessively disrupting the user experience, investigating how to balance effectiveness with perceived intrusiveness across usage contexts.- Conducting longitudinal, in-the-wild studies to understand how cumulative exposure to ACDPs and proposed interventions affect digital wellbeing over time.

The proposal will adopt a human-centered approach, building upon the scientific literature from Human-Computer Interaction, Cognitive Psychology, and Psychophysiology. The work plan is organized in four partially overlapping phases:- Phase 1 (months 0?6): literature review at the intersection of digital wellbeing measurement and ACDPs; focus groups and interviews with researchers, designers, and end users; identification of measurement gaps and intervention opportunities.- Phase 2 (months 3?24): design, development, and validation of multimodal measurement instruments and reusable experimental tools; controlled laboratory studies on the impact of specific ACDPs; development of heuristic-based methods for characterizing ACDPs.- Phase 3 (months 12?36): design, prototyping, and controlled evaluation of digital interventions informed by measurement findings; investigation of the effectiveness-intrusiveness trade-off.- Phase 4 (months 24?36): longitudinal in-the-wild deployments of measurement instruments and interventions; cross-platform generalization; consolidation into guidelines and open-source research tools for the HCI community.

Results are expected to be published at top HCI conferences (e.g., ACM CHI, ACM CSCW, ACM UIST) and in journals such as ACM Transactions on Computer-Human Interaction, International Journal of Human-Computer Studies, Proceedings of the ACM on IMWUT, and Computers in Human Behavior.

[1] A. Monge Roffarello, A. De Luca, Am I in Control? How the Design of the TikTok Feed Shapes Users' Sense of Agency, CHI EA 2026, https://doi.org/10.1145/3772363.3798790

Spatio-Temporal Data Science Applied to Earth Observation

The main objective of this research proposal is to study and design data-driven pipelines and deep learning models to analyze heterogeneous spatio-temporal data in the Earth Observation (EO) domain. Such data include, for instance, multi-spectral satellite imagery, time series from remote sensing platforms, and geo-referenced textual or environmental reports. Both descriptive and predictive problems will be considered, with applications such as environmental monitoring, disaster management, and climate analysis.

The main challenges that will be addressed are as follows.

Heterogeneity. Earth Observation systems inherently produce heterogeneous data characterized by different modalities, resolutions, and formats (e.g., satellite images, sensor measurements, textual reports). Each source provides a partial view of the observed phenomena, and meaningful insights can only be extracted through effective integration. To address this issue, innovative data integration techniques based on aligned latent spaces will be investigated. These approaches will enable the fusion of multi-modal EO data, preserving complementary information and improving the understanding of complex environmental processes.

Scalability. EO data are typically large-scale, due to the continuous acquisition of high-resolution satellite imagery and long-term monitoring data. Efficient big data pipelines and distributed machine learning techniques will be designed to ensure scalability for processing both historical archives of remote sensing/EO data and continuously acquired streams.

Timeliness. Timeliness is critical in EO applications such as natural disaster response (e.g., floods, wildfires), environmental hazard detection, and rapid mapping. Therefore, real-time and incremental machine learning algorithms will be developed to support near real-time analysis and decision-making based on continuously updated EO data streams.
 

The work plan for the three years is organized as follows.

1st year. Analysis of the state-of-the-art algorithms and data science pipelines for spatio-temporal EO data. Based on the strengths and limitations of existing approaches, a preliminary common data representation based on latent spaces will be designed to effectively integrate heterogeneous EO data sources. Novel algorithms will then be developed and validated using historical EO datasets, with applications including disaster-related event detection, environmental change analysis, and data retrieval.

2nd year. The proposed representations for heterogeneous spatio-temporal EO data will be further refined, with a particular focus on scalability and resource-aware learning. Efficient architectures and distributed processing strategies will be investigated to handle large-scale repositories.

3rd year. The timeliness aspect will be addressed by developing near real-time machine learning models for EO data analysis. The focus will be on incremental and streaming-based approaches capable of processing continuously acquired EO data for time-sensitive applications such as early warning systems.
 
The outcomes of the research activity are expected to be published at IEEE/ACM International Conferences and in any of the following journals:
- ACM Transactions on Spatial Algorithms and Systems
- IEEE Transactions on Knowledge and Data Engineering
- IEEE Transactions on Big Data
- IEEE Transactions on Emerging Topics in Computing
- ACM Transactions on Knowledge Discovery in Data
- Information sciences (Elsevier)
- Expert Systems with Applications (Elsevier)
- Machine Learning with Applications (Elsevier)
- IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

Agentic AI for Advanced Temporal Reasoning

https://dbdmg.polito.it/

https://smartdata.polito.it 

Context 
Temporal Reasoning (TR) requires a combination of various skills including mathematical and logical reasoning as well as commonsense knowledge [1,2]. Addressing TR on multimodal databases including, for instance, time series, audio tracks, documents, and images, involves training and fine-tuning multimodal learning architectures, such as transformers and Multimodal Large Language Models (MLLMs), to incorporate the notion of time and effectively handle time-evolving scenarios.  

Challenges 
Multimodal content misalignment: in real-world applications paired inputs in different modalities  (e.g., an audio track and a video content, a time series and its textual annotations) are sometimes partially misaligned due to, for instance, the presence of subtle errors in the audio speech transcription, issues in the audio-video synchronization, or inaccuracies in the human annotations. Since state-of-the-art Multimodal LLMs are usually trained on aligned image-text pairs, the quality of the MLLM outputs could degrade in the presence of content misalignment.    Long context: incorporating historical data in time-evolving scenarios entails prompting MLLMs with long-context inputs, which are either costly using large proprietary models or practically unmanageable using small ones. Tool calling for mathematical reasoning: to efficiently address mathematical reasoning for TR, a common practice is to use LLM calls to external tools. However, as a matter of fact, most MLLMs are specialized neither for TR nor for coding. 


Research objectives Benchmark state-of-the-art Agentic AI frameworks on TR tasks; Propose innovative approaches to detect and mitigate misalignment issues in paired multimodal inputs; Explore the adoption of recent approaches to long-context modeling and reasoning (e.g., [4,5]) for TR; Develop Agentic AI and Reinforcement Learning solutions incorporating innovative TR approaches; Apply the proposed solutions in real-world application domains (finance, IoT, predictive maintenance, telecommunications, energy).

Tentative work plan 
During the first year, the PhD student will study existing AI agent frameworks suited for TR tasks. Focusing on the issue of content misalignment between modalities in multimodal sources, the goal is to also propose ad hoc testing benchmarks, drift detection mechanisms, and mitigation strategies.  

In the second year, the PhD student will extend the study of Agentic AI methods, specifically addressing the issues of long-context modeling and reasoning for TR, exploring the application of novel methods to real-world scenarios.  

In the last year, the PhD student will propose new AI agents and RL approaches to tackle TR tasks and further investigate their application in real industrial scenarios. 

Funding 
The research activities will be carried out under the FIS 2 National Project ?TA-LLM - Large Language Models: a matter of time?, funded by MUR. CUP: E53C25001820001 

Bibliography 
[1] Large Language Models Can Learn Temporal Reasoning. Siheng Xiong, Ali Payani, Ramana Kompella, Faramarz Fekri. ACL 2024. https://aclanthology.org/2024.acl-long.563.pdf 
[2] Timedial: Temporal commonsense reasoning in dialog. Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui. ACL 2021.  
[3] H Liu, C Li, Q Wu, and YJ Lee. Visual instruction tuning. CoRR, abs/2304.08485, 2023. https://doi.org/10.48550/arXiv.2304.08485 
[4] Recursive Language Models. Alex L. Zhang, Tim Kraska, Omar Khattab. https://arxiv.org/abs/2512.24601 
[5] Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Minhao Cheng, Shuai Wang, Hong Cheng, Jingren Zhou. ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization. https://arxiv.org/abs/2509.13313

List of possible publication venues 
- Conferences: ACL, EMNLP, ACM Multimedia, NeurIPS, AAAI, KDD, IEEE ICDM, ECML PKDD, ACM CIKM 
- Journals: IEEE TKDE, ACM TKDD, IEEE TAI, ACM TIST, IEEE/ACM TASLP, TACL

The PhD candidate is expected to  

- Have the ability to critically analyze complex systems, model them and identify weaknesses; 
- be proficient in Python programming; 
- know data science fundamentals; 
- have a solid background on machine learning and deep learning; 
- have natural inclination for teamwork; 
- be proficient in English speaking, reading, and writing. 

Multimodal Temporal Reasoning using Tiny LLMs

https://dbdmg.polito.it/

https://smartdata.polito.it 

Context 
Multimodal Large Language Models (MLLMs) have achieved remarkable success on several multimodal learning understanding and generative tasks. However, they often fail to accurately interpret temporal relationships, such as event orders, temporal entities, and time differences [1,2]. Existing approaches have attempted to solve temporal reasoning issues with Reinforcement Learning techniques [3,4], but the proposed solutions are either suited for textual data only or require costly human annotation and model training.  

Challenge 
Small and tiny LLMs suffer from limited model capability to incorporate temporal information during training. Particularly while dealing with multimodal LLMs, the differences in performance between small and medium/large models become significant.   

Research objectives Benchmarking MLLMs' performance on temporal reasoning tasks; Release new multimodal temporal reasoning for testing and fine-tuning purposes. Propose new temporal reasoning strategies (e.g., teacher-student models, SFT approaches, RL techniques) tailored to tiny LLMs; Study the role of different modalities on temporal reasoning; Integrate the newly proposed solutions into existing frameworks; Define new performance metrics to capture models' effectiveness and efficiency in temporal reasoning; Develop Agentic AI solutions incorporating the newly proposed models; Assess MLLM's reasoning performance on temporal aspects. 


Tentative work plan 
During the first year, the PhD student will first study the performance of existing multimodal LLMs on established temporal reasoning tasks and envisage new possible datasets and testing benchmarks. Then, the student will start developing new temporal reasoning solutions, starting from the most established textual and visual models. 

In the second year, the PhD student will investigate the role of different modalities, entities, and layout information on temporal reasoning performance. New RL models and Agentic frameworks integrating the newly proposed models will be also investigated. 

Finally, in the last year the PhD student will analyze new domains, languages, and temporal aspects that have not been addressed by previously proposed models and propose systematic approaches for evaluating MLLM reasoning on temporal aspects.  

Funding 
The research activities will be carried out under the FIS 2 National Project ?TA-LLM - Large Language Models: a matter of time?, funded by MUR. CUP: E53C25001820001 

Bibliography 
[1] A Survey on Temporal Reasoning for Temporal Information Extraction from Text. Artuur Leeuwenberg, Marie-Francine Moens. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI-20). 
[2] A comprehensive review of temporal reasoning with multimodal LLMs: Foundations, challenges, and healthcare applications. Marouene Chaieb, Nour Ben Ameur, Wadii Boulila. Computer Science Review. Volume 61, August 2026, 100960 
[3] Time-R1: Towards Comprehensive Temporal Reasoning in LLMs. Zijia Liu, Peixuan Han, Haofei Yu, Haoru Li, Jiaxuan You. https://arxiv.org/abs/2505.13508 
[4] MTBench: A Multimodal Time Series Benchmark for Temporal Reasoning and Question Answering. Jialin Chen, Aosong Feng, Ziyu Zhao, Juan Garza, Gaukhar Nurbek, Cheng Qin, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, Rex Ying. https://arxiv.org/abs/2503.16858

List of possible publication venues 
- Conferences: ACL, EMNLP, ACM Multimedia, NeurIPS, AAAI, KDD, IEEE ICDM, ECML PKDD, ACM CIKM 
- Journals: IEEE TKDE, ACM TKDD, IEEE TAI, ACM TIST, IEEE/ACM TASLP, TACL

Continual Learning for Generative Models

https://dbdmg.polito.it/

https://smartdata.polito.it

Context 
Large Language Models (LLMs) are one of the most disruptive technologies of the last years as they are revolutionizing information access, creativity, and text processing tasks. Beyond understanding and generating text in natural language, recently proposed LLMs also support visual, video, and acoustic content as part of the instruction prompts or responses. Multimodal LLMs rely on large-scale generative pre-training, fine-tuning, and human alignment. These processes are inherently dynamic, as languages, data distributions, and user demands continually evolve. When a new model has been trained, it is practically infeasible to account for all possible future scenarios in advance. Therefore, pretrained, fine-tuned, and reward models need to be incrementally updated throughout their lifetime. Such a dynamic scenario is commonly denoted by Continual Learning (CL, in short) [1].  

Challenges 
Due to the significant costs of data annotation and training, a relevant aspect is the opportunistic reuse of past annotations/models/rewards to address new tasks/domains/scenarios. CL performance is often limited by catastrophic forgetting, where learning a new task usually results in a dramatic performance drop of the old tasks [2]. Generative models like textual LLMs and Multimodal LLMs not only demand the incremental updating of predictive models, but also the capability to produce novel content grounded in perceptual understanding [3].  

Research objectives Benchmarking Existing CL techniques for textual generative models; Extend CL techniques for textual LLMs towards multimodal scenarios; Transfer CL models from one domain to another, from one language to another, and from one modality to another; Propose new foundational CL models, including Reinforced CL techniques [4,5]; Study new strategies to fight catastrophic forgetting in challenging scenarios; Generalize CL approaches to make them agnostic to data modality and language; Explore new, challenging application scenarios. 


Tentative work plan 
During the first year, the PhD student will study existing generative CL techniques and compare them with CL solutions for discriminative models. Focusing on textual models first, the PhD will investigate new approaches for adding languages, domains, and tasks. In parallel, the research will explore the suitability of multimodal strategies to extend the transferability of CL techniques from one modality to another. 

In the second year, the PhD student will present new CL approaches, including innovative techniques based on Reinforced CL, and test them on benchmark and real-world data. The research will also address open issues related to catastrophic forgetting. 

In the last year, the PhD student will extend the prior work towards cross-modal, cross-lingual and modality-agnostic scenarios.  

Funding 
The research activities will be carried out under the FIS 2 National Project ?TA-LLM - Large Language Models: a matter of time?, funded by MUR. CUP: E53C25001820001 

Bibliography 
[1] L. Wang, X. Zhang, H. Su and J. Zhu, "A Comprehensive Survey of Continual Learning: Theory, Method and Application," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5362-5383, Aug. 2024, doi: 10.1109/TPAMI.2024.3367329. 
[2] M. McCloskey and N. J. Cohen, ?Catastrophic interference in connectionist networks: The sequential learning problem,? Psychol. Learn. Motivation, vol. 24, pp. 109?165, 1989, doi: https://doi.org/10.1016/S0079-7421(08)60536-8 
[3] Guo, H., Zeng, F., Zhu, F., Wang, J., Wang, X., Zhou, J., and Liu, C. L. (2025). Continual learning for generative ai: From LLMs to MLLMs and beyond. arXiv preprint arXiv:2506.13045. 
[4] Ju Xu and Zhanxing Zhu. 2018. Reinforced continual learning. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (NIPS'18). Curran Associates Inc., Red Hook, NY, USA, 907?916. 
[5] Continual Learning as Computationally Constrained Reinforcement Learning Kumar, S., Marklund, H., Rao, A., Zhu, Y., Hong, J. J., Yueyang L., Van Roy, B.2025, Foundations and Trends in Machine Learning 

List of possible publication venues 
- Conferences: ACL, EMNLP, ACM Multimedia, NeurIPS, AAAI, KDD, IEEE ICDM, ECML PKDD, ACM CIKM 
- Journals: IEEE TKDE, ACM TKDD, IEEE TAI, ACM TIST, IEEE/ACM TASLP,  TACL 

Time-Aware Reinforcement Learning from AI Feedback

Adversarial Robustness in Multi-Modal Foundation Models

Large-scale multi-modal AI models enable richer understanding and generation of content across different modalities. However, the security implications of cross-modal interactions remain poorly understood. Recent evidence suggests that adversaries can craft inputs where, for example, benign visual content paired with carefully manipulated audio or text can cause model misclassification or unsafe outputs. Unlike single-modality adversarial attacks, multi-modal attacks exploit the complex fusion mechanisms that integrate information across modalities, creating attack vectors that are difficult to detect and defend against using existing techniques.

The main objective of the proposed research is to advance understanding of adversarial vulnerabilities in multi-modal AI systems by establishing theoretical foundations for attack surfaces and developing practical defense mechanisms that can be deployed in real-world applications. In this research work, the candidate will leverage expertise in adversarial machine learning, information theory, and secure system design to develop both theoretical insights and practical solutions for multi-modal AI security.

The research activity will be organized in three phases:

Phase 1 (1st year): The candidate will conduct a comprehensive study of the attack surface of multi-modal AI models, focusing on architectures that combine vision, language, and audio modalities. This phase involves analyzing state-of-the-art multi-modal fusion mechanisms to identify potential vulnerability points where cross-modal interactions can be exploited. The candidate will develop a taxonomy of attack vectors specific to multi-modal systems, categorizing them by the exploited modality interactions, attack objectives, and required adversary capabilities.

The candidate will begin developing novel attack methods that exploit semantic inconsistencies between modalities. For example, attacks where visual and textual content appear individually benign but their combination triggers misclassification, or where subtle audio perturbations alter the interpretation of accompanying visual content. At this phase's end, preliminary results are expected to be published, including the attack taxonomy, initial attack methods, and theoretical characterizations of multi-modal vulnerability surfaces. During the first year, the candidate will also acquire necessary background through coursework in adversarial machine learning, information theory, and multi-modal AI architectures, supplemented by personal study.

Phase 2 (2nd year): In the second phase, the candidate will build upon the knowledge acquired during the first year to create a dataset of synthetic adversarial attacks. The dataset will cover several attack techniques, including cross-modal perturbation, semantic inconsistency exploitation, modality-specific backdoors. The dataset will target different types vulnerabilities and it will cover different application contexts. This dataset will serve as a benchmark for evaluating defense mechanisms and will be made publicly available to support the research community.

The dataset will be validated through an experimental campaign to demonstrate the effectiveness of the attack methods in exposing vulnerabilities. Applications in content analysis systems, customer service AI agents, and AI-assisted software development tools will serve as case studies.

Phase 3 (3rd year): Building on previous results, the candidate will develop new attack techniques capable of automatically synthesizing attack samples. These techniques will be implemented as a framework of parameterized attack templates to generate diverse adversarial examples across different multi-modal architectures. The attack generation framework will incorporate the results from Phase 2 to prioritize attack strategies with higher success probabilities. The candidate will refine and validate the attack mechanisms through extensive testing on multimodal models. The final phase will produce comprehensive documentation, open-source implementations of attack tools, and best practices guidelines for developing robust multi-modal AI systems. Dissemination activities will include publications, software releases, and potentially workshops or tutorials to transfer knowledge to practitioners.

The work may be conducted in collaboration with industry partners deploying multi-modal AI systems or relevant research initiatives focusing on AI safety and security.

The contributions produced by the proposed research can be published in conferences and journals belonging to the areas of Machine Learning and AI (e.g., NeurIPS, ICML, ICLR, CVPR, ACL, AAAI, JMLR, IEEE TPAMI), Cybersecurity (e.g., IEEE S&P, ACM CCS, USENIX Security, NDSS, ACM Transactions on Privacy and Security), and AI safety (e.g., AIES, FAccT).

The activities will be carried out within the scope of a research collaboration between the Italian Institute of Artificial Intelligence for Industry and Politecnico di Torino.

Privacy-Preserving Machine Learning over IoT networks.

Stack and Compilation Techniques for Hybrid Algorithms on Fault-Tolerant Quantum Architectures

Fairness-Aware Generative AI for Socio-technical Systems 

Generative AI is increasingly embedded in socio-technical systems that support communication, assessment, recommendation, profiling, moderation and decision preparation, such as customer-support chatbots, automated essay scoring in education, recommendation systems on digital platforms, and decision-support tools used in domains like hiring, healthcare, and public administration. In these settings, biased model behaviour can implicitly embed political views, produce unequal treatment, distorted representations and unfair downstream effects. This PhD position addresses the problem from a fairness-aware software and systems perspective, with the goal of developing audit methods, tools, workflows and design principles for the analysis and mitigation of bias in generative AI.

Recent work in this area has shown that unfair behaviour in LLMs can emerge in different forms. Prior studies attached to this line of research have shown, for example, that models may reproduce gendered stereotypes in generated descriptions and may also react differently to semantically equivalent inputs when those inputs vary by dialect, producing stereotype-bearing differences in adjectives, occupational associations, trust judgments and inferred background. These cases are not the exclusive focus of the PhD, but they provide initial concrete evidence that can be expanded upon in terms of depth or breadth during the research. that fairness risks in generative AI are broader than any single application domain and can arise from both social and linguistic signals.

The research will investigate fairness-aware generative AI as a general complex socio-technical challenge that integrates technical dimensions?such as training data balance, model architecture, and optimization strategies?with core concepts from the social sciences. It will examine how bias manifests across textual and multimodal outputs; how such behaviour can be evaluated through controlled empirical protocols; and how mitigation can be embedded into the architecture of AI-enabled systems. Particular attention will be given to the comparison between prompt-based mitigation and process-based mitigation, since recent evidence suggests that critique-and-revision workflows and multi-agent orchestration may provide more stable mitigation than single-pass prompting alone.

The work is expected to produce a fairness-aware framework for generative AI in socio-technical systems, combining empirical evaluation methods, reusable benchmark designs, mitigation workflows and software engineering guidance. The research plan may include the construction of matched-input evaluation protocols, cross-model comparative studies, analysis of representational and allocational harms, and the design of agentic or workflow-level controls to reduce biased outputs in realistic deployment settings. The broader aim is to support the development of trustworthy generative AI systems that are not only technically effective, but also fairer, more transparent and more robust in practice.

Phase 1 (Months 1-12) - Benchmark design and harm identification.
Develop a hiring-specific fairness framework for generative AI applied to CV screening, candidate ranking, and interview support. Construct matched-input candidate profiles that vary protected attributes while holding qualifications constant. Build a reusable benchmark and conduct pilot evaluations to identify representational harms in generated content and allocational harms in shortlisting and ranking outcomes.

Phase 2 (Months 13-24) - Comparative evaluation and mitigation.
Run large-scale cross-model studies measuring disparities in selection rates, ranking positions, feedback tone, and salary recommendations. Design and experimentally test mitigation strategies at the model level (data adjustments, constrained generation, fairness-aware fine-tuning) and at the workflow level (re-ranking procedures, structured human oversight). Quantify disparity reduction and analyze trade-offs with predictive performance.

Phase 3 (Months 25-36) - Deployment testing and operational consolidation.
Integrate mitigation workflows into simulated recruitment pipelines reflecting realistic organizational constraints. Evaluate robustness under distribution shifts and adversarial prompting. Consolidate results into a deployable framework including benchmark tools, mitigation pipelines, monitoring protocols, and engineering guidance for fair and compliant use of generative AI in hiring systems.

List of possible venues for publications
The position is aligned with publication venues at the intersection of software engineering, responsible AI and human-centred AI, including FSE, ICSE, ASE, ESEM, AIES, FAccT, EMSE and ACM TOSEM. The research can connect with collaborations involving academic groups, interdisciplinary centres, public bodies and industrial organisations interested in trustworthy AI, digital transformation and socio-technical evaluation of AI-enabled systems.