
Open positions in the 2027 spring session
Deadline for the Application: November 16th, 2026
Beginning of PhD: March, 2027
Candidates should contact contacts.smartdata@polito.it and check the official page for requirements and details.
In your email, please include:
- The position you are interested in
- Curriculum Vitae (CV)
- Master’s degree transcript
- Eventual academic references
Understanding Autonomous Agent Decisions: Explainability and Trustworthiness
Danilo Giordano (DAUIN)
Marco Mellia (DAUIN)
Matteo Boffa (DAUIN)
Description: Autonomous agents make decisions through interactions among planning, memory, tools, and communication components. Their executions produce observable evidence, including actions, tool calls, retrieved information, state changes, and stated explanations. This evidence offers opportunities to investigate agent behaviour, but interpreting it requires care: an agent’s explanation may be incomplete or inconsistent with the factors that actually influenced its actions.
The proposed research will investigate how execution evidence can support the interpretation and auditing of autonomous agents. A central challenge is to establish which observations justify a particular explanation or reliability assessment. Logs may document what happened without establishing why it happened, while reported confidence alone may provide insufficient evidence that a decision is sound. The candidate will develop methods to reconstruct decision paths, test behavioural explanations, and communicate the uncertainty of diagnostic conclusions.
Security will provide a central setting for this work. Attackers may manipulate memory, tool outputs, or communication channels, producing behaviour that resembles an ordinary mistake. The central research question is whether execution analysis, supported by controlled interventions, can distinguish competing explanations for unexpected behaviour and identify evidence of compromised components.
Expected contributions include methods for analysing execution traces, validating behavioural explanations, detecting signs of manipulation, and producing auditable accounts of agent decisions. Evaluation will measure the accuracy of these diagnoses and their supporting evidence, including false alarms and cases where the available observations do not support a firm conclusion.
Research Challenges and Key Questions:
- Evidence-Based Interpretation: What can execution traces establish about an agent’s decisions, and how can stated explanations be tested against observed actions and controlled interventions?
- Reliability Signals: Which observable signals support an assessment of decision reliability, and how should uncertainty be communicated when evidence is incomplete or conflicting?
- Security Diagnosis: Under what conditions can behavioural analysis distinguish ordinary errors from malicious manipulation and localise potentially compromised components?
- Auditable Decisions: How can evidence across components be reconstructed into verifiable accounts that support human investigation and security oversight?
Required background for the PhD candidate:
- Strong programming skills in Python.
- Knowledge of AI, LLMs, and autonomous agents.
- Basic knowledge of cybersecurity and adversarial attacks.
- Familiarity with explainable AI and uncertainty estimation
Network Operators as Active Members of Renewable Energy Communities: Joint
Optimization of Radio Access, Edge Intelligence and Local Energy
Daniela Renga (DET)
Greta Vallero (DET)
Michela Meo (DET)
Mobile networks are an enabling infrastructure for the digital transition, but they are among the largest electricity consumers of the ICT sector. Radio access network (RAN) densification in 5G and beyond, together with edge computing nodes hosting AI services close to users, places a growing, distributed and continuous energy demand on the grid. Most of this demand is still treated as a fixed load, managed by the operator in isolation from the local energy system.
At the same time, renewable energy communities are emerging as a model to increase local self-consumption, flexibility and the integration of renewable sources. Their members are mostly residential and commercial buildings; telecommunications infrastructure rarely takes part. Yet network sites have features that make them valuable community members: a demand spread across the territory, on-site batteries, rooftop space for photovoltaics and, above all, a flexibility that buildings lack. Base stations can switch carriers or enter sleep modes at low traffic, and edge nodes can shift, offload, or simplify their AI workloads.
This flexibility is not free. Reducing the energy of the access network or of edge inference affects served traffic, quality of service, latency, and the accuracy of AI models. Running a lighter model at the edge saves energy but may lower accuracy; offloading to the cloud saves local energy but adds latency and transmission costs. Deciding when and how the operator should adapt therefore requires models that capture network, AI and energy performance together.
The proposed research will study the RAN operator as an entity consuming energy for both the access network and edge computing, as an active member of renewable energy communities. Its central question is how network and edge AI operations can be jointly optimized with community energy flows, so that the operator meets service and accuracy requirements while improving the community’s self-consumption, peak demand, cost and emissions. The candidate will combine power modeling of network and edge infrastructure, measurements of AI inference on edge hardware, simulation of energy communities with real traffic and generation data, and optimization and learning-based control. Energy, environmental, economic, network and AI indicators will be measured jointly.
Research Challenges and Key Questions:
- Joint Performance Modeling: How can models jointly capture the energy consumption of RAN and edge infrastructure and its relationship with traffic, quality of service, latency, and AI accuracy?
- Service-Aware Flexibility: When and how should base stations adjust carriers or enter sleep modes, and edge nodes shift, offload, or simplify AI workloads, while meeting service and accuracy requirements?
- Community Energy Optimization: How can network and edge AI operations be coordinated with local renewable generation, batteries, and community energy flows to improve self-consumption and reduce peak demand, costs, and emissions?
- Adaptive Control: How can optimization and learning-based methods coordinate network, AI, and energy decisions as traffic, workloads, and renewable generation vary?
- Integrated Evaluation: How can measurements and simulations using real traffic and generation data jointly assess energy, environmental, economic, network, and AI performance?
Required background for the PhD candidate:
- Strong programming skills in Python for simulation and data analysis
- Knowledge of optimization
- Basic knowledge of mobile networks architecture
- Familiarity with machine learning for time-series forecasting
- Interest in energy systems and sustainability and knowledge of renewable energy communities are a plus.
Green Edge-Cloud Continuum: Energy and Carbon Trade-offs between Centralized
and Distributed Computing Infrastructure
Daniela Renga (DET)
Greta Vallero (DET)
Michela Meo (DET)
Data centers are the backbone of cloud services and, increasingly, of artificial intelligence (AI). The rapid growth of AI workloads, from model training to large-scale inference, is driving a steep rise in electricity demand, rack power density, and the need for cooling. Energy cost, carbon emissions, water use, and grid capacity are now first-order constraints on where and how computation is deployed.
Two architectural trends pull in different directions. Large, centralized data centers exploit economies of scale: high server utilization, efficient cooling, access to large renewable contracts and a favorable power usage effectiveness (PUE). Distributed infrastructures, made of regional sites, edge data centers and network-embedded nodes, bring computation closer to users and data, reducing latency and long-distance traffic, and can exploit local renewable generation and waste-heat reuse. Yet they usually run at lower utilization, with less efficient cooling and a larger fixed overhead per unit of computation. Which option is greener depends on the workload, location, energy mix, and network, and the answer changes over time with hardware, models, and grids.
Current comparisons are often limited to a single metric, such as PUE, or ignore parts of the picture: the energy of the network that connects users to the computation, the embodied carbon of hardware, the time-varying carbon intensity of electricity, the cost of keeping servers idle, and the quality of service that applications require. Without a complete and comparable view, operators lack the evidence to decide how to size, place, and operate their infrastructures.
The proposed research will study the energy and carbon efficiency of centralized and distributed data center architectures, with particular attention to AI inference and cloud workloads, and use the findings to improve how infrastructures are designed and operated. Its central question is under which conditions a big data center, a distributed infrastructure, or a combination of the two minimizes energy use and emissions while meeting service requirements. The candidate will combine power and carbon modeling of data centers and networks, measurements on real hardware, simulation with real workload and energy data, and optimization and learning-based control. The work will jointly measure energy, carbon, cost, latency, and service quality.
Research Challenges and Key Questions:
- Holistic Energy and Carbon Modeling: How can models jointly capture data center and network energy use, cooling and idle overheads, embodied hardware carbon, and time-varying electricity carbon intensity?
- Architectural Trade-offs: Under which workload, location, energy mix, and network conditions do centralized, distributed, or hybrid infrastructures minimize energy use and emissions while meeting service requirements?
- Infrastructure Design: How should operators size and locate computing resources to balance utilization, cooling efficiency, access to renewables, waste-heat reuse, and proximity to users?
- Adaptive Operation: How can optimization and learning-based control guide workload placement and resource allocation as demand and energy conditions change, while balancing energy, emissions, cost, latency, and service quality?
- Integrated Evaluation: How can hardware measurements and simulations using real workload and energy data provide complete, comparable assessments of alternative architectures, particularly for AI inference and cloud workloads?
Required background for the PhD candidate:
- Strong programming skills in Python for simulation and data analysis
- Knowledge of optimization
- Basic knowledge of mobile networks architecture
- Familiarity with machine learning for time-series forecasting
- Interest in energy systems and sustainability and knowledge of renewable energy communities are a plus.
Towards Reliable Multimodal Agents
Marco Mellia (DAUIN)
Matteo Boffa (DAUIN)
Danilo Giordano (DAUIN)
Description: Large language models (LLMs) increasingly power agents that gather information and take actions, such as searching the web, processing documents, or running code. These agents operate on growing volumes of text, images, audio, and video, often accumulating information over multiple steps and sessions. Their effectiveness depends on how this information is represented, organised, and made available within limited context and resource budgets.
The proposed research will investigate how input properties, context organisation, and memory policies affect agents’ ability to complete tasks accurately and efficiently. Relevant information may become difficult to access as inputs grow, be lost during compression, or be retrieved in a form that is poorly suited to the task. Multimodal inputs introduce further choices about representation, granularity, and alignment across sources. Providing more information or retaining more history may improve performance, but can also increase latency and cost.
The central research question is how to organise and manage information so that multimodal agents achieve the best task performance within a given resource budget. The candidate will combine controlled experiments with realistic workflows to measure the effects of input representation, context selection, compression, and memory retrieval. These findings will guide methods that adapt information provision to the task, modality, and available resources.
Expected contributions include benchmarks that isolate information-processing limits, measurements of quality–cost trade-offs, and practical methods for context and memory management.
Research Challenges and Key Questions:
- Input Representation: How do the format, length, position, and modality of information affect task performance, and which representations make relevant evidence easier to use?
- Context and Memory: What information should agents retain, compress, retrieve, or reread as workflows grow and tasks span multiple sessions?
- Performance Measurement: How can benchmarks isolate information-processing limits and measure their effects on accuracy, stability, latency, and cost?
- Resource-Aware Adaptation: How can agents adapt context and memory policies to task requirements and resource budgets across models and applications?
Required background for the PhD candidate:
- Strong programming skills in Python.
- Knowledge of AI Agents, Large Language Models and Multimodal Models.
- Knowledge of experimental design and statistical analysis.
- Basic knowledge of cybersecurity
For more details, you may contact: contact@smartdata.polito.it.
In your email, please include:
- The position you are interested into
- Curriculum Vitae (CV)
- Master’s degree transcript
- Eventual academic references