Leonardo Ranaldi
I am a Research Fellow at the University of Edinburgh and Researcher at the University of Rome Tor Vergata. My research interests lie in NLP and AI, focusing on LLM reasoning, self-learning, memory-augmented systems, multimodal reasoning, neuro-symbolic methods, and controllable and aligned models. I study the extent to which generative models can learn from interaction, acquire and use structured reasoning capabilities, and adapt their behaviour across contexts and over time, with the aim of steering generative architectures beyond statistical shortcuts towards reliable and more controllable systems.
I work in international contexts, collaborating with researchers across academia and industry. I am the PI of a Google for AI Research computing grant (Building Trustworthy AI), Co-PI of a Huawei–Edinburgh research grant (Adaptive Memory-Augmented Agentic Reasoning), and I have been WP Leader in the Horizon Europe project UTTER. I publish regularly in leading international AI and CL/NLP venues and journals, and I contribute to the research community as a Senior Area Chair for major conferences. I also organise international workshops, including ORACLE and MathNLP, tutorials on reasoning, neuro-symbolic modelling and agentic AI, and an international shared task on multilingual reasoning.
Research
My work treats adaptation as a reasoning problem, in which structure indicates what should remain invariant, evidence indicates when to revise, memory determines what persists and prediction determines how far to act. The essay Adaptive Reasoning Agency retells this line of work with its mechanisms, diagnostic evidence and open questions.
- Reasoning and abstraction. Quasi-symbolic and schema-guided representations that disentangle content from logical structure, and their counterpart in the activations of the model.
- Adaptive and memory-augmented agents. Agents that consolidate experience into reusable procedures, anticipate failure through world models, and share validated procedures across a team.
- Controllable and aligned models. Agreement under user pressure formulated as a decision policy, with oversight through dialectic reasoning.
- Multilingual and multimodal reasoning. Transfer of reasoning procedures across languages, modalities and model families.
Experience
Research Fellow and Co-PI
ILCC, School of Informatics, University of Edinburgh, UK. Mar 2024
- Develop and co-lead research in NLP focusing on reasoning, self-improving and memory-augmented models, multilingual and multimodal reasoning, and reliable agentic AI.
- Co-PI of the Huawei–Edinburgh project on Adaptive Memory-Augmented Agentic Reasoning, coordinating research activities and mentoring junior researchers.
- Co-supervise PhD students within the School of Informatics, and contribute to teaching, assessment and student supervision.
- Work Package Leader in Horizon Europe UTTER (UoE unit, WP9 Ethics and data), coordinating research activities across academic and industrial partners.
Researcher
Department of Digital Humanities, University of Rome Tor Vergata, Italy. Jun 2025
- Research in NLP and generative AI, with a focus on LLMs, reasoning, neuro-symbolic methods, alignment and adaptive AI.
- Design and deliver undergraduate and postgraduate teaching in AI, Data Science and Information Systems, including curriculum development, assessment and student supervision.
Postdoctoral Industrial Research Fellow
Idiap Research Institute, Switzerland. Feb 2023 – Sep 2024
- Developed neuro-symbolic NLP frameworks for robust human–model alignment, connecting mathematical reasoning with safe AI deployment.
PhD Candidate and Pre-Postdoc
University of Rome Tor Vergata, Italy. Sep 2020 – Feb 2023
- Researched symbolic interpretability for transformer-based models, providing transparent frameworks to audit model behaviour. Supervised master’s and PhD students.
Intern Researcher
Mashfrog Group, Italy. May 2020 – Nov 2020
- Conducted applied R&D in AI and NLP, collaborating with industry stakeholders to translate technical models into business-oriented solutions.
Education
Italian National Scientific Qualification (Abilitazione Scientifica Nazionale)
Italian Ministry of Universities and Research. Jul 2026
Qualified for appointment at Associate Professor level in Computer Science (01/B1) and Computer Engineering (09/H1).
PhD in Physical Sciences and Industrial and Innovation Engineering
University of Rome Tor Vergata and University Guglielmo Marconi, Italy. Sep 2020 – Sep 2023
Thesis: It is Time to Reason: Pre-trained Linguistic Models Out of the Way. Supervisors: Prof. Fabio Massimo Zanzotto and Prof. Francesca Fallucchi.
Master’s Degree in Computer Science, cum laude
University of Rome Tor Vergata, Italy. Sep 2018 – Jul 2020
Supervisor: Prof. Fabio Massimo Zanzotto.
Research leadership and funding
- Huawei–Edinburgh Joint Research Lab, Co-PI. University of Edinburgh, £210,000. Adaptive Memory-Augmented Agentic Reasoning, on adaptive memory, agentic reasoning and self-improving AI systems.
- Google for AI Research Grant, PI. $35,000. Building Trustworthy AI, on reliable reasoning and RAG models.
- Almawave, PI. University of Rome Tor Vergata, €20,000. Memory-Augmented LLM Architecture.
- Horizon Europe UTTER, Work Package Leader and Beneficiary. University of Edinburgh. Unified Transcription and Translation for Extended Reality.
- Other academic grants. STEP (Co-PI, €17,426), Public Engagement Tor Vergata (Co-PI, €16,953), DataLake@2023 (Beneficiary, €20,897), Cohere For AI Research Grant (€2,000).
Selected publications
The complete list is on Google Scholar and Scopus.
2026
- World-Model Evolving Agents. L. Ranaldi. EMNLP 2026.
- Abstract Activation Spaces for Content-Invariant Reasoning in Large Language Models. G. Maraia, M. Valentino, F. M. Zanzotto, L. Ranaldi. EMNLP 2026.
- Evolving Agents. L. Ranaldi. ACL 2026.
- Agentic Oversight via Dialectic Reasoning. L. Ranaldi, F. Ranaldi. ACL 2026.
- Thinking in Schemas: Robust Syllogistic Reasoning in LLMs. F. Ranaldi, L. Ranaldi, F. M. Zanzotto, S. Cohen. ACL 2026.
- Learning Multilingual Agentic Policy to Control Sycophancy. L. Ranaldi, G. Pucci. EACL 2026.
- Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Task. L. Ranaldi, B. Haddow, A. Birch. Findings of EACL 2026.
2025
- Advancing Oversight Reasoning across Languages for Audit Sycophantic Behaviour via X-Agent. G. Pucci, L. Ranaldi. EMNLP 2025.
- Improving Multilingual RAG Models through Dialectic Reasoning Argumentations. L. Ranaldi, F. Ranaldi, F. M. Zanzotto, B. Haddow, A. Birch. EMNLP 2025.
- Improving Chain-of-Thought Reasoning via Quasi-Symbolic Abstractions. L. Ranaldi, M. Valentino, A. Freitas. ACL 2025.
- R2-MultiOmnia: Multilingual Multimodal Reasoning. L. Ranaldi, F. Ranaldi, G. Pucci. ACL 2025.
- Multilingual Reasoning via Self-training. L. Ranaldi, G. Pucci. NAACL 2025.
2024
- Self-Refine Instruction-Tuning for Aligning Reasoning in Language Models. L. Ranaldi, A. Freitas. EMNLP 2024.
- Empowering Multi-step Reasoning across Languages via Program-Aided Language Models. L. Ranaldi, G. Pucci, B. Haddow, A. Birch. EMNLP 2024.
- Aligning Large and Small Language Models via Chain-of-Thought Reasoning. L. Ranaldi, A. Freitas. EACL 2024.
Teaching and supervision
Course Lecturer
University of Rome Tor Vergata. 2025 – present
- AI and Social Media Analytics (18 ECTS, 90 contact hours).
- Data Science and Python (6 ECTS, 30 contact hours).
- Database and Information Systems (6 ECTS, 30 contact hours).
Teaching Assistant
University of Edinburgh. 2024 – 2025
- Natural Language Understanding+ (NLU+), School of Informatics, contributing to assessment design, tutorials, office hours and student supervision.
Adjunct Professor
2022 – 2023
- Practical Natural Language Processing, University of Rome Tor Vergata (3 ECTS, 15 contact hours).
- Deep Learning, University Guglielmo Marconi (6 ECTS, 30 contact hours).
Mentoring and student supervision
- Co-supervisor of PhD students Elena Ruzzetti, Davide Venditti and Michele Mastromattei.
- Supervisor of more than five Master’s and Bachelor’s theses in Computer Science and Engineering, on neuro-symbolic AI, data processing pipelines and mechanistic interpretability.
- Mentor of junior and postdoctoral researchers within the Huawei–Edinburgh Joint Lab.
Academic service
Chairing and editorial work
- Senior Area Chair for EMNLP (2025, 2026) and EACL (2026, 2027).
- Area Chair for EMNLP (2023), EACL (2024), COLING (2025) and LREC-COLING (2024).
- Action Editor for ACL Rolling Review (2023–2025), Guest Editor for Future Internet and Big Data and Cognitive Computing, and reviewer for TACL and TMLR.
Workshops, shared tasks and tutorials
- ORACLE, Workshop on Open Reasoning Across Cultures and Languages (EMNLP 2026, EACL 2026).
- MathNLP, Workshop on Mathematical Natural Language Processing, fourth edition at EMNLP 2026.
- SemEval-2026 Task 11 on Multilingual Syllogistic Reasoning (ACL 2026).
- Tutorial on Mechanistic Interpretability and Neuro-Symbolic Modelling by Design (EKAW 2026).
- Tutorial on Neuro-Symbolic Augmented LLMs Meet Interpretability (CIKM 2026, ISWC 2026).