I welcome motivated master's students interested in AI, robotics, and applied machine learning; all with a focus on LLMs. Below you’ll find a selection of current project ideas. Each project is open-ended, you’re expected to bring your own perspective and interest and adapt it into a viable thesis plan. More information on how to apply is at the end.
Reinforcement learning from human feedback (RLHF) and its variants (DPO, RLOO, GRPO) rely on gradient-based optimisation. However, when models are large and fine-tuning resources limited, zeroth-order (gradient-free) methods offer a promising alternative.
Implement and benchmark zeroth-order preference optimisation on open-source LLMs.
Explore forward-mode differentiation in LLMs
Theoretical analysis of convergence and scalability
"Visualising Policy-Reward Interplay to Inform Zeroth-Order Preference Optimisation of Large Language Models" Galatolo et al. 2025
"Fine-tuning language models with just forward passes" Malladi et al. 2023
Most modern VLMs rely on pairing a vision encoder with a pretrained LLM; while easy to train, this methodology creates obstacles in allowing a seamless feedback loop between text and vision. Some work integrate tools to aid in the process [1] but this breaks the gradient flow.
Explore new VLM architectures to (better) integrate text and vision
Explore cross-model link modules in existing VLM architectures
Abductive reasoning is often overlooked when training and testing LLMs in favour of the more straightforward inductive/deductive reasoning. This creates a problem in particular scenarios, including but not limited to moral reasoning.
Evaluate LLMs for this problem and identify common failure modes
Understand common abductive mechanisms and improve on current sota LLMs
Bridge philosophy and AI for a work that is grounded in theory but useful to the real-world
"Beyond Ethical Alignment: Evaluating LLMs as Artificial Moral Assistants" Galatolo et al. 2025
"Towards a theory of abduction based on conditionals" Pfiser, 2022
Diabetes prevention programmes often fail to engage populations at risk. Social robots and AI agents offer a way to deliver tailored behavioural support, with dialogue, goal tracking, and personalised coaching.
Prototype an LLM-driven coaching agent.
Design small-scale user studies with simulated interventions.
Explore personalisation strategies based on user modelling and Sense-of-Coherence (SoC).
TBA
Personality is a very impactful factor for social agents (virtual or embodied) that interact with users. For this reason, embedding and manipulating personality in LLMs is a very important topic. However, current research so far has been confined to questionnaire assessment and manipulation through prompting, which is suboptimal
Investigate more robust ways to assess LLMs' embedded personality
Find new ways to manipulate expressed personality of LLMs beyond prompting
Do user studies with embodied or virtual agents to check the effects of personality manipulation
TBA
If you are interested, send me:
A short statement of interest (1 page max): which project got your attention? What are your interests? How are you going to apply them to this project? Please don't ask chatGPT to write this, I want to understand your interests.
Your CV
Relevant transcripts (informal is fine)
Contact: alessio.galatolo@it.uu.se