For Spring 2027, I will be accepting 2-3 motivated master's students interested in AI, Agents, 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. The listed directions are examples; a good thesis usually goes deep on one or two of them. More information on how to apply is at the end. Note that these projects can be expanded either in a master's thesis or in a project within the scope of a course (e.g., open course).
I expect my students to pursue a publication (workshop/full paper) either during or following the completion of the master's thesis.
Today's alignment pipelines assume abundant labels and a single "average" user. They also increasingly assume an LLM judge that can stand in for humans. Our recent results show that this breaks down for minority, cultural, or personal preferences. Aligned models encode mainstream preferences linearly in their activations, but that structure disappears when the target preferences differ, so LLM judges become poor proxies for those users. Even so, a handful of labelled examples combined with a model's internal representations can go a long way. This project asks the next question: what is the minimal signal needed to adapt a model to an individual or a community, and how should that signal be represented, propagated, and trusted? The scope runs from theory (when is a preference linearly decodable?) to building and evaluating a working personalisation system.
When and why preferences are linearly separable in activation space; transfer of preference probes across layers, models, and populations; links to the linear representation hypothesis.
Active learning or elicitation strategies that choose which few pairs to ask a user about; robustness to noisy, inconsistent, or drifting human labels.
Compare training-time personalisation (DPO/IPO on propagated labels, LoRA-per-user, mixtures of adapters) with training-free alternatives (steering vectors, decoding-time reweighting, hypothesis-based prompting).
Aggregating many small groups without erasing any of them, and measuring who wins and who loses (e.g. with PRISM's demographics).
Galatolo & Beloucif, Low-Resource Preference Adaptation of LLMs via Activation-Based Label Propagation, EMNLP 2026)
Kirk et al., The PRISM Alignment Dataset, NeurIPS D&B 2024; Kirk et al., The benefits, risks and bounds of personalizing the alignment of LLMs to individuals, Nature Machine Intelligence 2024
Sorensen et al., A Roadmap to Pluralistic Alignment, ICML 2024
Poddar et al., Personalizing RLHF with Variational Preference Learning, NeurIPS 2024; Jang et al., Personalized Soups, 2023
Garbacea & Tan, HyPerAlign: Interpretable Personalized LLM Alignment via Hypothesis Generation, arXiv 2505.00038, 2025; Drift: Decoding-time Personalized Alignments, Findings of EMNLP 2025
Park et al., The Linear Representation Hypothesis and the Geometry of LLMs, ICML 2024; Zou et al., Representation Engineering, 2023
Rafailov et al., DPO, NeurIPS 2023; Azar et al., A General Theoretical Paradigm to Understand Learning from Human Preferences (IPO), AISTATS 2024
Backpropagation is the memory bottleneck of LLM post-training. It is also what keeps alignment out of reach on consumer GPUs, edge devices, and federated settings. Zeroth-order (ZO) and evolution-strategy (ES) methods estimate updates from forward passes only. Until recently they were considered too noisy for anything beyond classification. Our work (ZOPrO, CA-ZODPO) showed that ZO preference optimisation becomes viable when the search is structured, for example by policy-reward dynamics or by the activation contrast between chosen and rejected responses. Recent work also reports that full-parameter ES fine-tuning of billion-parameter models can compete with RL. The big question is whether forward-only optimisation can become a real alternative paradigm for post-training, and what new possibilities that opens (non-differentiable rewards, on-device personalisation, privacy).
Geometry of the DPO/IPO loss (local curvature, Hessian structure) and what it implies for ZO convergence rates and for choosing the perturbation subspace.
Better search subspaces (activation-guided, low-rank, Hessian-informed, learned); variance reduction; ZO/ES for RL with verifiable rewards (GRPO-style reasoning tasks).
Forward-only training on a single consumer GPU or in a federated setup where each device personalises its own model. This connects naturally to Project 1.
Forward-only knowledge distillation (e.g. ZO with hidden-state alignment) to produce small deployable models.
When does ES/ZO actually beat first-order RL (reward hacking, long horizons, stability), and when is it just cheaper?
Galatolo et al., Visualising Policy-Reward Interplay to Inform Zeroth-Order Preference Optimisation of LLMs (ZOPrO), Findings of ACL 2025
Yuang & Galatolo, Contrastive Activation-Guided Zeroth-Order Preference Optimisation (CA-ZODPO), AXIOM @ NeurIPS 2026
Malladi et al., Fine-Tuning Language Models with Just Forward Passes (MeZO), NeurIPS 2023
Zhang et al., Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark, ICML 2024
Zhao et al., Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer (HiZOO), ICLR 2024
Qiu et al., Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning, ICML 2026
Salimans et al., Evolution Strategies as a Scalable Alternative to Reinforcement Learning, 2017; Spall, SPSA, 1992
Shao et al., DeepSeekMath (GRPO), 2024
VLMs often answer from language priors rather than from the image. They hallucinate objects, ignore fine detail, and perform worse as their reasoning chains get longer. In our ICRA work, a small feedback loop from the language model back to the vision encoder already improved robotics tasks. This suggests that perception should be iterative and query-driven, not a single fixed encoding. Several recent lines of work push in the same direction: "thinking with images", RL rewards for grounding, latent reasoning, and modality-specific routing inside the network. The big question is how to build and train VLMs whose reasoning depends on visual evidence in a measurable way, and that do this efficiently.
RL rewards that measure visual dependence (counterfactual image perturbations, consistency across views, grounding accuracy) without hurting general capabilities.
Iterative or latent "look-again" passes; modality-routed, MoE-like branches retrofitted onto encoder-free VLMs to control over-reliance on text.
Query-aware visual token pruning and re-attending. Which patches matter, and when do they need to come back?
Causal or interventional metrics of how much an answer actually depends on the image, as opposed to benchmark accuracy.
Galatolo, Cumbal et al., Lightweight Visual Reasoning for Socially-Aware Robots, ICRA 2026
Zheng et al., DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning, ICLR 2026
Liu et al., Visual-RFT: Visual Reinforcement Fine-Tuning, ICCV 2025
Tong et al., Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, CVPR 2024
Leng et al., Mitigating Object Hallucinations via Visual Contrastive Decoding, CVPR 2024
Chen et al., An Image is Worth 1/2 Tokens After Layer 2 (FastV), ECCV 2024
Lin et al., MoMa: Efficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts, 2024; Diao et al., Unveiling Encoder-Free VLMs (EVE), NeurIPS 2024
Hao et al., Training LLMs to Reason in a Continuous Latent Space (Coconut), COLM 2024
Reasoning models produce long, unstructured traces, but we mostly judge them by their final answer. In our own work, we turned traces into reasoning graphs and found that abductive leaps help but are underused, and that graph depth predicts quality. Separately, our moral-assistant benchmark found that abductive (moral) reasoning, i.e. inferring the values at stake in a situation, is where LLMs fail most. Both results point to the same gap: we lack tools to evaluate how models reason, and we lack ways to train that reasoning directly. This project moves from analysis to intervention. Can structural properties of reasoning become training signals, and can they make models better at reasoning under uncertainty, including in ethically loaded settings?
Use reasoning-graph properties (coverage, depth, valid abductive steps) as process rewards for RL. Does the model then reason better or just game the structure?
Benchmarks and training for hypothesis generation versus hypothesis selection, for example in diagnosis-like or scientific settings.
Do the extracted graphs reflect the model's actual computation? Combine with causal or attention interventions on chain-of-thought steps.
A candidate with a background in philosophy, logic, or cognitive science could formalise what a "good" abductive or moral argument is.
Nguyen & Galatolo, Ask Your Language Model Why: Analysing LLM Abductive Reasoning Through Graphs (under review, 2026)
Galatolo, Rappuoli et al., Beyond Ethical Alignment: Evaluating LLMs as Artificial Moral Assistants, ECAI 2025
Sun & Saparov, Do Language Models Follow Occam's Razor? Parsimony in Inductive and Abductive Reasoning, arXiv 2509.03345, 2025
Salimi et al., Wiring the 'Why': A Unified Taxonomy and Survey of Abductive Reasoning in LLMs, arXiv 2604.08016, 2026
Bhagavatula et al., Abductive Commonsense Reasoning, ICLR 2020
Lightman et al., Let's Verify Step by Step, 2023; Golovneva et al., ROSCOE, ICLR 2023
Turpin et al., Language Models Don't Always Say What They Think, NeurIPS 2023; Lanham et al., Measuring Faithfulness in Chain-of-Thought Reasoning, 2023
Hendrycks et al., Aligning AI with Shared Human Values (ETHICS), ICLR 2021; Jiang et al., Can Machines Learn Morality? The Delphi Experiment, 2021
Is your topic of interest not on this list? I am happy to supervise a project starting from your own ideas, as long as it's connected to LLMs / MLLMs or agents. If you are interested in this, provide the details of your idea alongside your statement of interest (see below).
If you are interested, send me:
A short statement of interest (1 page max): which project(s) got your attention? What are your interests? How are you going to apply them to this project? Please write this yourself (and not your favourite LLM); any generic-sounding statement will be discarded.
Your CV
Relevant transcripts (informal is fine)
Contact: alessio.galatolo@it.uu.se
Expression of interest: 10th of November
Project plan submission: 27th of November (17:00)
Projects and positions will be filled on a first-come, first-served basis. I encourage you to get in touch early.