Title:	Scalable Learning and Optimization for Graph-Structured Data and Reliable LLMs
DNr:	Berzelius-2026-267
Project Type:	LiU Berzelius
Principal Investigator:	Aristides Gionis <argioni@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2026-10-08 – 2027-05-01
Classification:	10201
Keywords:

Abstract

Our group develops scalable machine-learning and optimization methods for graphs, hypergraphs, and large language models. Over the next 6 months, the proposed project will use Berzelius to support four related research directions requiring large-scale experiments, baseline comparisons, scalability evaluations, hyperparameter tuning, parameter-sensitivity analyses, ablation studies, and repeated runs for reliable evaluation. First, we will conduct large-scale empirical benchmarking of new theoretical algorithms for consistent clustering, including but not limited to k-means, correlation clustering, and related graph optimization problems. We will also extend our graph-clustering research to hypergraphs and evaluate new CPU-based algorithms on large datasets. Second, we will evaluate a new graph-rewiring method on large datasets with long-range dependencies to determine whether it can mitigate over-squashing in graph neural networks. These experiments will require GPU-based model training across multiple datasets and random seeds, as well as CPU resources for preprocessing and graph construction. Third, we will study LLM capability and reliability by developing and comparing training optimizers, including Muon and spectrum-based methods; designing training objectives; and measuring uncertainty using models’ internal signals during training and inference. Experiments will cover models ranging from fewer than 100 million parameters to models with one billion parameters or more. Finally, we will investigate Graph Foundation Models and improve the efficiency of long-context LLM inference, with particular attention to KV-cache utilization. This work will require model training, fine-tuning, inference experiments, hyperparameter tuning, and systematic ablation studies.