About Publications Projects Contact
Dario Vajda

Dario Vajda

Undergraduate researcher at UL FRI working on making LLMs natively reason over graphs.

I am a final-year undergraduate student in the Mathematics and Computer Science programme at the Faculty of Computer and Information Science, University of Ljubljana, where I hold the highest GPA in my year. I have accepted an offer to pursue the MPhil in Advanced Computer Science at the University of Cambridge, starting in the next academic year.

Research Interests

I am passionate about bridging graph-structured reasoning and large language models — teaching LLMs to natively think over graphs.

Publications

Under Review Preprint

Teaching LLMs to See Graphs: Unifying Text and Structural Reasoning

D. Vajda

Introduced a novel architecture that enables pretrained LLMs to natively process graph topologies by injecting graph-aware attention biases, adding only 0.015% additional parameters. A 1B-parameter GTLM matched or exceeded 7B-parameter state-of-the-art models on Text-Attributed Graph benchmarks.

arXiv:2605.10247
Preprint · 2026

Building a Strong Instruction Language Model for a Less-Resourced Language

D. Vreš, T. Arčon, T. Petrič, D. Vajda, M. Robnik-Šikonja, I. L. Bajec

Pretrained and instruction-tuned GaMS3, a 12B-parameter LLM for Slovenian, covering data curation, training, and evaluation across diverse benchmarks.

arXiv:2603.01691
LUHME Workshop · ACL 2025

Improving LLMs for Machine Translation Using Synthetic Preference Data

D. Vajda, D. Vreš, M. Robnik-Šikonja

Fine-tuned an LLM for Slovenian machine translation using Direct Preference Optimization on programmatically generated synthetic preference data, outperforming both models used in dataset construction.

ACL Anthology

Projects

Sidenote

Local-first app for reading, annotating, and analyzing research papers with AI agents.

GitHub

genmoji

Trained and compared an Autoregressive model and a Diffusion Probabilistic model to generate emojis.

GitHub

miniGPT

GPT reimplemented from scratch to understand transformer architecture at a low level.

GitHub

Contact

I am always happy to discuss research, collaborations, or opportunities.

vajdadario [at] gmail [dot] com

Download CV