I study the limitations of language models, especially reasoning models, and how to make their deployment more reliable. My work spans uncertainty estimation, capability evaluation, chain-of-thought monitoring, and interpretability. I am a Research Fellow at EXAI and the University of Tartu, where I completed my PhD in Artificial Intelligence.
Preprint, 2026
We show that combining verbalized confidence and self-consistency with two samples improves AUROC by up to 12 points on average, outperforming either signal alone even at larger sampling budgets.
PhD thesis, University of Tartu, 2024
This thesis examines how transformer models organize languages and domains internally, including where shared representations emerge and where they remain fragmented.
AACL, 2022
Oral presentation.
We introduce Average Neuron-Wise Correlation (ANC), reducing the required neuron-wise comparisons from O(N²) to O(N), and show that multilingual models first align languages before specializing for prediction.
Baltic Journal of Modern Computing, 2022
We resolve conflicting findings about multilingual representations and show that most languages, including low-resource Baltic languages, share a cross-lingual space, while a few remain separate.
WMT, 2021
* Equal contribution.
We show that translation models recover latent data domains, then reuse their representations for unsupervised adaptation, avoiding a separate language model and improving in-browser English-to-Estonian translation by up to 1.6 BLEU.
Preprint, 2019
We use a single multilingual translation model for grammatical error correction and style transfer, without task-specific training data.
WMT, 2018
We propose compositional phrase embeddings for phrase-based unsupervised machine translation.
