Maksym Del

AI researcher focused on making reasoning language models more reliable. I work on 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.

Portrait of Maksym Del

Research

Sampling-efficiency curves for confidence estimation across reasoning models and benchmarks

How Uncertainty Estimation Scales with Sampling in Reasoning Models

Maksym Del, Markus Kängsepp, Marharyta Domnich, Ardi Tampuu, Lisa Yankovskaya, Meelis Kull, Mark Fishel

Preprint, 2026

arXiv

We show that 2 to 4 reasoning traces capture most of the benefit of sampling for black-box confidence estimation in reasoning models. This low sampling budget offsets the high cost of reasoning traces, establishing sampling as the practical default for confidence estimation.

Overview of an LLM-based pipeline for artificial error generation and grammatical error correction

To Err Is Human, but Llamas Can Learn It Too

Agnes Luhtaru*, Taido Purason*, Martin Vainikko, Maksym Del, Mark Fishel

Findings of EMNLP, 2024

arXiv paper

* Equal contribution.

We show how to use large language models effectively for artificial error generation and grammatical error correction.

Cover of Maksym Del's PhD thesis on multilingual and multi-domain transformer representations

Multilingual and Multi-Domain Representational Patterns Across Transformer-Based Models

Maksym Del

PhD thesis, University of Tartu, 2024

paper

This thesis studies how transformer models organize multilingual and multi-domain information in their internal representations.

Illustration for the True Detective abductive-reasoning benchmark

True Detective: A Deep Abductive Reasoning Benchmark Undoable for GPT-3 and Challenging for GPT-4

Maksym Del, Mark Fishel

*SEM, 2023

paper code

We introduce 191 short-form detective puzzles for evaluating abductive reasoning and find a 42-percentage-point accuracy gap between GPT-4 and top human solvers.

Layer-wise cross-lingual similarity patterns in multilingual language models

Cross-lingual Similarity of Multilingual Representations Revisited

Maksym Del, Mark Fishel

AACL, 2022

paper code

Oral presentation.

We introduce Average Neuron-Wise Correlation, reducing cross-lingual representation comparison from O(N^2) to O(N), and demonstrate consistent middle-layer interlingual structure across multilingual models.

Cross-lingual sentence representation similarities for Baltic and other languages

Similarity of Sentence Representations in Multilingual LMs: Resolving Conflicting Literature and a Case Study of Baltic Languages

Maksym Del, Mark Fishel

Baltic Journal of Modern Computing, 2022

paper code

We reconcile conflicting findings about cross-lingual structure in multilingual language models and examine the representation of Baltic languages.

Visualization of domain structure in neural machine translation representations

Translation Transformers Rediscover Inherent Data Domains

Maksym Del*, Elizaveta Korotkova*, Mark Fishel

WMT, 2021

paper code

* Equal contribution.

We show that translation transformers separate inherent data domains internally and use this structure for unsupervised adaptation, improving in-browser English-to-Estonian translation by up to 1.6 BLEU.

Zero-shot monolingual translation setup for grammatical error correction and style transfer

Grammatical Error Correction and Style Transfer via Zero-shot Monolingual Translation

Elizaveta Korotkova, Agnes Luhtaru, Maksym Del, Krista Liin, Daiga Deksne, Mark Fishel

Preprint, 2019

arXiv

We perform grammatical error correction and style transfer with a single multilingual translation model, without task-specific training.

Phrase-based unsupervised machine translation with compositional phrase embeddings

Phrase-based Unsupervised Machine Translation with Compositional Phrase Embeddings

Maksym Del, Andre Tattar, Mark Fishel

WMT, 2018

paper

We propose compositional phrase embeddings for phrase-based unsupervised machine translation.

Architecture and results of the C-3MA neural machine translation systems

C-3MA: Tartu-Riga-Zurich Translation Systems for WMT17

Matīss Rikters, Chantal Amrhein, Maksym Del, Mark Fishel

WMT, 2017

paper code

We describe the University of Latvia, University of Zurich, and University of Tartu neural machine translation systems submitted to the WMT17 shared task.