Maksym Del

I am a postdoc at the Estonian Center of AI Excellence and Natural Language Processing group at the University of Tartu.

Currently, I am working on improving practical reliability of frontier LLMs (uncertainty, CoT monitoring) and evaluating thier capabilites.

During my PhD I worked on mechanistic understanding of cross-lingual generalization and LLM evals.

My research interests include ensuring AI trustworthiness and safety more broadly by developing moodel control, monitoring, and interpretability measures.

Google Scholar  /  LinkedIn  /  GitHub  /  Twitter (X)  /  Email

Research

My research interests include ensuring AI trustworthiness and safety more broadly by developing moodel control, monitoring, robustness, and interpretability measures.

To Err Is Human, but Llamas Can Learn It Too

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

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

We show that GPT-4 cannot reliably solve short-form detective puzzles, even when given a chain of thought reasoning trace hinting at the correct answer.

Cross-lingual Similarity of Multilingual Representations Revisited

We demonstrate the universality of the internal cross-lingual structure across multilingual language models.

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

We address confsuion in prior work regarding the internal cross-slingual structure in multilingual language models.

Translation Transformers Rediscover Inherent Data Domains

We show that translation transformers keep internal domain representaions apart.

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

We present an approach that does both grammatical error correction and style transfer with a single multilingual translation model out-of-the-box.

Phrase-based Unsupervised Machine Translation with Compositional Phrase Embeddings

We propose compositional phrase embeddings for unsupervised machiene translation.

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

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

Design and source code from Jon Barron's website