Hi,
I'm Marek Kadlčík,
an AI researcher.

I'm interested in Neural Networks.

Publications

Language Models Learn Universal Representations of Numbers and Here's Why You Should Care

ACL (Main track), 2026

Štefánik*, Mickus*, Kadlčík*, Højer, Spiegel, Vázquez, Sinha, Kuchař, Mondorf, Stenetorp

We describe how universal and consistent the number representations are in pretrained LMs - across models, layers and contexts, and design a novel more accurate constrained probe.

Pre-trained Language Models Learn Remarkably Accurate Representations of Numbers

EMNLP (Main track), 2025

Marek Kadlčík, Michal Štefánik, Timothee Mickus, Michal Spiegel, Josef Kuchař

We demonstrate that pretrained LMs represent numbers with remarkable accuracy by designing a probing method that decodes representations with near-perfect accuracy.

Can Out-of-Distribution Evaluations Uncover Reliance on Prediction Shortcuts?

EMNLP (Findings), 2025

Michal Štefánik, Timothee Mickus, Michal Spiegel, Marek Kadlčík, Josef Kuchař

We find that cross-dataset evaluation, a common method for assessing models' robustness, correlates poorly with targeted evaluation of models' reliance on reasoning shortcuts.

Towards the Roots of the Negation Problem: A Multilingual NLI Dataset and Model Scaling Analysis

EMNLP (Findings), 2025

Tereza Vrabcová, Marek Kadlčík, Petr Sojka, Michal Štefánik, Michal Spiegel

We show that logical reasoning is a language-sensitive skill. We create a parallel NLI benchmark with paired hypotheses for evaluating the causal effect of negation on reasoning in logical entailment.

Self-Training Language Models for Arithmetic Reasoning

EMNLP (Findings), 2024

Marek Kadlčík*, Michal Štefánik*

We find that LMs can improve by autonomously attempting to solve reasoning problems, even when they know the problems from prior SFT. Preference optimization gives more robust models than rejection-sampling finetuning.

Concept-aware Data Construction Improves In-context Learning of Language Models

ACL (Findings), 2024

Michal Štefánik, Marek Kadlčík

Exposing few-shot language models to relevant demonstrations during training is more data-efficient and reduces over-reliance on semantic priors.

Calc-X and Calcformers: Empowering Arithmetical Chain-of-Thought through Interaction with Symbolic Systems

EMNLP (Main track), 2023

Marek Kadlčík, Michal Štefánik, Ondřej Sotolář, Vlastimil Martinek

Accuracy in solving math problems can be doubled by training the model to interact with an external calculator during inference.

Soft Alignment Objectives for Robust Adaptation of Language Generation

ACL (Main track), 2023

Michal Štefánik, Marek Kadlčík, Petr Sojka

We propose an alternative loss to cross-entropy with a soft target distribution that exhibits superior out-of-domain robustness when fine-tuning LMs.

Can In-Context Learners Learn a Reasoning Concept from Demonstrations?

ACL (NLRSE workshop) - Best Paper award, 2023

Michal Štefánik, Marek Kadlčík

We study the limitations of language models' understanding of functional relations in in-context few-shot learning.

A Whisper Transformer for Audio Captioning Trained with Synthetic Captions and Transfer Learning

DCASE Challenge submission (technical report), 2023

Marek Kadlčík*, Adam Hájek*, Jürgen Kieslich, Radosław Winiecki

Exploring pretraining audio captioning models with synthetic descriptions. Our models were downloaded 100,000+ times in total.

Resources and Few-shot Learners for In-context Learning in Slavic Languages

EACL (Slavic NLP workshop) - Best Paper Award, 2023

Michal Štefánik, Marek Kadlčík, Piotr Gramacki, Petr Sojka

We facilitate the training of in-context learners in underrepresented languages by transforming supervised datasets into an instructional format.

Experience

Quantitative Trading R&D, p-hat fund

Sep 2025 - Jun 2026

I was designing and training reinforcement learning (RL) strategies for automatic trading on cryptocurrency exchanges, including risk-sensitive RL optimization. I was also responsible for developing the core architecture of the entire RL training pipeline.

Research Internship, National Institute of Informatics Japan

Feb 2026 - Mar 2026

I attended a short research visit in NII Japan, contributing to preparation of the design and software infrastructure of AIMO Interpretability Challenge for NeurIPS 2026.

Machine Learning R&D, Gauss Algorithmic

Oct 2021 - Mar 2023

I worked on designing and applying state-of-the-art deep learning methods to automate email customer support (natural language processing) and to develop a single-camera 3D hand motion capture system (computer vision).

Assistant Teacher, Masaryk University

Sep 2019 - Dec 2021

Teaching helped me learn to communicate effectively and talk to an audience. My courses:
  • Neural Networks — course assistant (2021)
  • Introduction to Machine Learning — tutor (2020, 2021)
  • Non-imperative Programming — tutor (2019)

App Developer, Fenomen Multimedia

2016 - 2018

During my high school years, I worked part-time developing from scratch and deploying two mobile apps (Kotlin/Java and React), one reaching 10,000+ users.

Awards and Achievements

Spotlight Paper

2026

Received for paper LMs Learn Universal Representations of Numbers and Why You Should Care at the ICML Mechinterp Workshop.

Top 10 Reviewers

2026

I was recognized for an outstanding number of high-quality reviews according to ACL Rolling Review Statistics (Jan 2026).

Rector's Award

2025

I was honored with the Outstanding Student Award at Masaryk University for my research contributions in the area of language models and reasoning.

Outstanding Reviewer

2024

Received at EMNLP 2024 for high-quality reviews.

Dean's Award

2024

Received for my master's thesis, Improving Arithmetical Reasoning of Language Models for its contributions to the field of natural language processing.

Best Paper Award

2023

Received for paper Can In-context Learners Learn a Reasoning Concept from Demonstrations? at ACL Natural Language Reasoning Workshop.

100k+ Model Downloads

2023

My Whisper audio-captioning models have been downloaded over 100,000 times on Hugging Face.

Best Paper Award

2023

Received for paper Resources and Few-shot Learners for In-context Learning in Slavic Languages at EACL Slavic NLP Workshop.

Dean's Award

2021

Received for my bachelor's thesis, Risk-Sensitive Reinforcement Learning for its high quality.

Education

Masaryk University, PhD program

2024 - ongoing

I research language model reasoning, interpretability, and learning from feedback.

Masaryk University, AI & ML master's program

2021 - 2024

My thesis, Improving Arithmetical Reasoning of Language Models, won the Dean's Outstanding Thesis Award and led to two publications at international conferences.

Johannes Kepler University, Exchange program

2022 (and 2020)

I attended an exchange program at JKU to learn about deep learning, reinforcement learning, probabilistic modeling, and machine learning for natural language processing, computer vision, and audio processing.

Masaryk University, AI & NLP bachelor's program

2018 - 2021

My thesis, Risk Sensitive Reinforcement Learning, overviews the current state of the art of risk management in reinforcement learning. It won the Dean's Outstanding Thesis Award.

English - C1

2018

I passed the Cambridge exam for certification of English at the C1 level - advanced proficiency.

Socials

Google Scholar

Marek Kadlčík

GitHub

@prompteus

LinkedIn

@prompteus