| CARVIEW |
About me
I am a 4th year PhD Candidate at KAUST AI Initiative, supervised by Prof. Jürgen Schmidhuber . I am interested in scalable methods that improve reasoning in neural networks. In practice, it means that I mostly work on Mixture-of-Experts, analyzing and improving transformer architecture, LLMs and (goal-conditioned / hierarchical) reinforcement learning algorithms.
I have published in top ML conferences such as NeurIPS, ACL, ICLR or ICRA. My publications have been distincted as oral presentations in ACL and ICLR. The list of all my publications can be found below.
I studied simultaneously Mathematics (BSc+MSc) and Computer Science (BSc) at the University of Warsaw. I wrote my Master's thesis on improving BERT's mathematical skills under the supervision of Mateusz Malinowski and Henryk Michalewski
I have 4 years of industry experience. Apart from my 2024 Amazon research internship, I worked for over 3 years as a Deep Learning engineer before starting my PhD. During that time I worked with some of the biggest Polish companies like Allegro, LPP or AmRest.
Apart from deep learning, I enjoy bicycle trips, playing chess, and playing some instruments: I play guitar, and I have started learning to play the piano.
Publications
-
Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing
NeurIPS 2025 (Efficient Reasoning Workshop)
Piotr Piękos, Róbert Csordás, Jürgen Schmidhuber
-
Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine
Wenyi Wang*, Piotr Piękos*, Li Nanbo, Firas Laakom, Yimeng Chen, Mateusz Ostaszewski, Mingchen Zhuge, Jürgen Schmidhuber
-
Hyperbolic Residual Quantization: Discrete Representations for Data with Latent Hierarchies
NeurIPS 2025 (NEGEL Workshop)
Piotr Piękos, Subhradeep Kayal, Alexandros Karatzoglou
-
PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors
NeurIPS 2025
Yimong Chen, Piotr Piȩkos, Mateusz Ostaszewski, Firas Laakom, Jürgen Schmidhuber
-
SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention
NeurIPS 2024
Róbert Csordás, Piotr Piękos, Kazuki Irie, Jürgen Schmidhuber
-
Fast and Precise: Adjusting Planning Horizon with Adaptive Subgoal Search
ICLR 2023OralMichał Zawalski, Michał Tyrolski, Konrad Czechowski, Tomasz Odrzygóźdź. Damian Stachura. Piotr Piękos, Yuhuai Wu, Łukasz Kuciński, Piotr Miłoś
-
Measuring and Improving BERT's Mathematical Abilities by Predicting the Order of Reasoning
ACL 2021OralPiotr Piękos, Henryk Michalewski, Mateusz Malinowski
-
Mindstorms in Natural Language-Based Societies of Mind
CVMJ 2025
Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Róbert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, Louis Kirsch, Bing Li, Guohao Li, Shuming Liu, Jinjie Mai, Piotr Piękos, Aditya Ramesh, Imanol Schlag, Weimin Shi, Aleksandar Stanić, Wenyi Wang, Yuhui Wang, Mengmeng Xu, Deng-Ping Fan, Bernard Ghanem, Jürgen Schmidhuber
-
Utilizing a Malfunctioning 3D Printer by Modeling Its Dynamics with Artificial Intelligence
ICRA 2024
Renzo Cabalero*, Piotr Piękos*, Eric Feron, Jürgen Schmidhuber
-
Efficient Value Propagation with the Compositional Optimality Equation
GCRL NeurIPS 2023 Workshop
Piotr Piękos, Aditya Ramesh, Francesco Faccio, Jürgen Schmidhuber
* - equal contribution
Experience
-
May 2024 - October 2024
Amazon
Applied Scientist Internship
Creating hierarchical discrete representations from dense vector representations used for representations of product in the Amazon catalog.Lead to a publication about Hyperbolic Residual Quantization apart from being implemented in the internal Amazon systems.
-
June 2023 - Aug 2023
IDSIA
Research Internship
Working on sample efficient Goal-Conditioned RL and improving the speed of the Transformer. The result of that were two publications, "Efficient Value Propagation with the Compositional Optimality Equation " and "SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention".
-
Sept 2022 - Present
KAUST
PhD Student (GPA 4.0/4.0)
Supervised by Prof. Jürgen Schmidhuber. I am working on generalization, sample efficiency in context of language processing and reinforcement learning.
-
Jan 2022 - Jul 2022
AWARElab, IMPAN (Institute of Mathematics, Polish Academy of Sciences)
Researcher
Working on AdaSubS. Implementing experiments on the sokoban environment
-
Oct 2021 - Jan 2022
Allegro.PL
Research Engineer
Researching neural search methods for improving search recommendations on the Allegro.pl platform
-
July 2017 - June 2020
ITmagination
ML/DL Engineer
In ITmagination I created DL/ML solutions to help other companies. Example projects are: Object detection system for evaluating influencers, recommender system or a system to help with estimating staff needs in physical shops.
Awards
Best Poster Award — "Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing"
Oral Presentation — "Fast and Precise: Adjusting Planning Horizon with Adaptive Subgoal Search"
Best Paper Award — "Mindstorms in Natural Language-Based Societies of Mind"
Oral Presentation — "Measuring and Improving BERT's Mathematical Abilities by Predicting the Order of Reasoning"
Best Poster Award — "Measuring and Improving BERT's Mathematical Abilities by Predicting the Order of Reasoning"
Measuring and Improving BERT's Mathematical Abilities by Predicting the Order of Reasoning
Piotr Piękos, Henryk Michalewski Mateusz Malinowski
Before the rise of chain-of-thought prompting in LLMs and following papers we investigated the impact of the rationales on the reasoning and demonstrated that it can be useful to improve reasoning. We design specific losses that can help model improve coherence when trained on rationales. We also found that our method not only improves the raw accuracy score on mathematical tasks, but also makes the model more reliable.
Our method is completely data-driven, and therefore easily can be used to improve any general purpose language model with a dataset of rationales.
More details, link to the paper and as video describing the project can be found on https://bert-math.github.io
Fast and Precise: Adjusting Planning Horizon with Adaptive Subgoal Search
Michał Zawalski, Michał Tyrolski, Konrad Czechowski, Tomasz Odrzygóźdź, Damian Stachura, Piotr Piękos, Yuhuai Wu, Łukasz Kuciński, Piotr Miłoś
In rich environments, some decisions need to be carefully thought through before making the decision, whereas for some it is easy to predict a state that's far away which we want to achieve. We utilize this fact in the context of hierarchical reinforcement learning and design an algorithm that adaptively selects horizon for choosing the next subgoal (state to reach) planning.
Our algorithm proves to operates significantly better than baselines, when compared on the same budget. Moreover, the algorithm demonstrates much better out of distribution generalization on INT environment, where it has to prove mathematical inequalities.
More details can be found on the project website
Project due to its success has also appeared in several polish popular news services, for example: