News and updates

Research news, talks, papers, and group milestones

Recent updates from my research group, collaborations, publications, invited talks, and media coverage.

May 29, 2026 Latest update

ICML 2026 Accepted Papers

I am pleased to share that five papers have been accepted to ICML 2026 in Seoul, together with four papers at ICML 2026 workshops. I am grateful to my students and collaborators for their work on these projects.

Accepted conference papers:
1. Zero-Shot Off-Policy Learning, with Arip Asadulaev, Maksim Bobrin, Salem Lahlou, Dmitry Dylov, Fakhri Karray, and Martin Takáč.
2. WaveSSM: Multiscale State-Space Models for Non-stationary Signal Attention, with Ruben Solozabal, Velibor Bojkovic, Hilal Alquabeh, Klea Ziu, Kentaro Inui, and Martin Takáč.
3. From Optimization to Generalization under Heavy-Tailed Data: The Role of Gradient Clipping, with Aleksandr Shestakov, Martin Takáč, and Eduard Gorbunov.
4. Your Latent Reasoning is Secretly Policy Improvement Operator, with Arip Asadulaev, Rayan Banerjee, Fakhri Karray, and Martin Takáč.
5. CoRe: Collaborative Reasoning via Cross Teaching, with Kshitij Mishra, Mirat Aubakirov, Martin Takáč, Nils Lukas, and Salem Lahlou.

Workshop papers:
1. ψDAG: Projected Stochastic Approximation Iteration for DAG Structure Learning, with Klea Ziu, Slavomír Hanzely, Loka Li, Kun Zhang, Martin Takáč, and Dmitry Kamzolov.
2. Breaking the Martingale Curse: Multi-Agent Debate via Asymmetric Cognitive Potential Energy, with Yuhan Liu, Juntian Zhang, Yichen Wu, Martin Takáč, Salem Lahlou, Xiuying Chen, and Nils Lukas.
3. Decision-Making from Offline Datasets to Online Adaptation: Black-Box Optimization to Reinforcement Learning, with Alexey Zemtsov, Maksim Bobrin, Alexander Nikulin, Dmitry V. Dylov, Fakhri Karray, Vladislav Kurenkov, Martin Takáč, and Arip Asadulaev.
4. Revisiting the Form of Attention with Positional Encoding for Molecular Structures, with Yusei Ito, Aidar Alimbayev, Klea Ziu, Deepak Kumar, Kanta Ono, and Martin Takáč.
Archive

Earlier updates

Publications

Abdulla Jasem Almansoori Defends PhD Thesis


My PhD student Abdulla Jasem Almansoori defended his thesis, Adapters for Collaborative Learning: Personalization and Structure-Aware Optimization.

His dissertation studies parameter-efficient adaptation for collaborative learning, focusing on federated learning under heterogeneous client data and structure-aware optimization of lightweight adapters. The committee included Samuel Horvath, Praneeth Vepakomma, Martin Jaggi, and me.

Byzantine-Robust Federated Learning


MBZUAI featured our work on Bant, a Byzantine-robust approach to federated learning. The method combines a trusted trial function with trust scores to filter corrupted updates, with variants evaluated on image classification, ECG abnormality detection, and recommender system benchmarks.

The study was presented orally at the 40th Annual AAAI Conference on Artificial Intelligence in Singapore, with co-authors from MBZUAI and other institutions, including me. check the full article here.

Artem Agafonov Defended His PhD Thesis


My PhD student Artem Agafonov defended his PhD thesis, "Globally Convergent Quasi-Newton-Type Methods and Their Applications in Convex Minimization, Federated Learning and Variational Inequalities."

The committee members were Martin Takac, Samuel Horvath, Zhiqiang Shen, Junpei Komiyama, and Sebastian Stich.

MedNNS for Medical Imaging Model Selection


An MBZUAI news article describes MedNNS, a framework for selecting both neural network architectures and pretrained weights for new medical imaging datasets. The approach reframes model selection as a retrieval problem, using dataset and model embeddings to recommend candidates before expensive training begins.

The article notes that MedNNS is planned for presentation at MICCAI 2025 in South Korea and reports gains in speed, accuracy, and stability across MedMNIST tasks, while also discussing limitations for out-of-distribution datasets and future directions beyond classification. check the full article here.

ReCall and the reversal curse in LLMs


I co-authored a study on ReCall, a two-step prompting strategy that uses self-referencing causal cycles in text to help autoregressive language models retrieve information across sequence order.

The work examines the reversal curse in LLMs and will be presented at the 63rd Annual Meeting of the Association for Computational Linguistics in Vienna. check the full article here.

My first MBZUAI PhD Student Defense


On October 28, 2024, my first PhD student at MBZUAI defended his thesis, Machine Learning for Combinatorial Optimization. The committee included Larry Snyder, Karthik Nandakumar, Bin Gu, Zhiqiang Xu, and me.

The thesis examined machine learning approaches for combinatorial optimization, including reinforcement learning for stochastic vehicle routing and scheduling, curriculum learning, and prompting strategies for large language models.

Keynote at AAAI 2024 Workshop on Artificial Intelligence for Operations Research

I gave a keynote, "Beyond Conventional Boundaries, RL's Leap in Solving Optimization Problems," at the AAAI 2024 Workshop on Artificial Intelligence for Operations Research in Vancouver, Canada. AAAI 2024 Workshop on Artificial Intelligence for Operations Research.

Moving to MBZUAI


I am excited to share that starting in Fall 2021 I will move to MBZUAI in Abu Dhabi, UAE. I am grateful to my colleagues at Lehigh, from where I moved, for their support.

I have multiple openings in my research lab for those interested in joining me. www.MBZUAI.ac.ae.

Majid Jahani successfully defended his Ph.D. dissertation!


My PhD student Majid Jahani defended his Ph.D. dissertation.
His doctoral committee members included
Prof. Luis Nunes Vicente, Prof. Frank E. Curtis, Prof. Katya Scheinberg, Prof. Aryan Mokhtari, and Prof. Albert S. Berahas and of couse me.

I was invited speaker at the Robot Learning Workshop


Video from my presentation "Image Classification using Deep Reinforcement Learning" is available. here.

I had a brief discussion in Slovak National Radio


I had a brief discussion in Slovak National Radio. here.

I was a guest in Slovak national TV



Check 43:19 here.

I was a guest in discussion in TA3 (Slovak national TV)


I had an opportunity to discuss the future of AI with Prof. Farkaš and Vladimír Šucha, Director-General of the Joint Research Centre, the European Commission's science and knowledge service. Vladimír Šucha. here.

My interview in Slovak national TV news


Check 21:01 in the TV archive; more information is also available on Lehigh's website. here. here.

Youtube video about our recent work.


A video explaining our paper "Multi-Agent Image Classification via Reinforcement Learning" by Hossein K. Mousavi, Mohammadreza Nazari, Martin Takáč, and Nader Motee is available. here.

Lam won Dissertation Award for his PhD thesis!

I am extremely happy to inform that Lam has won the 2019 P.C. Rossin College of Engineering and Applied Science Elizabeth V. Stout Dissertation Award.
The title of his thesis was "A Service System with On-Demand Agents, Stochastic Gradient Algorithms and the SARAH Algorithm".

I was the recipient of the 2019 edition of the Richard P. Vinci Award for Educational Excellence of the RCEAS.

The prize was awarded today at the Rossin College Award Ceremony:
"Educational Excellence - This award recognizes a Rossin College faculty member who has demonstrated effective teaching AND/OR
enhanced the student learning experience by introducing innovative teaching methods into the classroom AND shows an outstanding commitment to the success of their students." https://engineering.lehigh.edu/news/article/rossin-awards-celebrate-excellence-across-lehighs-engineering-college.

New paper: Quasi-Newton Methods for Deep Learning: Forget the Past, Just Sample

With Albert S. Berahas and Majid Jahani
we posted a new paper about Sampled Quasi-Newton method for training Deep Neural Networks! arxiv.