His research spans machine learning, reinforcement learning, self-supervised learning and large language models. In 2014, he introduced implicit exploration (IX), an approach to exploration in adversarial multi-armed bandit problems. At DeepMind, he led the team that developed Bootstrap Your Own Latent (BYOL), a self-supervised representation-learning method. In 2023, he and his co-authors received an ICML Outstanding Paper Award for work on learning in zero-sum imperfect-information games. In 2024, he co-authored Nash Learning from Human Feedback (NLHF), which formulates preference learning using a pairwise preference model and a Nash-equilibrium objective.
Michal Valko began his research career with an internship at Intel Labs in Silicon Valley (2009–2010). In 2011, he joined the French national research institute Inria, where he secured a permanent position as a research fellow in 2012 and subsequently began lecturing on machine learning as part of the MVA master’s programme at ENS Paris-Saclay.
He co-founded the Google DeepMind research laboratory in Paris, where he worked from 2019 to 2024. Following this, he joined Meta AI in Paris as a chief scientist and lead researcher focusing on reinforcement learning and the post-training optimization of foundation models. In 2025, he joined Isara Labs as a founding researcher.
Research
Valko's early research included graph-based methods for online semi-supervised learning. His 2011 doctoral dissertation studied approximate graph methods for settings in which data arrive in a stream and computation and storage become limiting.
In 2014, with Tomáš Kocák, Gergely Neu and Rémi Munos, Valko introduced implicit exploration for bandit problems with side observations. The method incorporates exploration through biased loss estimates, and their NeurIPS paper gave a near-optimal regret algorithm that does not require the observation system to be known before an action is selected. Implicit exploration was later included in Valko's 2016 habilitation thesis, Bandits on Graphs and Structures.
Another line of his research concerned the scalability of graph and kernel methods. In 2017, Daniele Calandriello, Alessandro Lazaric and Valko introduced SQUEAK, a single-pass kernel approximation algorithm that avoids constructing the full kernel matrix. The work addressed the quadratic time and storage needed to construct and store such matrices; its running time is linear in the number of samples, up to its dependence on the effective dimension. In 2018, with Calandriello, Lazaric and Ioannis Koutis, he applied distributed spectral sparsification to large-scale graph learning. Their method constructs a sparsified graph using near-linear memory and extends effective-resistance approaches associated with Daniel Spielman and Nikhil Srivastava to a regularized setting used in machine learning.
At DeepMind, Valko led the team that designed and published Bootstrap Your Own Latent (BYOL), a self-supervised representation-learning method presented at NeurIPS in 2020. BYOL uses an online and a target neural network, with the online network trained to predict the target network's representation of another augmented view of the same image. Unlike the contrastive methods with which it was compared, BYOL does not require negative example pairs.
In 2023, Valko, Côme Fiegel, Pierre Menard, Tadashi Kozuno, Rémi Munos and Vianney Perchet studied learning in zero-sum imperfect-information games through self-play with trajectory feedback. They established a problem-independent lower bound on the number of game trajectories required to learn near-optimal strategies and introduced two follow-the-regularized-leader algorithms: Balanced FTRL, which matches the lower bound when the information-set structure is known in advance, and Adaptive FTRL, which adapts its regularization without requiring that prior knowledge. The paper received an ICML Outstanding Paper Award.
In 2024, Valko co-authored Nash Learning from Human Feedback (NLHF), an alternative to standard reinforcement learning from human feedback. Rather than learning a scalar reward model from pairwise preferences and then optimizing that reward, NLHF models preferences between pairs of responses directly and seeks a policy that forms a Nash equilibrium of the preference model. The work introduced Nash-MD, a mirror-descent algorithm whose last iterate converges to a regularized Nash equilibrium, and evaluated the approach on text summarization.
Also in 2024, Valko was personally recruited by Mark Zuckerberg to join Meta's generative-AI group as a principal Llama scientist, after five years at Google DeepMind. Later accounts in Denník N and Forbes reported that the recruitment did not involve a conventional interview; Valko said that Meta made an offer within minutes and that the company was urgently recruiting reinforcement-learning expertise for Llama 3. At Meta, he worked on reinforcement learning and post-training for Llama 3 and co-authored The Llama 3 Herd of Models, the technical report describing the model family, including its 405-billion-parameter model and work on multilinguality, coding, reasoning, tool use and multimodal capabilities.
Awards and recognition
Valko received Inria's Prime d'excellence scientifique for 2014–2017 and again for 2018–2021. In 2023, he and his co-authors received an ICML Outstanding Paper Award for Adapting to game trees in zero-sum imperfect information games. In 2024, the Slovak edition of Forbes included Valko in its Changemakers list, in a category covering artificial intelligence and large language models.
Scientific community and public engagement
Mathematics education
Valko was among the organizers of the Korešpondenčný matematický seminár (KMS) in the 2002/03 school year, alongside Richard Kollár and Alexander Erdélyi. KMS brought together the Stredoslovenský korešpondenčný matematický seminár (SKMS), the Bratislavský korešpondenčný matematický seminár (BKMS), and Bratislava-based organizers of STROM. Its organizers described the unification as a way to broaden participation, with separate categories for beginning and more experienced students, while continuing to operate on a voluntary, unpaid basis. The camp archive lists Valko among the leaders of KMS's BETA residential mathematics camps in 2003 and 2004.
Scientific organization, publishing and evaluation
His programme committee service has also included the 2019 Conference on Learning Theory and the 2021 workshop Graph Neural Networks and Systems (GNNSys).
At Inria, Valko was an elected member of the institute's evaluation committee, serving on researcher-recruitment, promotion and scientific-excellence-award committees. His responsibilities included national junior-researcher selection and a working group on deontological ethics; he also served on a faculty-recruitment committee at CMLA, ENS Paris-Saclay. In 2018, he participated in the working group for the creation of the RandOpt research team and the national committee for secondments at Inria.
Valko was programme co-chair of the 2018 CNRS summer school on Networks, Graphs, and Machine Learning (RESCOM), and organizing co-chair of the Optimizing Human Learning workshop that year. With Laura Toni, he co-organized the June 2019 UCL–French Embassy workshop on graph-based machine learning and sequential decision-making in London.
As a member of Inria's SequeL team, he helped organize the 2019 Reinforcement Learning Summer School in Lille. The two-week school combined lectures and practical sessions on reinforcement learning and bandit algorithms, selecting 110 participants from about 300 applications. He also served on the programme committee of the 2015 European Workshop on Reinforcement Learning (EWRL). SequeL hosted EWRL in Lille in 2008, 2015 and 2018; its successor team, Scool, was announced as the organizer of the October 2026 Lille edition.
Teaching and research mentorship
Valko has taught Graphs in Machine Learning in the MVA master's programme and co-supervised doctoral research on sequential decision-making, bandit algorithms and graph learning. His doctoral-examination service includes thesis juries at Toulouse III University in 2017 and 2018, and serving as the opponent at a doctoral defence at Chalmers University of Technology in 2019.
His involvement in the Eastern European Machine Learning Summer School (EEML) has included mentoring, lecturing and facilitating journal clubs between 2020 and 2022, and co-organising the 2023 edition in Košice. EEML seeks to strengthen machine-learning communities in Eastern Europe and offers need-based financial support for participation.
Valko serves on the advisory board of GHOST Day, an applied-machine-learning conference in Poznań. He has also lectured at DataFest Yerevan in 2023 and 2025, and at Yerevan State University's seventh Mathematics and Applications summer school, held in Tsaghkadzor, Armenia, in 2025.
The programme of the 2023 AI for Africa TechInnovDays in Morocco included two sessions by Valko on representation learning and reinforcement learning.
Scientific and venture advisory work
Valko's startup mentoring has included companies applying artificial intelligence to healthcare, materials discovery and mineral exploration. In a 2025 interview with Forbes, he discussed his work with Sword Health, Entalpic and Lithosquare, and his angel investment in Entalpic. He has also advised the materials-science company Altrove and the healthcare-technology company Tortus.
His work with venture-capital firms has included venture-partner roles at KAYA VC, Formula VC and Sparkle Ventures.
Public engagement and volunteering
Valko is an ambassador of the ESET Science Award. He has also volunteered with senior citizens, an activity described by both the award's organizers and Forbes Slovensko.
In 2025, he represented the technology sector in a panel on responsible artificial intelligence at the BratislavAI Forum's High-Level Summit on AI, organized by Slovakia's Ministry of Education in cooperation with the OECD. The panel, chaired by education minister Tomáš Drucker, included representatives of the OECD, United Nations, UNESCO and European Commission.
A draft programme published by the National Bank of Slovakia also listed Valko and Drucker for the summit's closing remarks.
In 2026, he won the AI Social Impact of the Year category in the ASAI AI Personality of the Year awards.
References
123"2023 Awards". International Conference on Machine Learning. Retrieved 26 August 2026.
↑Calandriello, Daniele; Lazaric, Alessandro; Valko, Michal (2017). Distributed Adaptive Sampling for Kernel Matrix Approximation. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics. Vol.54. PMLR. pp.1421–1429.
↑Munos, Rémi; Valko, Michal; etal. (2024). Nash Learning from Human Feedback. Proceedings of the 41st International Conference on Machine Learning. Vol.235. PMLR. pp.36743–36768.
