Debmalya Mandal

Assistant Professor of Computer Science, University of Warwick

I study how AI systems should learn and decide when people, incentives and societies are part of the loop — from aligning models with human feedback to designing fair collective decisions.

I am a member of the Foundations of AI & ML group at Warwick DCS. Before Warwick, I was a postdoc at the Max Planck Institute for Software Systems with Goran Radanovic and at Columbia’s Data Science Institute with Daniel Hsu. I received my PhD in Computer Science from Harvard, advised by David C. Parkes.

Google Scholar CV Research statement GitHub

Research

AI and multi-agent systems, with an emphasis on theory and modelling and applications to societal problems.

Learning from human feedback

Making RLHF and preference-based alignment hold up when feedback is corrupted, strategic, shifted or multilingual.

See: Distributionally robust RLHF, Strategyproof RLHF

All papers on this topic

RL in strategic and changing worlds

Performative reinforcement learning, principal–agent problems and multi-agent RL, where the learner’s policy reshapes its own environment.

See: Performative RL, Stochastic principal–agent problems

All papers on this topic

Robustness

Learning guarantees that survive corrupted data, poisoning attacks and distribution shift, from offline and multi-agent RL to learning from human feedback.

See: Corruption robust RLHF, Corruption-robust zero-sum Markov games

All papers on this topic

Social choice and information elicitation

Voting with limited communication, surprisingly popular voting and peer prediction — getting good collective answers from people.

See: Efficient and thrifty voting, Surprisingly popular voting

All papers on this topic

Fairness and causality

Fairness guarantees beyond the training distribution, socially fair RL, and causal effect analysis in multi-agent systems.

See: Fairness beyond the training data, Agent-specific effects

All papers on this topic

News

All news

Sep 2026 Five papers accepted at NeurIPS 2026, including Inference-Time Nash Alignment and Meta-Learning Preferences for Multilingual LLM Alignment.
May 2026 Distributionally Robust RLHF accepted at ICML 2026.
Jan 2026 One paper accepted at ICLR 2026 and two at AISTATS 2026.
Oct 2025 Rafid Ameer Mahmud joins the group as a PhD student, co-supervised with Paolo Turrini.
Sep 2025 Two papers accepted at NeurIPS 2025: Strategyproof RLHF and Stochastic Principal-Agent Problems.

Group

Group page

Jiaying Lin
PhD student, since Oct 2024. [research topic]
Rafid Ameer Mahmud
PhD student, since Oct 2025. Co-supervised with Paolo Turrini

Selected publications

All publications

ICML

Distributionally Robust Reinforcement Learning with Human Feedback

Debmalya Mandal, Paulius Sasnauskas, and Goran Radanovic

Human feedback · Reinforcement learning · Robustness

In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026

arXiv Bib Code

@inproceedings{mandal2026drrlhf,
  title = {Distributionally Robust Reinforcement Learning with Human Feedback},
  author = {Mandal, Debmalya and Sasnauskas, Paulius and Radanovic, Goran},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year = {2026},
}

JAIR

Surprisingly Popular Voting Recovers Rankings, Surprisingly!

Hadi Hosseini, Debmalya Mandal, Nisarg Shah, and Kevin Shi

Social choice

Journal of Artificial Intelligence Research (JAIR). Supersedes the IJCAI 2021 version (PDF) , 2026

Bib HTML Code Website

@article{hosseini2026spvoting,
  title = {Surprisingly Popular Voting Recovers Rankings, Surprisingly!},
  author = {Hosseini, Hadi and Mandal, Debmalya and Shah, Nisarg and Shi, Kevin},
  journal = {Journal of Artificial Intelligence Research (JAIR)},
  year = {2026},
}

AISTATS

Corruption Robust Reinforcement Learning with Human Feedback

Debmalya Mandal, Andi Nika, Parameswaran Kamalaruban, Adish Singla, and Goran Radanovic

Human feedback · Reinforcement learning · Robustness

In Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS), 2025

Oral arXiv Bib

Oral presentation (top 2% of accepted papers).

@inproceedings{mandal2025crrlhf,
  title = {Corruption Robust Reinforcement Learning with Human Feedback},
  author = {Mandal, Debmalya and Nika, Andi and Kamalaruban, Parameswaran and Singla, Adish and Radanovic, Goran},
  booktitle = {Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS)},
  year = {2025},
}

ICML

Performative Reinforcement Learning

Debmalya Mandal, Stelios Triantafyllou, and Goran Radanovic

Reinforcement learning

In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023

arXiv Bib

@inproceedings{mandal2023performative,
  title = {Performative Reinforcement Learning},
  author = {Mandal, Debmalya and Triantafyllou, Stelios and Radanovic, Goran},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning (ICML)},
  year = {2023},
}

NeurIPS

Ensuring Fairness Beyond the Training Data

Debmalya Mandal, Samuel Deng, Suman Jana, Jeannette M. Wing, and Daniel Hsu

Fairness and causality · Robustness

In Advances in Neural Information Processing Systems (NeurIPS), 2020

Bib PDF Code

@inproceedings{mandal2020fairness,
  title = {Ensuring Fairness Beyond the Training Data},
  author = {Mandal, Debmalya and Deng, Samuel and Jana, Suman and Wing, Jeannette M. and Hsu, Daniel},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2020},
}

NeurIPS

Efficient and Thrifty Voting by Any Means Necessary

Debmalya Mandal, Ariel D. Procaccia, Nisarg Shah, and David P. Woodruff

Social choice

In Advances in Neural Information Processing Systems (NeurIPS), 2019

Oral Bib PDF Slides

Oral presentation (one of 36 out of 1428 accepted papers).

@inproceedings{mandal2019thrifty,
  title = {Efficient and Thrifty Voting by Any Means Necessary},
  author = {Mandal, Debmalya and Procaccia, Ariel D. and Shah, Nisarg and Woodruff, David P.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2019},
}

Service

Full list

Area chair: NeurIPS 2024–2026, ICML 2025–2026, IJCAI 2025
Top reviewer: NeurIPS 2019 and 2021, ICML 2020

MB 1.15, Mathematical Sciences Building
University of Warwick, Coventry CV4 7AL, UK

Email is the best way to reach me.