Reinforcement Learning Researcher
PhD Candidate in Computer Science, Harvard University
Data to Actionable Knowledge Lab · advised by Finale Doshi-Velez
I’m a final-year PhD candidate in Computer Science at Harvard University (cross-registered at MIT), advised by Finale Doshi-Velez in the Data to Actionable Knowledge Lab. My research focuses on developing data-efficient reinforcement learning algorithms that extract insights from offline healthcare datasets to assist clinicians. I’m the inaugural recipient of the Yiming Zhang (Jump Trading) Graduate Fellowship in Artificial Intelligence, awarded for exceptional research potential in reinforcement learning. Through industry internships, I built an evaluation framework for AI agents at Waymo, forecast equity covariance matrices at Voloridge Investment Management, and applied RL to battery management systems (BMS) at Schneider Electric. Before Harvard, I earned an Honors degree in Statistics with a minor in Computer Science from Concordia University in Montreal, graduating with a 4.27/4.30 GPA (top 1%).
Introduces strategic link scores, which measure how much an earlier action depends on a planned follow-up. They are used to explain black-box RL agents and to make decision-support recommendations more robust, across maze, chess, healthcare, traffic, and crisis-negotiation environments.
@article{hyk2026strategicallylinked,
title = {Strategically-Linked Decisions in Long-Term Planning and Reinforcement Learning},
author = {H{\"u}y{\"u}k, Alihan and Raedler, Jonas B and Benac, Leo and Doshi-Velez, Finale},
journal = {Reinforcement Learning Journal},
volume = {7},
year = {2026}
}
Learns environment dynamics from offline, near-optimal expert trajectories by turning the expert’s choices into constraints on a Bayesian posterior. This improves downstream decisions in synthetic tasks and in ICU management of hypotension.
@inproceedings{pmlr-v300-benac26a,
title = {Bayesian Inverse Transition Learning: Learning Dynamics from Near-Optimal Trajectories},
author = {Benac, Leo and Sharma, Abhishek and Parbhoo, Sonali and Doshi-Velez, Finale},
booktitle = {Proceedings of The 29th International Conference on Artificial Intelligence and Statistics},
pages = {2242--2250},
year = {2026},
volume = {300},
series = {Proceedings of Machine Learning Research},
publisher = {PMLR},
url = {https://proceedings.mlr.press/v300/benac26a.html}
}
Decision Points RL (DPRL) improves on a behavior policy only at frequently visited state–action pairs. Its safety guarantees depend on those visit counts rather than on the size of the state–action space, and it improves both safety and performance on synthetic and real-world tasks.
@inproceedings{pmlr-v258-sharma25a,
title = {Decision-Point Guided Safe Policy Improvement},
author = {Sharma, Abhishek and Benac, Leo and Parbhoo, Sonali and Doshi-Velez, Finale},
booktitle = {Proceedings of The 28th International Conference on Artificial Intelligence and Statistics},
pages = {2935--2943},
year = {2025},
volume = {258},
series = {Proceedings of Machine Learning Research},
publisher = {PMLR},
url = {https://proceedings.mlr.press/v258/sharma25a.html}
}
A two-stage inverse RL approach that recovers clinical priorities by pruning ICU trajectories that deviate from clinician consensus. The benefit of pruning varies across diseases and demographic groups.
@article{bovenzi2025pruning,
title = {Pruning the Path to Optimal Care: Identifying Systematically Suboptimal Medical Decision-Making with Inverse Reinforcement Learning},
author = {Bovenzi, Inko and Carmel, Adi and Hu, Michael and Hurwitz, Rebecca and McBride, Fiona and Benac, Leo and Tello Ayala, Jos{\'e} Roberto and Doshi-Velez, Finale},
journal = {AMIA Annual Symposium Proceedings},
volume = {2024},
pages = {202--211},
year = {2025},
pmid = {40417587},
url = {https://pmc.ncbi.nlm.nih.gov/articles/PMC12099440/}
}
Recorded decision data, such as ICU treatment records, often lacks information the decision-maker actually saw, which can make good decisions look suboptimal to inverse RL. This work estimates how much may be missing by finding the smallest observation changes under which the expert’s actions look optimal.
@misc{benac2026quantifyingpotentialobservationmissingness,
title = {Quantifying Potential Observation Missingness in Inverse Reinforcement Learning},
author = {Leo Benac and Abhishek Sharma and Alihan Huyuk and Finale Doshi-Velez},
year = {2026},
eprint = {2605.12831},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2605.12831}
}
A benchmark for multi-party negotiations in which agreement forms through a series of binding, partial commitments. It combines a configurable game generator, game instances grounded in a real climate-negotiation exercise, and baseline solvers. No single solver wins in every regime.
@misc{benac2026benchmarkmultipartynegotiationgames,
title = {A Benchmark for Multi-Party Negotiation Games from Real Negotiation Data},
author = {Leo Benac and Jonas Raedler and Zilin Ma and Finale Doshi-Velez},
year = {2026},
eprint = {2603.14066},
archivePrefix = {arXiv},
primaryClass = {cs.MA},
url = {https://arxiv.org/abs/2603.14066}
}
Two estimators of conditional value-at-risk (CVaR) for bandits whose losses drift over time. They plug into standard policies such as ε-greedy and, in simulations, outperform approaches that ignore non-stationarity.
@misc{benac2021riskaversenonstationarymultiarmed,
title = {Risk averse non-stationary multi-armed bandits},
author = {Leo Benac and Fr{\'e}d{\'e}ric Godin},
year = {2021},
eprint = {2109.13977},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2109.13977}
}
The precursor to BITL: a gradient-free, constraint-based posterior over transition dynamics learned from expert data, which yields high-performing policies with much lower variance across datasets.
@misc{benac2023bayesianinversetransitionlearning,
title = {Bayesian Inverse Transition Learning for Offline Settings},
author = {Leo Benac and Sonali Parbhoo and Finale Doshi-Velez},
year = {2023},
eprint = {2308.05075},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2308.05075}
}
@inproceedings{benac2023inversetransitionlearning,
title = {Inverse Transition Learning for Characterizing Near-Optimal Dynamics in Offline Reinforcement Learning},
author = {Benac, Leo and Parbhoo, Sonali and Doshi-Velez, Finale},
booktitle = {ICML 2023 Workshop on Counterfactuals in Minds and Machines},
year = {2023}
}
Research focus: reinforcement learning. Advisor: Finale Doshi-Velez. Yiming Zhang (Jump Trading) Graduate Fellow.
GPA 4.27/4.30 (top 1%). Eric O’Connor Mathematics Medal for the highest GPA among graduating Mathematics students.
Required course for Harvard’s Data Science master’s program. Co-organized course content, lectures, and the course website; facilitated in-class discussions and graded assignments.
Co-organized course content and lectures, held weekly office hours, and supervised 50+ students on research projects.
Basketball has been a constant for me: I played in high school and college, reaching the Arizona state championship in 2015. I’m now a student manager and practice player for Harvard’s Division I women’s basketball team, the 2025 Ivy League champions. I’m also interested in real estate investing.
The best way to reach me is by email at lbenac at g dot harvard dot edu. I’m based in Cambridge, MA, and I’m happy to talk about research, collaborations, or roles starting in 2027.