Harsh Satija

Timeline

1. 2018–2019

   Decoupling dynamics and reward

   In April 2018, he coauthors the preprint Decoupling Dynamics and Reward for Transfer Learning with Amy Zhang and Joelle Pineau, followed by an ICLR workshop paper in 2019. Their approach separates learning the task representation, forward dynamics, inverse dynamics and reward function so reinforcement learning knowledge transfers across tasks.

2. 2020

   Backward value functions at ICML

   In 2020, Satija, Philip Amortila and Joelle Pineau publish Constrained Markov Decision Processes via Backward Value Functions at the International Conference on Machine Learning. The method translates cumulative cost constraints into state-based constraints to define a safe policy improvement procedure.

3. 2021

   Multi-Objective SPIBB at NeurIPS

   With Philip S. Thomas, Joelle Pineau and Romain Laroche, he presents Multi-Objective SPIBB at NeurIPS in 2021 for offline reinforcement learning with multiple reward signals. The method ensures a new policy performs at least as well as its baseline on each objective and is tested on learning critical-care policies for administering intravenous fluids and vasopressors in sepsis.

4. June 2022

   ICML workshop on responsible decision-making

   In June 2022, he helps organize an ICML workshop on responsible decision-making in dynamic environments, focusing on long-term welfare, safety and fairness.

5. September 2022–April 2023

   Group fairness in reinforcement learning

   His work on group fairness in reinforcement learning is selected for an oral presentation at EWRL in September 2022 and accepted in Transactions on Machine Learning Research in April 2023 with Matteo Pirotta, Alessandro Lazaric and Joelle Pineau. The study shows that reinforcement learning algorithms can learn fair policies without violating fairness requirements during training.

6. 2023

   McGill PhD on RL alignment

   He completes a PhD at McGill University in 2023 in collaboration with Mila under the supervision of Joelle Pineau. His thesis, Towards alignment of Reinforcement Learning agents, examines safety, robustness and fairness in reinforcement learning.

7. September 2025

   Failure modes in multi-agent debate

   In September 2025, he coauthors Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate with Andrea Wynn and Gillian Hadfield. The study demonstrates that debate can decrease accuracy over time as models shift from correct to incorrect answers in response to peer reasoning, even when stronger models outnumber weaker ones.

8. As of October 2026

   Vector Institute fellowship and LLM norms

   As of October 2026, Satija is a Vector Distinguished Postdoctoral Fellow at the Vector Institute and a postdoctoral fellow in the Normativity Lab, collaborating primarily with Gillian Hadfield on overseeing AI agents. In October 2026, he coauthors Reading the Room: Foundations, Design, and Challenges of Normative Competence in LLMs, finding that baseline LLM agents fail to learn norms even when useful and indiscriminately copy idiosyncratic noise alongside enforced rules.