Portrait of Mo Samsami

Mohammad Reza Samsami

Research Engineer @ Google DeepMind

I’m Mo, a Research Engineer at Google DeepMind in RL Engineering Team. I earned my master’s at Mila and have interned at Google Research, ServiceNow Research, and EPFL. I also co-founded Zette AI. Before that, I studied at Sharif University of Technology, where as an undergrad I co-founded Shenasa, a multidisciplinary student community on cognitive science.

My research goal is building generally intelligent machines that can plan under uncertainty in complex environments. I’m particulary interested in devloping machines that learn from experience, form rich internal models of the world, simulate possible futures, and choose actions with foresight. My current focus is on Genie.

Research Perspective

  1. Human and animal intelligence is based on a huge amount of background knowledge about how the world operates, which has been developed through observation and interaction. Internal models of the world guide judgments about what is likely, plausible, or impossible.
  2. Inspired by that, my work centers on world models: neural networks that transform raw sensory streams into semantic, spatio‑temporal representations. With these, machines can predict outcomes, reason about choices, and plan over long horizons.
  3. Reinforcement learning provides a paradigm that can use world models to come up with an intelligent behavior. Despite progress in AI, general‑purpose decision‑making in open‑ended, dynamic settings is still a challenging problem.
  4. One obstacle is complexity: real world is visually, temporally, and interactively rich, which makes modeling it a central challenge. I advocate integrating perception across modalities (like video, language, and action) to develop agents that operate reliably in realistic settings, especially where physical interaction matters.
  5. In this spirit, I work on Genie to help make scalable world models practical and to bring these ideas closer to meaningful real‑world impact.

Some Publications

Mastering Memory Tasks with World Models

Mohammad Reza Samsami*, Artem Zholus*, Janarthanan Rajendran, Sarath Chandar
2024
ICLR 2024 (oral, top‑1.2%)

My first paper on world models, asking a simple question: if an agent could recall better, could it make better decisions? We focused on the memory bottleneck in world models and introduced R2I, an RL agent with improved memory. R2I not only achieves superhuman performance on memory-heavy tasks but also remains competitive across diverse domains.

Too Big to Fool: Resisting Deception in Language Models

Mohammad Reza Samsami, Mats Leon Richter, Juan A. Rodriguez, Megh Thakkar, Sarath Chandar, Maxime Gasse
2024

I also explored world models in the context of LLMs, asking: do larger models, with stronger overall performance, develop richer internal representations of the world? Our experiments suggest they do, showing that larger models are notably more resilient to misleading in-context cues, thanks to their ability to integrate prompt information with a stronger underlying world model.

Interpretability in Action: Exploratory Analysis of VPT, a Minecraft Agent

Karolis Jucys*, George Adamopoulos*, Mehrab Hamidi, Stephanie Milani, Mohammad Reza Samsami, Artem Zholus, Sonia Joseph, Blake Richards, Irina Rish, Özgür Şimşek
2024

Not a world model paper per se, but closely related in spirit: examining how a foundation-model agent perceives the world and acts based on its perception. Through mechanistic interpretability, we discovered patterns in how the VPT agent links recent observations to decisions, as well as scenarios where its perception led to misaligned behavior, like killing villagers.

Rendering‑Aware Reinforcement Learning for Vector Graphics Generation

Juan A Rodriguez, Haotian Zhang, Abhay Puri, Aarash Feizi, Rishav Pramanik, Pascal Wichmann, Arnab Mondal, Mohammad Reza Samsami, Rabiul Awal, Perouz Taslakian, Spandana Gella, Sai Rajeswar, David Vazquez, Christopher Pal, Marco Pedersoli
2025

In the same spirit of agents learning through interaction with a world or its proxy, this work frames SVG generation as an RL loop: the model writes code, renders it, and optimizes for visual fidelity, learning how its actions shape outcomes without requiring differentiable rendering.

Causal Imitative Model for Autonomous Driving

Mohammad Reza Samsami, Mohammadhossein Bahari, Saber Salehkaleybar, Alexandre Alahi
2021

From an earlier chapter of my work, when I was particularly interested in applying causality to enhance decision-making systems. We explored how explicitly modeling causal structure can lead to more reliable policies. In the context of autonomous driving, this approach helped train policies that avoided common pitfalls such as inertia and collisions.