Dan Haramati

I am a PhD candidate in the Computer Science department at Brown University advised by Prof. George Konidaris studying generalization and transfer in Reinforcement Learning (RL).

I am particularly interested in compositional generalization—the ability to understand and produce novel combinations of known components or concepts—for its tractability and utility in solving increasingly complex tasks. My recent work studies how factored structure can facilitate this type of generalization.

I am currently exploring world modeling and model-based RL, specifically addressing questions such as: How can one learn transferable world models? What level of abstraction is suitable for world modeling? When is learning a world model fundamentally preferable to model-free approaches?

Prior to pursuing my PhD I completed a M.Sc. in Electrical and Computer Engineering advised by Prof. Aviv Tamar and a B.Sc. in Electrical Engineering and Physics at the at the Technion - Israel Institute of Technology.

Outside of research, I enjoy hiking, snowboarding, basketball and traveling. Before beginning my academic journey, I spent two defining years traveling in East Asia, Europe and America.

Email / Google Scholar / Twitter / LinkedIn / Github / CV

News

HECRL and LPWM have been accepted to ICLR 2026.
Hierarchical Entity-Centric Reinforcement Learning (HECRL) and Latent Particle World Models (LPWM) have been accepted to the World Modeling Workshop at Mila 2026 as Oral Presentations (top 7%).
Excited to present our entity-centric decision-making line of work at RLDM 2025.
Our paper on entity-centric diffusion-based behavioral cloning (EC-Diffuser) has been accepted to ICLR 2025.
I have officially started my PhD at Brown University! (September 2024)
Our workshop paper on decision-making with Deep Latent Particles has been accepted to the GenAI4DM Workshop at ICLR 2024.
ECRL has been accepted to ICLR 2024 as a Spotlight (top 5%).
Our paper on Entity-Centric Reinforcement Learning (ECRL) has been accepted to the GCRL Workshop at NeurIPS 2023 as a Spotlight Talk.

Selected Publications

	Hierarchical Entity-centric Reinforcement Learning with Factored Subgoal Diffusion Dan Haramati, Carl Qi, Tal Daniel, Amy Zhang, Aviv Tamar, George Konidaris Fourteenth International Conference on Learning Representations (ICLR 2026) World Modeling Workshop at Mila 2026 Oral (top 7%) paper / code (coming soon) / project page TL;DR: We present a hierarchical entity-centric framework for offline goal-conditioned RL that produces entity-factored diffusion-generated subgoals for an RL agent, yielding consistent performance gains on long-horizon, image-based, sparse-reward tasks.
	Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling Tal Daniel, Carl Qi, Dan Haramati, Amir Zadeh, Chuan Li, Aviv Tamar, Deepak Pathak, David Held Fourteenth International Conference on Learning Representations (ICLR 2026) World Modeling Workshop at Mila 2026 Oral (top 7%) paper / code (coming soon) / project page TL;DR: We present a self-supervised object-centric world model that learns keypoints and masks directly from videos, supports multi-modal conditioning, scaled to real-world multi-object datasets.
	EC-Diffuser: Multi-object Manipulation via Entity-centric Behavior Generation Carl Qi, Dan Haramati, Tal Daniel, Aviv Tamar, Amy Zhang Thirteenth International Conference on Learning Representations (ICLR 2025) paper / code / project page TL;DR: We present a diffusion-based goal-conditioned behavioral cloning method that operates on object-centric representations, enabling zero-shot generalization to novel compositions of objects.
	Entity-centric Reinforcement Learning for Object Manipulation from Pixels Dan Haramati, Tal Daniel, Aviv Tamar Twelfth International Conference on Learning Representations (ICLR 2024) Spotlight (top 5%) Goal-Conditioned Reinforcement Learning Workshop, NeurIPS 2023 Spotlight Talk paper / code / project page / talk TL;DR: We introduce a structured visual goal-conditioned RL framework for multi-object manipulation, demonstrating agents that generalize from 3 to over 10 objects.

Site based on this template.