Research

Research themes

Agentic and multi-agent systems

I develop decentralized architectures that help groups of autonomous agents coordinate across different timescales and levels of abstraction. This includes hierarchical organizations, populations of interacting LLM agents, and methods for improving the reliability of long-running agentic processes.

Open-endedness and emergence

I study how novelty, cooperation, division of labor, governance, and cumulative culture can emerge in persistent populations of artificial agents. I am particularly interested in how collective behavior changes when agents share resources, environments, and information over long periods.

Embodied and applied intelligence

I approach intelligence as an ongoing loop between perception, action, memory, and learning, rather than as a process confined to static datasets. My current applied interests also include the use of AI to help prevent whale–ship collisions.

Exploration and adaptive learning

I design reinforcement-learning and quality-diversity methods that discover diverse behaviors, learn useful representations, and exploit sparse rewards with limited prior task knowledge.

Recent and selected work

TerraLingua: Emergence and Analysis of Open-endedness in LLM Ecologies

arXiv preprint, 2026

We introduce TerraLingua, a persistent ecology in which LLM-based agents survive, communicate, reproduce, and create artifacts that outlive them. Using an “AI Anthropologist,” we trace the emergence of cooperation, division of labor, governance attempts, and branching cultural lineages.

[Paper] · [Project] · [Code] · [Dataset]

TAG: A Decentralized Framework for Multi-Agent Hierarchical Reinforcement Learning

arXiv preprint, 2025

We introduce the TAME Agent Framework, a fully decentralized framework for building multi-agent hierarchies of arbitrary depth. Its LevelEnv abstraction standardizes information flow between levels while preserving loose coupling and support for heterogeneous agents.

[Paper] · [Code] · [Documentation]

Solving a Million-Step LLM Task with Zero Errors

arXiv preprint, 2025

We introduce a massively decomposed agentic process that combines focused micro-agents with multi-agent voting and error correction. The resulting system successfully completed a task requiring more than one million dependent LLM steps without an error.

[Paper]

Zero-shot Model-based Reinforcement Learning using Large Language Models

ICLR 2025

We investigate whether pretrained language models can predict continuous environment dynamics directly in context. Our Disentangled In-Context Learning method handles multivariate states and control signals, enabling applications in model-based policy evaluation and data-augmented off-policy reinforcement learning.

[Paper] · [Code]

Position: A Call for Embodied AI

ICML 2024 Position Paper track

We argue that progress toward general intelligence will require systems grounded in interaction, with perception, action, memory, and learning as integrated components. The paper proposes a conceptual framework and research agenda for embodied AI.

[Paper]

Discovering and Exploiting Sparse Rewards in a Learned Behavior Space

Evolutionary Computation, 32(3), 2024

We introduce STAX, an algorithm that learns its own low-dimensional behavior space from high-dimensional observations while exploring and exploiting newly discovered rewards. This reduces the prior task knowledge required in sparse-reward settings.

[Paper]

Earlier foundations

• Learning in Sparse Rewards Settings through Quality-Diversity Algorithms — PhD thesis, 2022.

• Sparse Reward Exploration via Novelty Search and Emitters — GECCO 2021.

• Unsupervised Learning and Exploration of Reachable Outcome Space — ICRA 2020.

• Virtual-to-real Deep Reinforcement Learning: Continuous Control of Mobile Robots for Mapless Navigation — IROS 2017.

Complete publication record

For the complete and most up-to-date list, see my Google Scholar, arXiv, DBLP, or ORCID profiles.