Angeliki Lazaridou is a research scientist at Google DeepMind in London, known for her work on grounding large language models in interactive environments and multi-agent systems. Her research sits at the intersection of natural language processing, reinforcement learning, and multimodal learning, with a particular emphasis on how artificial agents can acquire and use language through communication and embodiment rather than static text corpora.
Lazaridou completed her PhD in computational linguistics at the University of Toronto, where she was advised by Suzanne Stevenson and worked on semantic representation and lexical acquisition. She subsequently held a postdoctoral position at the University of Oxford before joining DeepMind (now part of Google DeepMind) in 2016. At DeepMind, she became a leading figure in the study of emergent communication, exploring how agents develop their own languages to solve collaborative tasks.
Emergent communication and multi-agent learning
Lazaridou's early work at DeepMind, often in collaboration with Karl Moritz Hermann and others, demonstrated that reinforcement learning agents could develop compositional communication protocols from scratch. In a series of influential papers, she showed that agents trained to refer to objects in a grid world could evolve languages that exhibited systematic structure, though these languages often differed from human languages in their efficiency and robustness. Her 2017 paper "Multi-Agent Cooperation and the Emergence of (Natural) Language" (with Igor Mordatch) became a foundational reference for the field, showing that agents could learn to communicate through discrete symbols while solving cooperative tasks.
This line of research extended to investigating the factors that influence the emergence of human-like linguistic properties, such as compositionality and ambiguity. Lazaridou's work highlighted the trade-offs between communicative success and language complexity, and she contributed to understanding when and why agents develop shared conventions. Her findings have implications for designing more interpretable and controllable AI systems, as well as for theories of language evolution.
Grounding language in multimodal and embodied contexts
A central theme of Lazaridou's research is grounding - connecting linguistic symbols to perceptual and physical experiences. She has worked on aligning vision-language models with textual representations, enabling agents to reason about images and video through language. In a 2021 paper, she and colleagues introduced a method for grounding language models in visual worlds by training them on paired image-text data, improving their ability to answer visual questions and generate descriptive captions.
More recently, Lazaridou has focused on grounding large language models in interactive environments, such as text-based games and simulated worlds. This work aims to move beyond static pretraining data by allowing models to learn from feedback, exploration, and social interaction. She has been involved in projects that use reinforcement learning to fine-tune language models for goal-directed behavior, including tasks like navigation and tool use, which require both linguistic understanding and planning.
Contributions to AI safety and evaluation
Lazaridou has also contributed to the evaluation and safety of large language models. She has co-authored work on measuring the factual accuracy and consistency of generated text, and on detecting and mitigating biases in model outputs. Her perspective emphasizes that grounding in real-world interactions can reduce hallucination and improve reliability, as models are forced to align their language with observable states.
In 2023, she participated in discussions and publications about the challenges of aligning generative AI systems with human values, particularly in multi-agent settings where models must coordinate with each other and with people. She has argued for the importance of studying communication as a means of achieving shared goals, rather than merely as a pattern-matching task.
Selected publications and recognition
Lazaridis's work has been published in top venues including the Conference on Neural Information Processing Systems (NeurIPS), the International Conference on Learning Representations (ICLR), and the Association for Computational Linguistics (ACL). Her papers have received thousands of citations, and she is a frequent invited speaker at workshops on emergent communication and interactive AI.
She has served as an area chair for major NLP and machine learning conferences and has mentored numerous interns and students at Google DeepMind. Her research has been featured in popular science outlets, highlighting the surprising ways in which AI agents develop their own languages.
Current directions
As of 2025, Lazaridis continues to explore how large language models can be embedded in embodied agents that perceive, act, and communicate in dynamic environments. Her recent projects involve multi-agent reinforcement learning with human-like negotiation and cooperation, as well as the use of world models to improve language understanding. She remains a vocal advocate for interdisciplinary approaches that combine insights from linguistics, cognitive science, and computer science.
Her work is widely seen as bridging the gap between static language models and the interactive, grounded intelligence required for real-world applications such as robotics, virtual assistants, and collaborative AI systems.