STAI CDT PhD student Matt Barker presented the research paper, ‘Translating Latent State World Model Plans into Natural Language’, at the 2025 Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM), held from 11–14 June at Trinity College Dublin, Ireland.
RLDM is a unique international forum that bridges the gap between AI and cognitive science, bringing together researchers across disciplines—from machine learning and robotics to neuroscience and behavioral economics. With its dual focus on both artificial and natural intelligence, the conference provides a rare platform for cross-disciplinary dialogue on how humans, animals, and algorithms learn and make decisions.
Matt’s talk and publication are about a method for translating a black-box agent’s internal world model representations into natural language.
As Matt explains, “These agents (think a robot etc.) have their own model about how their actions impact the world. This allows them to make a plan and ‘imagine’ what would happen if they executed their plan e.g., ‘If I do this, then this would happen’. However, since they use neural networks to learn their model, their internal plans are unintelligible to use humans, essentially being a big list of numbers, e.g., ‘If I do this, then this big list of numbers will change to this list of numbers’, etc. Our method allows this internal representation to be translated into a natural language (i.e., plain English) description of what the agent plans to do and imagines will happen to the world before it acts in the world”.
Matt particularly enjoyed the multi-disciplinary element of the conference which gave him the opportunity to hear viewpoints and ideas from across disciplines. The conference also provided Matt with some more experience of presenting his work. He said, “The talk was a great experience! I was quite nervous beforehand as it was my first proper talk, but it went really well and people asked lots of interesting questions”.
Reflecting on his PhD journey so far, Matt shared: “The PhD is great! I love the freedom that research offers, it’s great to be able to explore ideas that are of interest to you. The CDT is great as well, it’s especially nice having such a big cohort of other students in the same boat as you, it makes it a much for social and collaborative experience than a lot of other PhDs”.
Huge congratulations to Matt on this achievement! You can watch his presentation on Youtube here or you can read the full publication below:
Barker, M. and Leonetti, M., 2025. Translating Latent State World Model Plans into Natural Language. Extended abstract presented at the Multi-Disciplinary Conference on Reinforcement Learning and Decision Making (RLDM 2025).