Overview
Revna views games as interactive worlds with structure.
In Game AI research, models must go beyond language understanding. They need to grasp rules, track state, make strategic decisions, and anticipate the consequences of actions.
These properties also connect to applications in next-generation AI architectures and robotics.
Research Areas
Reinforcement Learning-Language Model Hybrid Game Agents - Game AI with Explainability
By combining the action optimization of reinforcement learning with the explanatory capabilities of language models, we aim to build agents that can explain in language why a given move is the right one.
This can lead to coaching AI for games such as chess, shogi, and poker that presents optimal moves together with reasons, as well as new game experiences with built-in explainability.
AI That Adapts Quickly Across Different Games
Traditional reinforcement learning often needs to relearn from scratch for each game, which makes adaptation slow.
By incorporating rule understanding through language models, Revna is researching AI that can adapt quickly even to games that are similar but not identical.
We aim to build more generalizable models by leveraging shared intuitions across game types, such as intuitions common to action games or forms of understanding common to puzzle games.
Why Games
Games are research environments that combine real-world-like complexity with clearly definable rules and evaluation criteria.
For that reason, they are especially well suited to developing and testing core AI capabilities such as reasoning, memory, planning, adaptation, and interaction.
Application Directions
- Explainable game agents
- Learning support and coaching AI
- General-purpose agents that adapt across different environments
- Applications to robotics and new model architectures