Company

Language models excel at text-based tasks, but they still struggle in structured environments governed by complex rules and strategic constraints. These include games, as well as real-world systems such as industrial plants, where effective decision-making requires a deep understanding of physical processes and operational structure.

Deep reinforcement learning systems can achieve superhuman performance in highly specific environments such as Chess or Poker, but they are typically limited in two important ways: they cannot clearly explain their reasoning in language, and they often require millions of training simulations to reach strong performance. Unlike human players, they cannot read the rules of a new board game and quickly learn how to play.

REVNA's mission is to develop novel model architectures that combine language reasoning with the structural understanding required for these non-textual domains. Our current research focuses on game agents, particularly agents that can explain their strategies and integrate language-based reasoning with traditional reinforcement learning methods.

We are also working with construction companies to develop agents that can understand and reason over P&ID graphs, along with historical sensor and maintenance data, in order to support maintenance operations.