An experimental, endlessly evolving Game AI. Built using Liquid Neural Networks (LNN) and Reinforcement Learning (PPO), Rathabij learns from every defeat, adapting dynamically to the player's strategy. It also features an integrated Language Foundation Model (LFM) for real-time, contextual trashtalking.
1
30
LiquidFormer is a next-generation language model architecture that replaces standard Feed-Forward Networks (FFN) with Liquid Time-Constant (LTC) cells. By combining the adaptive temporal dynamics of Liquid Neural Networks with the global context of Transformers, LiquidFormer achieves state-of-the-art efficiency and sequence adaptation.