Cuong Xuan Chu

Knowledge graph memory can achieve superior factual recall without the token costs of traditional retrieval methods, but its effectiveness hinges on overcoming the challenge of adapter selection.

MADA-RL boosts compact model reasoning accuracy by 2% with 16 times fewer trainable parameters, redefining how critics learn from generators.