Progressive Training for Explainable Citation-Grounded Dialogue: Reducing Hallucination to Zero in English-Hindi LLMs

This paper introduces XKD-Dial, a training method that teaches AI language models to generate dialogue responses in both English and Hindi while citing the sources of their information. The key innovation is a four-stage progressive training process that dramatically reduces false information (hallucination) to zero for smaller models, while the researchers also analyze how the models learn to cite sources through three different explainability techniques. The work demonstrates that well-designed training can make smaller models perform as well as larger ones and prevent the models from 'forgetting' their original abilities when learning new languages.