XKD-Dial grounds English-Hindi dialogue, cuts hallucinations

This looks less like classic RAG and more like training the model to treat citation as a behavior, which is appealing when you want grounded responses without bolting on retrieval at inference time.

• –The staged curriculum is the real story: multilingual adaptation, English citation SFT, bilingual SFT, then citation-aware GRPO.
• –The 0.0% hallucination figure is impressive, but it’s still a setup-specific result, not proof that the problem is solved generally.
• –Compared with RAG, this should be lighter at serving time, but it won’t replace retrieval for freshness, external documents, or open-world facts.
• –The explainability analyses matter because they try to show how citation behavior forms, not just whether the final metric improves.