Llama-Omni, Attention heads, Unified View of Preference Learning, scalable IRL for LLM training, OneGen one-pass retrieval and generation, role of in-context learning.