The field of task-oriented dialogue systems and large language models is moving towards more interpretable, modifiable, and accurate models. Researchers are exploring new frameworks and methods to enable non-technical experts to define, test, and refine system behavior with minimal effort. One of the key directions is the development of novel frameworks that can convert expert knowledge into executable conversation logic, allowing for true zero-shot specification of task-oriented dialogue systems. Another important area of research is the improvement of large language models' accuracy, domain-specific reasoning, and interpretability in vertical domains. This is being achieved through the introduction of new preference alignment methods, such as retrieval-augmented and chain-of-thought enhanced alignment, which systematically construct binary preference datasets incorporating external knowledge support and explicit Chain-of-Thought reasoning. Additionally, there is a growing interest in building consistent dialogues for large language models from scratch, using skeleton-guided multi-turn dialogue generation, and synthesizing millions of diversified and complicated user instructions with attributed grounding. Noteworthy papers in this area include: CoDial, which introduces a novel framework for interpretable task-oriented dialogue systems, and RACE-Align, which proposes a retrieval-augmented and chain-of-thought enhanced preference alignment method for large language models. ConsistentChat and SynthQuestions also present significant contributions, with ConsistentChat providing a framework for building consistent dialogues and SynthQuestions synthesizing millions of diversified user instructions.