The field of large language models is moving towards improving faithfulness, expressiveness, and adaptability. Recent developments focus on resolving knowledge conflicts, integrating external knowledge, and enhancing model sensitivity to contextual information. Notable advancements include the introduction of novel decoding algorithms, such as confidence- and context-aware adaptive decoding, and the development of frameworks that enable continuous steering of model sensitivity to contextual knowledge. These innovations aim to break the trade-off between faithfulness and expressiveness, allowing large language models to generate more accurate and informative responses. Noteworthy papers include: KL-based self-distillation for large language models, which achieves state-of-the-art performance in vocabulary expansion. CoCoA, a novel token-level algorithm for principled conflict resolution, demonstrates superior performance in question answering and summarization benchmarks. Collaborative Decoding (CoDe) breaks the trade-off between faithfulness and expressiveness by dynamically integrating output probabilities generated with and without external knowledge. GUARD, a self-adaptive decoding method, effectively balances coherence with diversity in LLM outputs, exhibiting substantial improvements in generation speed.