The field of large language models (LLMs) is moving towards more efficient and effective reasoning methods. Recent developments have focused on reducing token usage and latency while maintaining accuracy. One direction is the use of intra-request branch orchestration, which guides the model to terminate, duplicate, or continue branches based on predictions. Another approach is the use of adaptive computation methods, such as parallel thinking in latent space, which enables the model to fork or delete residual streams to perform additional thinking. These methods have shown significant improvements in token usage and latency reduction. Notably, some papers have proposed new frameworks and algorithms that leverage the strengths of multiple models, such as fusion methods and interdependent generation techniques. These advances have the potential to unlock new capabilities in LLMs and improve their overall performance. Noteworthy papers include: DUCHESS, which reduces cost and latency without sacrificing accuracy through intra-request branch orchestration. Chain-in-Tree, which introduces a plug-in framework that adaptively decides when to branch during search, reducing token generation and runtime by up to 85 percent. Thoughtbubbles, which proposes a transformer variant that natively performs parallel adaptive computation in latent space, outperforming standard decoder LMs and non-adaptive parallel computation approaches.