Recent advancements in language model (LM) technology have introduced innovative prompting strategies aimed at enhancing the reasoning and generation capabilities of these models. A notable development is the Retrieval Augmented Thoughts (RAT) strategy, which iteratively revises the chain-of-thought reasoning of large language models (LLMs) using retrieved information. This technique has shown significant improvements in long-horizon generation tasks by incorporating context-aware reasoning. Additionally, the Graph of Thoughts (GoT) paper highlights the utility of aggregating multiple thoughts from different reasoning paths, indicating a trend towards more sophisticated prompting methods. The prompt expansion approach, pioneered by HyperWriteAI, has been adopted by major players, including in the development of image models like DALL-E 3 and Ideogram 1.0, marking a step-change in output quality. Furthermore, a new paper introduces an unsupervised consistency training method that reduces misleading explanations caused by unverbalized biases in CoT prompting, enhancing the reliability of LLMs.