Surveying Techniques from Alignment to Reasoning

View the PDF file for the paper entitled Long Language Models after Training: Surveying techniques from alignment to think

PDF HTML (experimental) view

a summary:The appearance of large language models (LLMS) has mainly transformed the natural language processing, making it indispensable through areas ranging from conversation systems to scientific exploration. However, their pre -trained structures often reveal restrictions in specialized contexts, including restricted thinking capabilities, moral doubts, and optimal performance. These challenges require post-training language models (Polms) to address these defects, such as Openai-o/O3 and Deepseek-R1 (combined with large thinking models, or LRMS). This paper displays the first comprehensive survey of polms, which systematically tracks its development through five basic models: installation, which enhances the accuracy of the task; Alignment, which guarantees moral cohesion and compatibility with human preferences; Thinking, which develops multi -steps inference despite the challenges in the design of rewards; Efficiency, which improves resource use amid increasing complexity; Integration and adaptation, which extends the capabilities through various methods with tackling cohesion issues. Drawing the progress made from ChatGPT alignment strategies to innovative developments to think about Deepseek-R1, we explain how Polms benefit from data groups to reduce prejudices, deepen thinking capabilities, and enhance the ability to adapt to the field. Our contributions include a pioneering creation of the development of Polm, the classification of data technology organizer, and a strategic business schedule that emphasizes the role of LRMS in improving thinking efficiency and field flexibility. As a first survey of its scope, this work unifies the recent developments in Polm and establishes a strict intellectual framework for future research, and enhances the development of LLMS that excels in accuracy, moral durability, and diversity through scientific and societal applications.