CMU LLM Inference - Beam Search and Variants

Learn beam search algorithms and their variants for large language model inference in this lecture from CMU's Advanced NLP course. Explore the fundamental concepts of beam search as a decoding strategy, examine various modifications and improvements to the basic algorithm, and understand the inadequacies of mode-based approaches in natural language generation. Discover how beam search balances computational efficiency with output quality in LLM inference, analyze different beam search variants that address specific challenges in text generation, and gain insights into when and why traditional mode-seeking methods may fall short in producing high-quality natural language outputs.