Large Concept Model - Beyond Token-Based LLMs

Attend this plenary conference talk exploring the Large Concept Model as an alternative to token-based Large Language Models for achieving Advanced Machine Intelligence. Learn about the limitations of current token-based LLMs, including their lack of explicit reasoning and planning, hierarchical processing, and multilingual capabilities that are crucial characteristics of human intelligence. Discover how the Large Concept Model addresses these shortcomings by training on a multimodal and multilingual sentence representation space using diffusion-based methods. Explore the model's strong performance on generative tasks, its impressive zero-shot multilingual capabilities, and various explored variants including initial attempts at hierarchical text processing. Gain insights from Loïc Barrault, a Research Scientist at Meta AI with extensive experience in statistical and neural machine translation, multimodal processing, and lifelong learning methods, whose recent work includes contributions to NLLB200 for 200-language translation, Seamless-M4T for speech-to-speech translation across 100 languages, and reasoning in embedding spaces through the Large Concept Model.