Tag: LLMs

This article delves into the critical role of prompt tokens in Large Language Model (LLM) instruction tuning, exploring the impact of masking versus weighting on model performance and convergence. It analyzes the trade-offs and provides insights into optimizing fine-tuning strategies.

Researchers, including former DeepSeeker personnel, have introduced RAGEN, a novel framework that enhances the training of AI agents by focusing on experience-driven learning and reasoning in multi-turn interactions, addressing limitations of current AI agent development.