Language model alignment and tool use

2025 • Keynote • PDF

Lecture given in Cornell Tech's Applied Machine Learning course. Abstract. Despite the power and impressive capabilities of today’s language models, many deficiencies still remain, including misinformation, harmful information, confabulations, hallucinations, suboptimal reasoning and logic, bias, fairness, and sycophancy. The field is focusing many resources on addressing these, and the lecture will cover how alignment (e.g., DPO, RLHF, PPO, GRPO) and tool use (e.g., RAG, code execution, MCP) are parts of the solution.