Mini-course (6 hours) on "Applied Large Language Models and Agentic Systems" within the course Natural Language Processing in the Master Programme in Artificial Intelligence A.Y. 2026 - 2027.
This mini-course covers the practical side of large language models, from building agentic systems to improving them with reinforcement learning (RL) based post-training. Students first learn how an agent works under the hood: the agent loop, tool use, guardrails, memory, multi-agent delegation and evaluation. They then post-train the model behind an agent with RL (GRPO with LoRA), with a focus on reward design and on spotting reward hacking. The course is built around two hands-on workshops on AWS. Students build a customer-service agent, then fine-tune a small open-source model with RL so it makes better tool-use decisions
Schedule:
Thursday 17/12, 12.30-14.30 (1.01): From Evaluating Agents to Rewarding Them: Evals and Reward Design
******** CSS e script comuni siti DOL - frase 9957 ********