Using PDDL Planning to Ensure Safety in LLM-based Agents – Agustín Martinez Suñé
Using PDDL Planning to Ensure Safety in LLM-based Agents
Agustín Martinez Suñé – Ph.D. in Computer Science | Postdoctoral Researcher (Starting Soon), OXCAV, University of Oxford
Large Language Model (LLM)-based agents have demonstrated impressive capabilities but still face significant safety challenges, with even the most advanced approaches often failing in critical scenarios. In this talk, I’ll explore how integrating PDDL symbolic planning with LLM-based agents can help address these issues. By leveraging LLMs' ability to translate natural language instructions into PDDL formal specifications, we enable symbolic planning algorithms to enforce safety constraints throughout the agent’s execution. Our experimental results demonstrate how this approach ensures safety, even under severe input perturbations and adversarial attacks—situations where traditional LLM-based planning falls short. This work suggests a potential pathway for deploying safer autonomous agents in real-world applications. This work is a collaboration with Tan Zhi Xuan (MIT).
GS AI seminars
The monthly seminar series on Guaranteed Safe AI brings together researchers to advance the field of building AI with high-assurance quantitative safety guarantees.
Apply to speak
Video recordings
Mailing list
Series info website
Feedback
Donate 🧡