Ivan had the honor of attending an invitation-only visionary workshop #235 on LLM-guided assurance and synthesis for CPS in Shonan, Japan. He presented the lab’s work on calibrating chain-of-thought confidence by discovering temporal patterns with Signal Temporal Logic.
Materials:
Some of the prominent debates at the workshop included:
- What does the probability of LLM choices have to do with the probability of LLM mistakes?
- Are world models necessary for intent?
- Where do specifications for LLMs come from, and are they truly separate from data?
- What is the equivalence class of semantically valid formalizations of natural language?
- How to combine the perfection of formal methods and the magic of AI?
- How is the explainability of LLM states different from the explainability of LLM outputs?
- Should we prioritize syntactic or semantic robustness in reasoning?
- How to establish multifaceted connections between the modalities of sensing (camera/lidar data), reasoning (language, both natural and formal), and control (actions)?
- How do you expect the robot to clean dishes well if you did not teach it?
- Does Lean have enough support for future, not-yet-existent mathematics?
- What aspects of agent-based CPS engineering should we trust more, and which – less?