Introduction This blogpost summarises the findings of my (lite) systematic literature review of Schelling coordination in LLMs that I undertook as part of the Apart Research Fellowship. If LLMs can identify and exploit Schelling points, they might coordinate without leaving observable traces of communication. This matters for AI safety because...
Author's Note: This work originated as a team project during the Apart Research AI Control Hackathon, where my teammates did the majority of the foundational work. I've continued developing these ideas as part of the Apart Lab Studio program, but the core insights and initial implementation were primarily their contributions....
This blogpost summarises the findings of my (lite) systematic literature review of Schelling coordination in LLMs that I undertook as part of the Apart Research Fellowship.
If LLMs can identify and exploit Schelling points, they might coordinate without leaving observable traces of communication. This matters for AI safety because many proposed safety measures assume that multiple instances of a model can't effectively coordinate. If this assumption breaks down, approaches that rely on one model monitoring another become vulnerable.
This review examines what current models can actually do, what risks this poses, and when coordination capability translates into safety-relevant threats.
This blogpost makes heavy use of concepts from the AI Control research agenda, particularly the ‘untrusted... (read 2169 more words →)
Author's Note: This work originated as a team project during the Apart Research AI Control Hackathon, where my teammates did the majority of the foundational work. I've continued developing these ideas as part of the Apart Lab Studio program, but the core insights and initial implementation were primarily their contributions. My motivation for publishing this article is to extend my own thinking on these concepts and to help disseminate what I believe are valuable ideas that deserve broader attention. I'm grateful to have been part of the original team and for their permission to build upon our shared work.
As AI systems become increasingly powerful, ensuring their safe deployment remains a critical challenge.... (read 2473 more words →)