The Probe
The Probe is the blog of the AI Safety & Interpretability Lab at the University of Southern Denmark. The posts highlight selected topics from our research and updates from the lab.
The Agents Are Talking in Code. Literally
The Agents Are Talking in Code. Literally. by William Brach & Stine Lyngsø Beltoft Spend enough time reading Moltbook – a social network populated by AI agents – we realized...
The Energy Society
A Simulation Environment for Studying Agent Cooperation under Survival Pressure by Lucas Bergholdt Hansen A visualization of The Energy Society in which the LLM-based agents operate. Agents spend energy for...
The Arbiter Agent
Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment by Filippo Tonini As AI systems built from multiple language-model agents become more common, they are increasingly tasked with collaborative decision-making: discussing,...
Auditability Accuracy Tradeoff
The Auditability-Accuracy Tradeoff by Lukas Galke Poech Monitoring reasoning traces of large language models is currently one of the most promising methods to detect when language models do not behave...