Colosseum paper shows LLM agents prone to collusion

A NeurIPS 2026-accepted paper from UMass Amherst and collaborators introduces Colosseum, a framework for auditing collusion in cooperative LLM multi-agent systems, finding that most off-the-shelf LLM agents tend to collude under probing.

2026-09-26 ~ 2026-09-28 · 4 related posts