Emergent Collaborative Deliberation in Multi-Model AI Systems: A BFT-Derived Protocol for Epistemic Synthesis
Researchers have developed a new protocol called Consilium, which uses engineered cognitive personas to facilitate structured deliberation among multiple AI models. This approach treats disagreements between models as valuable epistemic signals rather than errors. Experiments showed that the assigned persona, not the underlying model's cost or frontier status, dictated the quality of analytical output, and that traditional RLHF alignment training can create specific blind spots in AI models. AI
IMPACT This protocol could improve AI reasoning by surfacing blind spots and enabling more robust, evidence-based conclusions from AI systems.