“The Bradley-Terry model is the formal name for the pairwise preference comparison method used by companies like Anthropic and OpenAI in their RLHF processes.”