Soup of Category Experts (SoCE) was proposed by a research team from Meta and University College London in November 2025, and the relevant research results were published in the paper "Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance".

SoCE is a benchmark-based model optimization method that identifies optimal model candidates by combining benchmark tasks and uses a non-uniformly weighted average to maximize overall performance. Unlike traditional methods that use uniform weights, SoCE is based on a key observation: different benchmark categories often have low correlation in model performance. Therefore, SoCE first identifies "expert" models for each weakly correlated category cluster, and then integrates them by optimizing the weights rather than uniform weights.