🚀 New from Meta AI Research: Souper-Model! By smartly averaging multiple model weights using a method called SoCE (Soup Of Category Experts), the team achieved strong performance across multilingual, tool-calling, and math tasks — without retraining a whole new model. 🔍 Key insight: benchmarks break down into categories that poorly correlate — so pick the “expert” model for each category and weight them non-uniformly for best gains. 📈 The result: meaningful boosts on major benchmarks. Read more here: https://lnkd.in/gtyErEhS #AI #LLM #ModelSoup #MetaAI #MachineLearning
