RECAST: Training a small router to search or compute the evidence a language model needs

RECAST trains a small language model to decide step by step whether to look up text, query a table, or write code, so a larger model gets the right context for questions that need calculation. The trained router averages 75.6 percent success across six benchmarks, almost 16 points above a prior system that uses a large Gemini model to search. #RetrievalAugmentedGeneration #LargeLanguageModels #ReinforcementLearning #NaturalLanguageProcessing #QuestionAnswering #ArXivSignals