Tag
This paper formulates the decision of when to use search in LLMs as an instance-level search-routing problem, using counterfactual supervision to train and improve routing policies, achieving macro-F1 improvements on Gemma and Qwen models.