RASER: Recoverability-Aware Selective Escalation Router for Multi-Hop Question Answering
Researchers have developed RASER, a new system designed to optimize multi-hop question-answering by reducing the number of expensive retrieval calls. RASER selectively escalates to more complex retrieval methods only when necessary, based on six features derived from an initial one-shot RAG. This approach significantly cuts down token costs, using 41-49% fewer tokens than always-pruning methods while maintaining competitive accuracy across various LLMs and benchmarks. AI
IMPACT Reduces computational costs for complex question-answering tasks, making LLM applications more efficient.