RASER: Recoverability-Aware Selective Escalation Router for Multi-Hop Question Answering

Researchers introduce RASER, a cost-efficient routing system for multi-hop question-answering that reduces token consumption by 51-59% compared to always-escalating methods while maintaining competitive accuracy. The system leverages six features from one-shot retrieval to intelligently decide whether additional retrieval rounds are necessary, eliminating wasteful LLM calls.

RASER addresses a fundamental inefficiency in retrieval-augmented generation (RAG) systems: the assumption that every multi-hop question requires expensive multi-round retrieval. The research reveals that single-pass RAG successfully answers a substantial portion of complex questions, making aggressive escalation strategies economically wasteful. This insight becomes increasingly critical as organizations deploy LLM-based systems at scale, where token costs directly impact operational margins.

The technical innovation centers on building lightweight routers that evaluate question answering quality without triggering additional LLM calls. RASER-2 implements binary decision logic (stop or escalate), while RASER-3 adds explicit cost-accuracy trade-off mechanisms across three strategies. By extracting decision signals from the initial RAG attempt—rather than performing redundant LLM reasoning—the system maintains computational efficiency while preserving answer quality.

For developers and companies operating budget-constrained LLM pipelines, RASER represents a practical optimization that could reduce inference costs by roughly half without sacrificing state-of-the-art performance. The approach generalizes across six different language models and multiple benchmarks, suggesting robust applicability. This work exemplifies the emerging pattern where success in AI infrastructure depends not just on model capability, but on intelligent system design that optimizes resource allocation.

The implications extend to production deployments where token budgets directly affect profitability. As LLM costs remain substantial, routing mechanisms that reduce unnecessary computation become competitive advantages. Future work likely focuses on refining decision criteria and extending such approaches to other multi-step reasoning tasks.

• →RASER reduces token consumption to 41-49% of baseline multi-retrieval methods while maintaining competitive F1 scores
• →The router makes routing decisions without extra LLM calls, using six features extracted from initial one-shot RAG
• →Empirical analysis shows many multi-hop questions are answered correctly by single-pass retrieval, eliminating need for escalation
• →System generalizes across six different LLMs and three QA benchmarks, indicating robust applicability
• →Cost-conscious routing mechanisms are becoming critical infrastructure for production LLM deployments