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

Abstract

Multi-hop question-answering architectures typically incur high costs by executing expensive retrieval operations for every query. Common techniques involve decomposing questions, performing multiple retrieval cycles, or navigating through bridge entities. These methods depend heavily on repeated Large Language Model (LLM) calls to rewrite or break down the query, resulting in significant additional token consumption. This approach becomes inefficient when operating under strict LLM budget constraints. Our analysis reveals, however, that a substantial portion of multi-hop questions are already resolved correctly by a single-shot Retrieval-Augmented Generation (RAG) process. Consequently, initiating an additional retrieval step for every query represents a wasteful use of resources.

To address this, we present RASER (Recoverability-Aware Selective Escalation Router), a suite of cost-effective routers derived from one-shot RAG and leveraging six specific features. RASER-2 functions as a binary decision maker, determining whether to halt processing or escalate to the supplementary retrieval action known as PRUNE. RASER-3 offers a more granular choice among one-shot RAG, PRUNE, and iterative retrieval via IRCoT. While it utilizes the same foundational features, RASER-3 explicitly incorporates a trade-off between cost and accuracy. Crucially, neither router requires an extra LLM invocation to make its determination.

Evaluated across six distinct LLMs and three multi-hop QA benchmarks, both RASER variants remain competitive with current state-of-the-art (SOTA) baselines in terms of F1 scores. Furthermore, they demonstrate significant efficiency, consuming only 41-49% of the tokens required by the "always-prune" strategy, and using fewer resources than baselines relying on iterative or decomposition-based retrieval.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC