QUBO-Optimized Evidence Selection for Retrieval-Augmented Question Answering with Unconventional Solvers

Researchers formulate evidence selection in retrieval-augmented question answering as a Quadratic Unconstrained Binary Optimization (QUBO) problem to address the limitations of top-k ranking and costly LLM-based selectors.

The method constructs an energy function balancing relevance, requirement coverage, support strength, redundancy, complementarity, and compactness. It is evaluated on HotpotQA against baselines including BM25, maximal marginal relevance, and various LLM-based set selectors. The QUBO selector achieves competitive exact-match and token-F1 performance while allowing context selection to be handled by Ising/QUBO-compatible solvers.

This approach opens a path toward RAG pipelines where LLMs are reserved for semantic processing and answer generation, while combinatorial evidence selection is offloaded to specialized hardware.