DBRR retrieval recovers 23.8% more evidence than top-k on HotpotQA

A new paper from Thomson D. Nguy, posted to arXiv (2608.18448), challenges the assumption that more context requires a larger retrieval budget. The proposed method, Dual-Bounded Relational Recall (DBRR), splits a fixed retrieval ceiling between relevance-selected seeds and bounded graph-adjacent context—pulling in evidence connected to top-ranking items rather than just the items themselves. In a controlled comparison against flat top-k retrieval using the same relevance-ranking stage and identical maximum retrieval units and tokens, DBRR achieved complete recovery of the official HotpotQA supporting-evidence set for every question.

Across 7,405 FullWiki questions, the primary DBRR allocation boosted complete-evidence recovery by 23.8 percentage points over its matched flat baseline (95% bootstrap interval 0.2269 to 0.2489). The effect was concentrated in bridge questions—those requiring multi-hop reasoning—which saw a 28.7-point increase, while comparison questions improved by only 4.2 points. DBRR improved 1,952 questions, tied on 5,261, and harmed just 192. Diagnostic tests confirmed real relationships outperformed random-neighbor and degree-preserving shuffled-graph controls. The implication is clear: under the same context budget, how you allocate context around ranked items matters as much as which items rank highest.