More Context, Same Budget: Dual-Bounded Relational Recall Beyond Top-K Retrieval

Abstract

More context does not require a larger retrieval budget. Under the same ceiling, a retrieval system can recover more of the evidence a question requires by following relationships between evidence that flat top-k ranking leaves behind.

The paper tests Dual-Bounded Relational Recall (DBRR), which allocates a fixed retrieval budget between relevance-selected seeds and bounded graph-adjacent context, against matched flat top-k retrieval using the same relevance-ranking stage and the same maximum number of retrieval units and tokens. The outcome is complete recovery of the official HotpotQA supporting-evidence set for each question.

Across 7,405 FullWiki questions, the Primary DBRR allocation increased complete supporting-evidence recovery by 23.8 percentage points over its matched flat baseline (paired risk difference 0.2377; question-level bootstrap 95% interval 0.2269 to 0.2489). It improved 1,952 questions, tied on 5,261, and harmed 192. Bridge questions drove the effect, with a 28.7-point increase; comparison questions showed a smaller 4.2-point difference.

In a prespecified, evaluation-only diagnostic population, real relationships also outperformed random-neighbor and degree-preserving shuffled-graph controls. The result concerns complete supporting-evidence recovery under the stated frozen design. It is not a claim about answer accuracy or general superiority across retrieval settings.

Fields and Methods

information retrieval, relational retrieval, multi-hop question answering, retrieval evaluation, evidence recovery.

• fixed-budget allocation between relevance-selected seeds and bounded graph-adjacent context
• matched flat top-k retrieval baseline with the same relevance-ranking stage, retrieval-unit ceiling, and token ceiling
• complete recovery of official HotpotQA supporting-evidence sets across 7,405 FullWiki questions
• paired risk difference with question-level bootstrap interval
• prespecified evaluation-only relationship diagnostics using random-neighbor and degree-preserving shuffled-graph controls

Collaborator Profile

Information-retrieval researchers, retrieval-augmented generation researchers, question-answering evaluation specialists, graph-retrieval researchers, and reproducibility-minded NLP researchers interested in evidence completeness under fixed context budgets.

Validation Needed

• Replicate the fixed-budget evidence-recovery result on additional datasets and retrieval corpora.
• Test whether the result persists with different relation structures, relevance rankers, and retrieval-unit definitions.
• Evaluate the separate relationship between complete supporting-evidence recovery and answer accuracy; this paper does not establish that relationship.
• Audit sensitivity to allocation choices and to bridge-versus-comparison question structure in follow-up studies.

Publication Posture

This is a public institute preprint. It has passed Radiant's internal disclosure review but has not been peer reviewed. Its permanent arXiv record is arXiv:2608.18448 in cs.IR, cross-listed cs.CL. The arXiv record is the primary public source.

Citation

Thomson D. Nguy. More Context, Same Budget: Dual-Bounded Relational Recall Beyond Top-K Retrieval. arXiv:2608.18448 [cs.IR, cs.CL].

Research Correspondence

For collaboration inquiries or research correspondence about this work, contact contact@theradiantinstitute.org.