Budget-Aware Evidence Selection for Post-Hoc LLM Attribution

Post-hoc attribution for Large Language Models (LLMs) aims to support, or justify, LLM’s generated response by providing a concise set of passages that are retrieved from an external corpus that support (i.e., logically \textit{entail}) the output. Natural Language Inference (NLI) has become a central mechanism for post-hoc attribution of LLM outputs, allowing retrieved passages to be assessed according to whether they entail a generated answer. However, existing NLI-based approaches typically score passages independently and select the highest-scoring candidates. This becomes problematic when attribution operates under a strict citation budget: several highly entailing passages may provide redundant evidence, while complementary passages that collectively offer stronger support may be excluded. We therefore study budget-constrained attribution as a selection problem built on top of NLI-based attribution. We introduce \model{} (\textbf{Resp}onse \textbf{Att}ribution), which selects a subset of $k$ passages from an initially retrieved pool using a submodular objective that combines NLI-based answer entailment, retrieval relevance, and redundancy reduction. This formulation preserves NLI as the primary attribution signal while explicitly reasoning about the utility of the selected evidence set under a fixed budget. Experiments on question-answering datasets with two open-source LLMs show that \model{} consistently outperforms standard baselines, yielding stronger post-hoc attribution in budget-constrained settings.We make our code and examples available at the project's \href{https://anonymous.4open.science/r/RESPATT-803E/}{\color{blue}repository}