CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering

Long-form question answering (LFQA) systems often struggle with "evidence-boundary overrun," where models generate claims that are not fully supported by the retrieved documents. The authors investigate whether this is a structural issue caused by the implicit, joint generation of claims and citations, and propose a method to make this attribution process explicit.

The authors introduce CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that separates the identification of evidence from the generation of the final answer.

1. Cognitive Map Induction Model (CMI): This model takes a query and retrieved documents to construct an explicit "cognitive attribution map." This map consists of answer-centered subgraphs that link specific semantic answer units to their supporting documents, effectively pruning irrelevant information and defining the evidence boundary before any text is written.
2. Structured Citation Reasoning Model (SCR): This model takes the generated map as input to produce the final long-form answer. Because the map already defines which documents support which claims, the model is constrained to generate citations that are strictly aligned with the pre-validated evidence.

This approach transforms the complex, combinatorial task of assigning citations to claims into a structured, map-guided generation process.

Experiments on the ASQA, ELI5, and ExpertQA datasets demonstrate that CAGE achieves state-of-the-art performance in citation-grounded generation. By contracting the "attribution space"—the set of possible claim-document assignments—CAGE significantly reduces the risk of unsupported claims. The results show consistent improvements across various model sizes, suggesting that explicit attribution mapping provides a robust mechanism for enhancing the verifiability of LLM outputs.