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

2026-07-27 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

CAGE (Cognitive Attribution Graphs for Citation Generation) is a novel two-stage framework designed to improve faithful inline citation generation in long-form question answering systems. It addresses the challenge of "attribution ambiguity" and "evidence-boundary overrun," where existing LLM outputs often attach citations that are only topically related but lack direct support for specific claims. CAGE first employs a Cognitive Map Induction Model to explicitly construct answer-centered support subgraphs, linking each semantic answer unit to its supporting documents. Subsequently, a Structured Citation Reasoning Model translates these units into sentence-level claims, ensuring map-aligned citations. This approach aims to resolve combinatorial claim-document assignments and clarify evidential boundaries. Evaluated on datasets like ASQA, ELI5, and ExpertQA, CAGE achieved state-of-the-art performance, demonstrating the efficacy of its attribution-space contraction and map-guided citation generation.

Key takeaway

For NLP Engineers developing long-form question answering systems, CAGE offers a robust approach to enhance citation fidelity. If you are struggling with "attribution ambiguity" or "evidence-boundary overrun" in your LLM outputs, consider implementing an explicit cognitive attribution mapping stage. This method can significantly improve the verifiability of your generated answers by ensuring claims are directly supported by cited evidence, as demonstrated by state-of-the-art performance on ASQA, ELI5, and ExpertQA.

Key insights

CAGE uses explicit cognitive attribution maps to align semantic answer units with supporting documents, ensuring faithful inline citations.

Principles

• Resolve attribution ambiguity explicitly.
• Align semantic units with supporting documents.
• Contract attribution space for fidelity.

Method

CAGE employs a Cognitive Map Induction Model to build answer-centered support subgraphs, followed by a Structured Citation Reasoning Model that generates sentence-level claims with map-aligned citations.

Topics

• Long-Form Question Answering
• Inline Citation Generation
• Large Language Models
• Attribution Ambiguity
• Cognitive Attribution Graphs
• Natural Language Processing

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Explicit Evidence Grounding via Structured Inline Citation Generation — LLMs excel at document retrieval but struggle with precise evidence span identification for faithful inline citations.
• VERICITE: Evaluating Sentence-Level Citation Faithfulness in Retrieval-Augmented Medical Question Answering — RAG systems in medical QA exhibit low sentence-level citation faithfulness, often due to LLM generation behavior.
• TraceBack: Multi-Agent Decomposition for Fine-Grained Table Attribution — TraceBack provides fine-grained, cell-level attribution for table QA using a multi-agent decomposition framework.
• ScholarSum: Student-Teacher Abstractive Summarization via Knowledge Graph Reasoning and Reflective Refinement — ScholarSum integrates hierarchical knowledge graphs and iterative student-teacher refinement for fluent, factually faithful scientific summarization.
• Faithful by Construction: Claim-Anchored Attribution for Multi-Document Summarization — A modular, claim-anchored summarization framework inherently builds fine-grained attribution and faithfulness into the generation process.
• MultAttnAttrib: Training-Free Multimodal Attribution in Long Document Question Answering — MultAttnAttrib offers training-free, multimodal attribution for long document QA, outperforming prompting and matching frontier models with reduced latency.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.