How Do LLMs Cite? A Mechanistic Interpretation of Attribution in Retrieval-Augmented Generation

• URL: http://arxiv.org/abs/2606.28358v1
• Date: Tue, 09 Jun 2026 07:06:13 GMT
• Title: How Do LLMs Cite? A Mechanistic Interpretation of Attribution in Retrieval-Augmented Generation
• Abstract summary: We show how a large language model decides whether to attach an inline citation while answering a factoid question.<n>We show that amplifying or attenuating only those critical heads and eliminating 69% of spurious ones on PopQA without harming answer accuracy.<n>The results reveal a potential disconnect between the model's apparent reasoning and its internal computational pathway.
• Score: 3.4562767039451674
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Retrieval-Augmented Generation (RAG) aims to enhance the trustworthiness of Large Language Models (LLMs) by grounding their outputs in external documents, often using inline citations for verifiability. However, the faithfulness of these citations -- whether the model genuinely uses a source to generate an answer -- remains a critical, unverified assumption. This paper offers the first mechanistic account of how a large language model decides whether to attach an inline citation while answering a factoid question. Using the Llama-3.1-8B-Instruct model in a controlled experimental environment based on the PopQA dataset, we employ an activation patching approach. We map the underlying mechanism responsible for citation, discovering that it is not a single, localized component but a distributed, multi-stage "attributional ensemble" of attention heads and MLP layers. We show that amplifying or attenuating only those critical heads and MLPs repairs over 90% of missed citations and eliminates 69% of spurious ones on PopQA without harming answer accuracy. Although gains on the multi-document HotpotQA benchmark are modest, the same component set still moves citation rates in the intended direction, indicating that the underlying mechanism is not dataset-specific. The results reveal a potential disconnect between the model's apparent reasoning and its internal computational pathway, suggesting that inline citations can create a false sense of security.

Related papers

• Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research [75.8770212601511]
  Deep research (DR) systems produce long-form cited reports by orchestrating multiple agents that search and synthesize information from the web.<n> Citations are the primary mechanism for evaluating the faithfulness of these reports, yet current DR systems exhibit poor citation recall.<n>We propose an evaluation method that pinpoints which agent introduced each error by locally testing agent invocations for faithfulness and verifiability relative to their own inputs.
  arXiv  Detail & Related papers  (2026-08-25T09:34:18Z)
• Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs [8.721564756242431]
  CITETRACE is a dataset that traces the full citation chain from user query through retrieved source to generated answer.<n>We design a three-dimension evaluation framework that scores each citation on intent-purpose alignment, source suitability, and answer-source fidelity.<n>Across our pool, 30.6% of citations distort their sources and 27.1% originate from domain-inappropriate sources.
  arXiv  Detail & Related papers  (2026-05-27T14:54:05Z)
• Are Finer Citations Always Better? Rethinking Granularity for Attributed Generation [75.27453112832093]
  We show that enforcing fine-grained citations degrades attribution quality by 16-276% compared to the best-performing granularity.<n>Our analysis suggests that fine-grained (sentence-level) citations disrupt necessary semantic dependencies for attributing evidence to answer claims.<n>Strikingly, citation-optimal granularity leads to substantial gains in attribution quality while preserving or even improving answer correctness.
  arXiv  Detail & Related papers  (2026-04-01T22:16:23Z)
• CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era [51.63024682584688]
  Large language models (LLMs) introduce a new risk: fabricated references that appear plausible but correspond to no real publications.<n>We present the first comprehensive benchmark and detection framework for hallucinated citations in scientific writing.<n>Our framework significantly outperforms prior methods in both accuracy and interpretability.
  arXiv  Detail & Related papers  (2026-02-26T19:17:39Z)
• Probing for Knowledge Attribution in Large Language Models [45.47366023067617]
  Large language models (LLMs) often generate fluent but unfounded claims, or hallucinations.<n> Proper mitigation depends on knowing whether a model's answer is based on the prompt or its internal weights.<n>We show that a probe, a simple linear classifier trained on model hidden representations, can reliably predict contributive attribution.
  arXiv  Detail & Related papers  (2026-02-26T09:21:12Z)
• VeriCite: Towards Reliable Citations in Retrieval-Augmented Generation via Rigorous Verification [107.75781898355562]
  We introduce a novel framework, called VeriCite, designed to rigorously validate supporting evidence and enhance answer attribution.<n>We conduct experiments across five open-source LLMs and four datasets, demonstrating that VeriCite can significantly improve citation quality while maintaining the correctness of the answers.
  arXiv  Detail & Related papers  (2025-10-13T13:38:54Z)
• Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models [44.31597857713689]
  We introduce Active Indexing for the first stage, which creates generalizable, source-anchored bindings.<n>Experiments with Qwen-2.5-7B&3B show that Active Indexing consistently outperforms a Passive Indexing baseline.<n> internal citations complement external ones by making the model more robust to retrieval noise.
  arXiv  Detail & Related papers  (2025-06-21T04:48:05Z)
• SelfCite: Self-Supervised Alignment for Context Attribution in Large Language Models [51.90867482317985]
  SelfCite is a self-supervised approach to generate fine-grained, sentence-level citations for statements in generated responses.<n>The effectiveness of SelfCite is demonstrated by increasing citation F1 up to 5.3 points on the LongBench-Cite benchmark.
  arXiv  Detail & Related papers  (2025-02-13T18:55:13Z)
• On the Capacity of Citation Generation by Large Language Models [38.47160164251295]
  Retrieval-augmented generation (RAG) appears as a promising method to alleviate the "hallucination" problem in large language models (LLMs)
  arXiv  Detail & Related papers  (2024-10-15T03:04:26Z)
• Model Internals-based Answer Attribution for Trustworthy Retrieval-Augmented Generation [8.975024781390077]
  We present MIRAGE --Model Internals-based RAG Explanations -- a plug-and-play approach using model internals for faithful answer attribution in question answering applications. We evaluate our proposed approach on a multilingual QA dataset, finding high agreement with human answer attribution.
  arXiv  Detail & Related papers  (2024-06-19T16:10:26Z)
• ALiiCE: Evaluating Positional Fine-grained Citation Generation [54.19617927314975]
  We propose ALiiCE, the first automatic evaluation framework for fine-grained citation generation. Our framework first parses the sentence claim into atomic claims via dependency analysis and then calculates citation quality at the atomic claim level. We evaluate the positional fine-grained citation generation performance of several Large Language Models on two long-form QA datasets.
  arXiv  Detail & Related papers  (2024-06-19T09:16:14Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.