PaperlayerView exact revision on arXiv

Source-linked AI summary

VeriCite: Towards Reliable Citations in Retrieval-Augmented Generation via Rigorous Verification

Haosheng Qian, Yixing Fan, Jiafeng Guo, Ruqing Zhang, Qi Chen, Dawei Yin, Xueqi Cheng

arXiv:2510.11394v1cs.IR

TL;DR

RAG systems can hallucinate when retrieval introduces irrelevant information, while existing citation methods face annotation, computation, and multi-citation limitations. VeriCite verifies generated claims and selected evidence across three stages, and experiments show improved citation quality while maintaining comparable answer correctness.

  • Problem

    RAG can incorporate irrelevant retrieved information into answers, and existing intrinsic or extrinsic citation methods face practical and multi-citation limitations.

    Source: §1 · p. 1

  • Method

    VeriCite generates an initial answer, selects and NLI-verifies useful evidence, then refines the response using validated statements and citations.

    Source: §1 · p. 2

  • Results

    VeriCite significantly improves citation quality while maintaining answer correctness comparable to strong baselines across five LLMs and four datasets.

    Source: §6 · p. 7

  • Takeaways & Limitations

    Pre-screening and pre-attributing verified evidence can reduce citation difficulty for the generator while preserving answer accuracy.

    Source: §1 · p. 2

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG) has emerged as a crucial approach for enhancing the responses of large language models (LLMs) with external knowledge sources. Despite the impressive performance in complex question-answering tasks, RAG still struggles with hallucinations. Attributing RAG-generated content through in-line citations has demonstrated potential in reducing hallucinations and facilitating human verification. Existing citation generation methods primarily rely on either fine-tuning the generator or employing post-processing approaches for citation matching. However, the former approach demands substantial annotated data and computational resources, while the latter often encounters difficulties in managing multiple citations and frequently produces suboptimal results. In this paper, we introduce a novel framework, called VeriCite, designed to rigorously validate supporting evidence and enhance answer attribution. Specifically, VeriCite breaks down into a three-stage generation: 1) The initial answer generation first generates a response based on all available contexts and has its claims verified through the NLI model; 2) the supporting evidence selection assesses the utility of each document and extracts useful supporting evidences; 3) the final answer refinement integrates the initial response and collected evidences to produce the final, refined answer.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.

1 Introduction

RAG can expand LLM capabilities but retrieved noise can worsen hallucinations, making attributable responses important for trust and verification. VeriCite addresses citation-generation limitations by rigorously verifying evidence before refining the final answer.

  • Retrieved irrelevant information can cause LLMs to incorporate noise and generate incorrect answers, exacerbating hallucinations.Source: §1 · p. 1
  • Inline citations improve traceability, error diagnosis, and transparency by anchoring statements to source materials.Source: §1 · p. 1
  • Intrinsic attribution requires substantial annotation and computation for fine-tuning, while in-context learning is sensitive to examples and generalizes poorly.Source: §1 · p. 2
  • Extrinsic matching is efficient with similarity metrics but threshold-sensitive, while NLI methods struggle when one statement needs multiple citations.Source: §1 · p. 2
  • VeriCite uses initial answer generation, supporting evidence selection, and final answer refinement, with NLI verification applied before citation-bearing synthesis.Source: §1 · p. 2
  • Across datasets and LLMs, VeriCite maintains answer accuracy comparable to baselines while significantly improving citation quality by pre-screening and pre-attributing useful evidence.Source: §1 · p. 2

2 Related Work

Prior attribution work divides into intrinsic methods that generate citations during answering and extrinsic methods that add citations through post-processing. The literature also develops automated and human-centered ways to evaluate citation faithfulness and quality.

  • Intrinsic attribution uses fine-tuning or prompting to produce integrated citations, including systems such as WebGPT, WebGLM, APO, LongCite, ALCE, and VTG.Source: §2 · p. 2
  • Extrinsic attribution first generates an answer, matches it to retrieved passages, and inserts citations using similarity metrics or NLI entailment judgments.Source: §2 · p. 2
  • Citation-generation research relies on both human assessment and automated metrics to evaluate attribution performance.Source: §2 · p. 2
  • AIS evaluates source attribution and content consistency manually, while AutoAIS approximates these judgments with sentence-level NLI.Source: §2 · p. 3
  • ALCE introduced citation recall and precision metrics alongside a benchmark covering fluency, correctness, and citation quality, while CAQA adds four-category attribution judgments.Source: §2 · p. 3

3 VeriCite framework

VeriCite treats citation-supported question answering as a staged process: generate an answer, select and verify evidence, then synthesize a fluent response from validated statements and citations.

3.1 Task Formulation

  • Given a query and top-k passages, the task is to generate an answer whose statements may cite sets of retrieved passages.Source: §3.1 · p. 3

3.2 Initial Answer Generation

  • Initial answer generation concatenates the query and retrieved passages, uses few-shot citation instructions, and decomposes statements for NLI-based entailment verification.Source: §3.2 · p. 3
  • Unsupported initial statements are discarded when the cited passages do not entail them, filtering likely hallucinated content.Source: §3.2 · p. 3

3.3 Supporting Evidence Selection

  • Supporting evidence selection addresses retriever-generator misalignment by evaluating each passage’s utility and generating evidence from retained passages.Source: §3.3 · pp. 3–4
  • Generated evidence is decomposed into statements and checked against its original passage with NLI, retaining only entailed statements.Source: §3.3 · p. 4
  • Verified evidence statements receive citation markers before summarization, decoupling attribution from generation and reducing citation demands on the LLM.Source: §3.3 · p. 4

3.4 Final Answer Refinement

  • Final refinement synthesizes validated statements by improving logical flow, removing redundancy, consolidating citations, and preserving semantic content with retrieved passages as grounding context.Source: §3.4 · p. 4

4 Experiments

Experiments evaluate VeriCite across four datasets and five open-source LLMs against citation-generation baselines. Overall, VeriCite substantially improves citation quality while generally matching or exceeding baselines in answer correctness, with dataset-specific limitations.

4.1 Datasets and Models

  • The evaluation covers ASQA, ELI5, HotpotQA, and MuSiQue using EM or Claim Recall for correctness and Citation F1 for citation quality.Five models are evaluated: Llama3-8B-Instruct, Gemma-2-9B-it, GLM-4-9B-Chat, and two Qwen2.5 variants.Source: §4.1 · p. 5

4.2 Baselines

  • Baselines include Vanilla, abstractive Summary, extractive Snippet, and fine-tuned preference-optimization APO, using top-5 passages, two few-shot examples, TRUE verification, and greedy decoding.Summary reduces redundancy, Snippet preserves source expressions, and APO combines supervised fine-tuning with preference optimization.Source: §4.2 · p. 5, §4.3 · p. 5

4.4 Main Results

  • On ASQA, VeriCite outperforms all baselines in answer correctness across five models, while citation F1 gains occur for Llama3 and Qwen2.5 but not Gemma-2 or GLM-4.GLM-4 achieves a 4.54% correctness increase over the best Vanilla baseline.Source: §4.4 · p. 5
  • 11.41% average Citation F1 improvement over the best baseline occurs across all five models on ELI5, despite lower answer correctness.This dataset contains predominantly non-factual, complex explanatory questions.Source: §4.1 · p. 4, §4.4 · p. 5
  • On multi-hop QA, only Qwen2.5 surpasses all baselines in answer correctness, whereas all models significantly outperform the strongest baselines in Citation F1.The remaining models fall slightly below their respective best correctness baselines, indicating challenges with cross-passage information integration.Source: §4.4 · p. 5
  • VeriCite matches or exceeds the best baselines in answer correctness, with particularly notable gains for Qwen2.5 and GLM-4, while citation quality improves across all five models.Both Qwen2.5 parameter scales show similar improvements, suggesting potential applicability to larger-scale models.Source: §4.4 · p. 5

5 Analysis

The analysis shows that both answer-generation stages contribute complementary content, while NLI verification is especially important for citation reliability. Among verifier choices, NLI offers the best performance-cost balance.

5.1 Ablation Study

  • Ablation experiments use Llama3-8B-Instruct on ASQA and remove one of three components: initial generation, evidence selection, or NLI verification.The no-verification variant assumes all generated statements are supported by retrieved passages.Source: §5.1 · pp. 5–6
  • Removing either initial answer generation or supporting evidence selection substantially reduces answer correctness, whereas removing NLI verification severely degrades citation quality.The two generation stages contribute complementary statements, while verification primarily protects citation reliability.Source: §5.1 · pp. 5–6

5.2 Discussion of Verifier

  • Using a general LLM as verifier decreases both answer correctness and citation quality relative to NLI, while DeepSeek-R1 marginally improves correctness with nearly equivalent citation quality.The verifier comparison uses Llama3-8B-Instruct for answer generation and 200 ASQA samples.Source: §5.2 · pp. 6–7
  • NLI is selected as the optimal verifier because it balances performance gains with resource expenditure more effectively than the tested LLM and DeepSeek-R1 alternatives.The LLM and NLI verifiers have comparable parameter sizes and operational costs, while DeepSeek-R1 has substantially higher practical deployment costs.Source: §5.2 · p. 7

6 Conclusion

VeriCite enhances citation quality in RAG through three sequential stages while maintaining answer correctness comparable to strong baselines. Ablations support its core components, and the verifier analysis motivates using an NLI model.

6 Conclusion

  • VeriCite combines initial answer generation, supporting evidence selection, and final answer refinement to improve citation quality while preserving competitive answer correctness.Ablation studies support the necessity of the core components, and verifier analysis supports the NLI design choice.Source: §6 · p. 7

A Prompt

The appendix provides prompt templates for four stages: initial answer generation, supporting evidence checking, supporting evidence extraction, and final answer refinement.

A.1 Initial Answer Generation

  • The initial-answer stage is labeled as a prompt template for generating the first response.Source: §A.1 · p. 8
  • The initial-answer prompt instructs the model to use relevant passage information while ignoring irrelevant content and preserving comprehensive, accurate details.Source: §A.1 · p. 8
  • The initial-answer prompt requires citations for every factual claim and at least one cited passage in each sentence.Source: §A.1 · p. 8

A.2 Supporting Evidence Check; A.3 Supporting Evidence Extraction; A.4 Final Answer Refinement

A.2 Supporting Evidence Check

  • The supporting-evidence check prompt asks whether each passage helps answer the question and restricts the output to “Yes” or “No.”Source: §A.2 · p. 8

A.3 Supporting Evidence Extraction

  • The supporting-evidence extraction prompt directs the model to retain relevant details from a passage while omitting irrelevant information.Source: §A.3 · p. 8

A.4 Final Answer Refinement

  • The final-refinement prompt supplies answer statements with citations and their original references for synthesis.Source: §A.4 · p. 8
  • The final-refinement prompt asks the model to summarize the statements and merge their citations into the answer.Source: §A.4 · p. 8

Related papers

Loading 2510.11394v1…