Grounding Answers in Retrieved Passages: A Study of Hallucinations in Retrieval-Augmented Language Models

Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering

This paper establishes the effectiveness of passage-conditioned generative open-domain QA on Natural Questions and TriviaQA. It evaluates answer performance and evidence aggregation rather than sentence-level support.

How to position yourself: Cite it as prior evidence for retrieval-enhanced generative QA, distinguishing the draft's focus on unsupported sentences and entity-mismatched retrieval.

From its abstract: We obtain state-of-the-art results on the Natural Questions and TriviaQA open benchmarks.

The Power of Noise: Redefining Retrieval for RAG Systems

This paper directly examines how retrieved passage relevance affects RAG performance and reports harm from highly ranked but irrelevant passages. That finding closely relates to the draft's loss of grounding benefits under entity-mismatched retrieval, although the reported outcome is accuracy rather than sentence support.

How to position yourself: Cite it for harmful effects of nonrelevant retrieved context, distinguishing the draft's entity-specific mismatch analysis and entailment-based outcome.

From its abstract: the retriever's highest-scoring documents that are not directly relevant to the query (e.g., do not contain the answer) negatively impact the effectiveness of the LLM.

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

Self-RAG addresses factuality through adaptive retrieval and reflection on both passages and generated answers. Its treatment of irrelevant retrieval and evidence use directly overlaps the draft's explanation of residual unsupported statements.

How to position yourself: Cite it as an alternative approach to controlling retrieval and grounding, distinguishing the draft's diagnostic comparison using a dense retriever and BART.

From its abstract: Our framework trains a single arbitrary LM that adaptively retrieves passages on-demand, and generates and reflects on retrieved passages and its own generations

Disentangling Faithfulness Hallucinations in Retrieval-Augmented Generation: A Systematic Benchmark and Analysis

This paper directly studies evidence-unsupported generation and separates retrieval errors from generation errors. It closely overlaps the draft's central question, while covering different datasets, model scales, and retrieval strategies.

How to position yourself: Cite it as a directly related empirical decomposition of RAG faithfulness failures, distinguishing the draft's BART setup and related-entity mismatch finding.

From its abstract: By disentangling retrieval and generation errors, we characterize how different pipeline components contribute to hallucinations in RAG systems.

Hybrid Retrieval for Hallucination Mitigation in Large Language Models: A Comparative Analysis

This preprint explicitly evaluates the relationship between retriever effectiveness and hallucination reduction, comparing sparse, dense, and hybrid retrieval. It overlaps the draft's central concern but studies retrieval alternatives rather than primarily comparing retrieval with no retrieval.

How to position yourself: Cite it for evidence linking retrieval relevance to hallucination rates, distinguishing the draft's sentence-level entailment measure and entity-mismatched evidence condition.

From its abstract: This study evaluates the relationship between retriever effectiveness and hallucination reduction in LLMs using three retrieval approaches

Retrieval Augmented Generation: A Simplified Approach to Its Operation and Understanding

This study also evaluates RAG for QA and compares it with BART, reporting improved factual accuracy. It is a close empirical precedent, although the abstract does not establish an equivalent sentence-entailment evaluation.

How to position yourself: Cite it for prior evidence that RAG improves QA factuality over BART; the draft distinguishes itself through retrieved-passage entailment and entity-mismatch analysis.

From its abstract: the obtained results show that the RAG model outperforms both (BERT and BART) in terms of factual accuracy and semantics.

When Attention Escapes: Diagnosing and Mitigating Hallucination in Retrieval-Augmented Generation

This paper directly investigates hallucinations caused by generators underusing retrieved evidence. It offers a mechanistic diagnosis and intervention relevant to the generator-side failure raised in the draft.

How to position yourself: Cite it for evidence-underutilization failures and attention-based diagnosis; the draft instead uses sentence entailment and examines incorrect-entity retrieval.

From its abstract: Retrieval-Augmented Generation (RAG) mitigates hallucinations by grounding generation in retrieved evidence, yet failures persist when models underutilize the retrieved context.

Hallucination Reduction in Large Language Models with Retrieval-Augmented Generation Using Wikipedia Knowledge

This study also examines hallucination reduction through Wikipedia-based RAG and reports gains in response quality. It is a close precedent, although it uses Mistral and does not describe the draft's entailment metric or entity-mismatch test.

How to position yourself: Cite it for prior empirical evidence on Wikipedia-grounded hallucination reduction; the draft adds a BART comparison and sentence-support analysis under entity mismatch.

From its abstract: By dynamically integrating relevant information, the Mistral model demonstrates substantial improvements in precision, recall, and overall quality of responses.

Rowen: Adaptive Retrieval-Augmented Generation for Hallucination Mitigation in LLMs

Rowen explicitly recognizes that irrelevant retrieved information can introduce hallucinations, closely matching the draft's concern about related but incorrect entities. It proposes adaptive retrieval to balance external evidence against parametric knowledge.

How to position yourself: Cite it for the risk of retrieval-induced hallucinations and adaptive mitigation; the draft quantifies lost grounding benefits under entity-mismatched retrieval.

From its abstract: While incorporating external information can help fill knowledge gaps, it also introduces the risk of irrelevant information, thereby increasing the likelihood of external hallucinations.

Natural Questions: A Benchmark for Question Answering Research

This paper introduces Natural Questions, the benchmark on which the draft reports its central result. It is the primary dataset reference rather than a competing hallucination study.

How to position yourself: Cite it to identify the evaluation dataset and its provenance; the draft uses that dataset to study retrieval-grounded answer support.

From its abstract: We present the Natural Questions corpus, a question answering data set. Questions consist of real anonymized, aggregated queries issued to the Google search engine.

ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability

ReDeEP investigates why RAG outputs can conflict with accurate retrieved evidence, directly addressing generator-side failures. It separates external-context use from parametric knowledge through mechanistic analysis.

How to position yourself: Cite it for evidence that relevant retrieval does not ensure faithful generation; the draft evaluates observable sentence support and incorrect-entity retrieval rather than internal mechanisms.

From its abstract: However, even with accurate and relevant retrieved content, RAG models can still produce hallucinations by generating outputs that conflict with the retrieved information.

Detecting and Repairing Hallucinations in Retrieval-Augmented Generation

This preprint studies unsupported claims in retrieved-source-conditioned answers and evaluates post-generation repairs. Its claim-level source checking is relevant, but it does not primarily test retrieval against a no-retrieval generator.

How to position yourself: Cite it for claim-level grounding checks and repair trade-offs, contrasting its repair interventions with the draft's measurement of retrieval's effect.

From its abstract: we split each flagged answer into individual factual claims, check each against the retrieved source

Investigating the Robustness of Retrieval-Augmented Generation at the Query Level

This preprint isolates RAG components to study failures caused by query perturbations. It offers a related robustness perspective, but query variation differs from the draft's focus on retrieved passages about the wrong entity.

How to position yourself: Cite it for component-level robustness analysis, contrasting query perturbations with the draft's examination of entity-mismatched evidence.

From its abstract: We study each module in isolation as well as their combined effect in an end-to-end question answering setting

Seven Failure Points When Engineering a Retrieval Augmented Generation System

This experience report documents RAG failure points across three application domains and attributes limitations to both retrieval systems and language models. It supports the draft's component-level framing without directly measuring its proposed grounding outcome.

How to position yourself: Cite it for practical evidence of failures across the retrieval-generation pipeline, distinguishing the draft's controlled open-domain QA comparison.

From its abstract: However, RAG systems suffer from limitations inherent to information retrieval systems and from reliance on LLMs.

Hallucination in Large Language Models and Retrieval-Augmented Generation: Mechanisms, Mitigation, and Evaluation

This review distinguishes retrieval reliability, evidence use, and evaluation as factors in RAG hallucinations. Its framing closely matches the draft's motivation, but it supplies a synthesis rather than a directly comparable experiment.

How to position yourself: Cite it for the distinction between factuality and evidence faithfulness, positioning the draft as a specific empirical study of sentence support and entity mismatch.

From its abstract: Errors may arise during query formulation, document retrieval, evidence aggregation, and answer grounding.

Reducing hallucination in structured outputs via Retrieval-Augmented Generation

This paper reports that retrieval reduces hallucinations in generated workflows and improves out-of-domain generalization. It provides related evidence for retrieval's benefits, but structured enterprise outputs differ from open-domain answer sentences.

How to position yourself: Cite it as evidence of hallucination reduction in another output setting, distinguishing the draft's Natural Questions evaluation and sentence-level entailment measure.

From its abstract: our proposed system significantly reduces hallucinations in the output and improves the generalization of our LLM in out-of-domain settings.

Retrieval-Augmented Generation for Large Language Models: A Survey

This survey organizes RAG architectures, their components, and evaluation challenges. It supplies broad context for the draft's retrieval-generator pipeline rather than a direct test of sentence grounding.

How to position yourself: Cite it for the broader RAG design space, distinguishing the draft's focused empirical analysis of unsupported answer sentences.

From its abstract: It meticulously scrutinizes the tripartite foundation of RAG frameworks, which includes the retrieval, the generation and the augmentation techniques.

Why Retrieval Doesn't Cure Everything: A Review of Hallucination in Retrieval-Augmented Generation

This review covers both retrieval failures and unfaithful generation, including cases where correct evidence is present. It also discusses NLI-based faithfulness scoring, making it relevant to both the draft's question and its measurement approach.

How to position yourself: Cite it for existing failure taxonomies and NLI-based grounding evaluation, distinguishing the draft's empirical retrieval comparison and entity-mismatch result.

From its abstract: surveys the detection methods (NLI-based faithfulness scoring, LLM-as-judge, internal-state probing, evidence-graph consistency)

Dense Passage Retrieval: Is it Retrieving?

This paper mechanistically investigates dense passage retrieval training and identifies limits on what trained retrievers can retrieve. It offers background on possible retrieval failures but does not directly evaluate answer grounding or entity confusion.

How to position yourself: Cite it for limitations of dense retrieval, distinguishing its mechanistic retriever analysis from the draft's downstream grounding evaluation.

From its abstract: We also uncover a limitation in this training style: the internal knowledge of the pre-trained model bounds what the retrieval model can retrieve.

Evaluation of Retrieval-Augmented Generation: A Survey

This survey examines evaluation of retrieval and generation, including relevance, accuracy, and faithfulness. It provides methodological context for interpreting the draft's NLI-based support measure without establishing that measure's validity.

How to position yourself: Cite it to situate sentence entailment within RAG evaluation dimensions, distinguishing the draft's specific metric and controlled QA comparison.

From its abstract: we examine and compare several quantifiable metrics of the Retrieval and Generation components, such as relevance, accuracy, and faithfulness

Retrieval-Augmented Generation and Hallucination in Large Language Models: A Scholarly Overview

This overview addresses RAG as a way to reduce hallucinations, including its effectiveness and remaining limitations. It supplies background rather than a directly comparable experiment.

How to position yourself: Cite it for an overview of RAG and hallucination mitigation; the draft contributes a specific sentence-support evaluation and an entity-mismatch finding.

From its abstract: This article explores the underlying causes of hallucinations in LLMs, the mechanisms and architectures of RAG systems, their effectiveness in reducing hallucinations

Retrieval-Augmented Generation: A Comprehensive Survey of Architectures, Enhancements, and Robustness Frontiers

The survey covers retrieval quality, grounding fidelity, and robustness to noisy evidence, all relevant to the draft's entity-mismatch result. Its scope is broader than the draft's controlled comparison.

How to position yourself: Cite it to situate retrieval quality and grounding as distinct RAG challenges; the draft measures their consequences in a particular QA pipeline.

From its abstract: it introduces new challenges in retrieval quality, grounding fidelity, pipeline efficiency, and robustness against noisy or adversarial inputs.

Survey of Hallucination in Natural Language Generation

This survey provides a broad account of hallucination measurement and mitigation, including generative question answering. It is useful for positioning the draft's operational definition of unsupported output.

How to position yourself: Cite it for the broader hallucination problem and evaluation landscape; the draft focuses specifically on retrieved-evidence support in open-domain QA.

From its abstract: a general overview of metrics, mitigation methods, and future directions

A Comprehensive Survey of Retrieval-Augmented Generation (RAG): Evolution, Current Landscape and Future Directions

This survey explains RAG architectures and their application to knowledge-intensive tasks. It offers general context but does not describe the draft's grounding metric or entity-mismatch experiment.

How to position yourself: Cite it for the broader development and architecture of RAG; the draft supplies a focused empirical analysis of unsupported answer sentences.

From its abstract: The study explores the basic architecture of RAG, focusing on how retrieval and generation are integrated to handle knowledge-intensive tasks.

An Intelligent Biomedical Retrieval Pipeline for Reducing Hallucination Risk in Retrieval Augmented Generation

This biomedical retrieval pipeline addresses weak or missing evidence as a source of hallucination risk. Unlike the draft, it evaluates evidence retrieval rather than support in generated answers.

How to position yourself: Cite it for retrieval-side strategies to improve evidence availability; the draft directly measures generated-answer support in open-domain QA.

From its abstract: These findings suggest lower hallucination risk through better evidence access rather than providing direct end-to-end proof on generated answers.

To Retrieve or Not to Retrieve? Uncertainty Detection for Dynamic Retrieval Augmented Generation

This paper studies uncertainty-based decisions about when to retrieve in long-form QA. It offers a possible extension to the draft's retrieval setup rather than directly testing sentence-level grounding gains.

How to position yourself: Cite it for adaptive retrieval as an alternative to deterministic retrieval; the draft isolates grounding differences between retrieval and no-retrieval systems.

From its abstract: dynamically performing retrieval by invoking it only when the underlying LLM lacks the required knowledge can be more efficient.

Improve Dense Passage Retrieval with Entailment Tuning

This work uses entailment to improve dense retrieval, whereas the draft uses entailment to evaluate generated sentences. It suggests a retrieval-side method that could improve the availability of supporting passages.

How to position yourself: Cite it for entailment-aware dense retrieval as a potential improvement; the draft applies NLI after generation rather than during retriever training.

From its abstract: We observed that a major class of relevance aligns with the concept of entailment in NLI tasks.

Retrieval-Augmented Generation for Question Answering and Beyond: A State-of-the-Art Review

This review covers RAG architectures, retrieval strategies, and applications to QA. It provides general background rather than direct evidence about unsupported sentences or entity mismatch.

How to position yourself: Cite it for the broader RAG design landscape; the draft focuses on a specific dense-retrieval and BART pipeline with entailment-based evaluation.

From its abstract: The primary RAG concepts have been systematically summarized in this paper, including system architecture, retrieval strategies, embedding techniques, reranking strategies and knowledge-aware generation frameworks.

Stable-RAG: Mitigating Retrieval-Permutation-Induced Hallucinations in Retrieval-Augmented Generation

Stable-RAG shows that answer reliability can depend on passage ordering even when the retrieved set contains the gold document. This complements the draft's finding that evidence composition, specifically entity mismatch, can eliminate retrieval gains.

How to position yourself: Cite it for another retrieval-context failure mode beyond missing evidence; the draft tests entity identity rather than document-order sensitivity.

From its abstract: LLM answers vary substantially across permutations of the retrieved set, even when the gold document is fixed in the first position.

Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering

This paper improves customer-service QA using knowledge-graph retrieval that preserves relationships within historical issues. It concerns retrieval quality, but neither its domain nor its stated evaluation directly addresses unsupported sentences in open-domain QA.

How to position yourself: Cite it only as an application illustrating structure-aware retrieval, distinguishing the draft's Wikipedia passages and entailment-based grounding evaluation.

From its abstract: This integration of a KG not only improves retrieval accuracy by preserving customer service structure information but also enhances answering quality