12,445 papers · continuously updated · last export: 10 Aug 2026livingmeta.ai
← Browse all papers
AI evidence extraction

From Illusion to Insight: A Taxonomic Survey of Hallucination Mitigation Techniques in LLMs

Ioannis Kazlaris, Efstathios Antoniou, Konstantinos Diamantaras, Charalampos Bratsas · AI · 2025

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10
Relevance
2/4
Quality (LMQS)
C
Evidence
1
Citations
1.70
FWCI
Top 10%
Impact

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.3390/ai6100260

Methodology & findings

Study design

Systematic literature review following PRISMA 2020 guidelines with 5-stage hierarchical framework: (1) Literature Retrieval from major electronic archives (Google Scholar, ACM Digital Library, IEEE Xplore, Elsevier, Springer, ArXiv) with cutoff date 12 August 2025; (2) Two-stage screening process (title/abstract review followed by full-text assessment); (3) Paper-level tagging assigning mitigation strategies; (4) Thematic clustering consolidating tags into six broad categories; (5) Content-specific retrieval using a custom RAG system based on Mistral language model.

Main result

The paper reveals that "a consistent pattern emerges: Training-and-learning methods buy reliability upfront: they are expensive offline but light at inference time, and they tend to travel well across tasks when supervision encodes uncertainty rather than certainty. Architectural modifications give direct control-steering attention, shaping decoding, or adding retrieval-yet they live or die by the timing and quality of what is injected into the model. Prompt and input optimization is the most economical knob to turn and often delivers surprising gains, but it is sensitive to phrasing and can drift over long conversations." Additionally, the taxonomy identifies that "Prompt Engineering dominates due to its low cost, flexibility, and ability to guide model behavior without retraining" and that "Self-reflection appears as an emergent ability mainly at larger scales (≈70B params); in smaller models it can backfire-second-guessing correct answers and increasing hallucinations."

Research paradigm

Positivist/Empiricist - systematic classification and categorization of computational mitigation techniques

Author conclusions

The authors conclude: "Our review reveals that large language model hallucinations are caused by interacting mechanisms rather than a single failure. The originality of our work lies in a method-oriented taxonomy that organizes mitigation strategies based on the operational levers they activate across the model lifecycle-including training, architecture, decoding, retrieval, quality control, and diagnostics-instead of by task or application. This framework goes beyond simple classification; it exposes shared interfaces between methods, clarifies hybrid techniques, and converts a previously scattered literature into a coherent design space. Ultimately, by mapping these mechanisms to evaluation targets and deployment guardrails, the taxonomy acts as an analytical scaffold that enables cumulative and testable progress, making it a dynamic tool instead of a static survey." The authors further propose: "Based on this analytical framework, we propose a nine-point research agenda focused on reducing large language model hallucinations and improving trustworthiness: Standardized evaluation must first be developed, moving toward uncertainty-aware metrics that explicitly link factuality with calibration, requiring the model to either abstain or show its sources for every claim."

Risk of bias

Language bias: primarily English-language sources (non-English papers may be underrepresented); Publication bias: peer-reviewed and preprint sources may overrepresent positive findings; Temporal bias: cutoff date of 12 August 2025 may exclude recent developments; Selection bias: exclusion of non-academic sources (blogs, opinion pieces) may miss practitioner insights; Method attribution bias: papers implementing multiple strategies make causal attribution difficult; Task-specificity bias: hallucination manifestations vary by task, limiting generalizability; Author selection: review conducted by specific research team without explicit inter-rater reliability metrics; Selection bias: Literature restricted to peer-reviewed sources; exclusion of non-academic sources (blogs, opinion pieces) may miss practitioner insights; Language bias: Cutoff date of 12 August 2025 means very recent work may be missed; search limited to English-indexed databases; Tagging bias: Multiple tags assigned to papers implementing multiple methodologies may introduce subjective classification; Publication bias: Preference for published work over null/negative results; Temporal bias: Sharp rise in publications since 2022 may skew taxonomy toward recent approaches; Selection bias: restriction to peer-reviewed sources and English-language literature may exclude relevant non-English or preprint work; Classification bias: subjective assignment of papers to categories when methods employ multiple strategies; Publication bias: likely toward positive results and published research, excluding null findings or unpublished negative results; Temporal bias: cutoff date of August 2025 may miss concurrent work; Database bias: reliance on major academic databases may miss grey literature or industry publications

Limitations

  • The authors state: "A key difficulty is that the majority of the papers implement a number of strategies (oftentimes referred to as frameworks), and therefore it is not always clearly discernible what the primary mitigation method is and to what extent its results can be safely attributed to it." Additionally, "Defining hallucinations is challenging because their classification depends heavily on the task, the logical relation between input and output, and pragmatic factors such as ambiguity, vagueness, presupposition, or metaphor." The review also notes: "Despite significant progress in LLM development, hallucinatory outputs continue to raise serious concerns about the reliability and trustworthiness of machine-generated text
  • These risks are particularly acute in critical domains such as healthcare, law and defense, where considerable harm may be caused to individuals."

Open questions raised

  • Knowledge Limitations and Updating: Hallucinations occur due to outdated or incomplete training data; more robust strategies for continuous knowledge updating and noise-aware retrieval needed
  • Computational Trade-offs and Latency Constraints: Many mitigation strategies add significant computational overhead; future research must focus on efficient mitigation techniques balancing reliability with speed
  • Robustness Under Distribution Shifts and Adversarial Inputs: LLMs vulnerable to hallucinations under data/prompt distribution shifts; improved uncertainty estimation and new training methods required
  • Interpretability and Attribution Difficulties: Challenging to understand how transformers create hallucinations; improved interpretability tools and causal analysis crucial
  • Lack of Standardized Benchmarks and Metrics: Hallucination evaluation fragmented with ad hoc datasets and inconsistent metrics; shared benchmarks including various hallucination types and languages needed
  • Ethical and Epistemological Concerns: Need to navigate balance between factual reliability and creative generation; requires both technical safeguards and governance frameworks
Data: ANAH: Bilingual hallucination dataset with sentence-level annotations; BoolQ: Yes/no question-answering dataset; DiaHalu: First dialogue-level hallucination evaluation benchmark; FACTOR: Factuality Assessment Corpus for Text and Reasoning; FACTSCORE: Fine-grained, atomic-level factuality metric dataset; FELM: Factuality Evaluation of Large Language Models benchmark; FEVER: 185,445-claim fact extraction and verification dataset; FRANK: Fine-grained error annotations for abstractive summarization; HADES: Reference-free hallucination-detection dataset for QA; HalluEditBench: Verified hallucinations across multiple domains; HalluLens: Hallucination-focused benchmark with intrinsic and extrinsic tasks; HALOGEN: Multi-domain hallucination benchmark; HaluEval: Large-scale hallucination tendency benchmark; HaluEval 2.0: Enhanced version with 8770 questions from diverse domains; HaluEval-Wild: Real-world user interaction hallucination evaluation; HDMBench: Hallucination detection benchmark across knowledge-intensive tasks; Head-to-Tail: 18,000 question-answer pairs segmented by knowledge popularity; HotpotQA: Multi-hop reasoning and information retrieval capability evaluation; LegalBench: Legal reasoning and interpretation benchmark; NQ (Natural Questions): Large-scale real questions with Wikipedia answers; PubMedQA: Medical question answering benchmark; RAGTruth: Nearly 18,000 naturally generated responses for word-level hallucination analysis in RAG; SelfCheckGPT: Zero-resource, black-box hallucination detection benchmark; TriviaQA: 650,000+ question-answer-evidence triplets; TruthfulQA: Benchmark for adversarial truthfulness assessment; UHGEval: Large-scale Chinese-language unconstrained generation hallucination benchmark; ANAH - bilingual hallucination dataset with sentence-level annotations; BoolQ - question answering dataset; DiaHalu - dialogue-level hallucination evaluation benchmark; FACTOR - Factuality Assessment Corpus for Text and Reasoning; FACTSCORE - fine-grained, atomic-level factuality metric; FELM - Factuality Evaluation of Large Language Models benchmark; FEWL - Factuality Evaluation Without Labels; FRANK - fine-grained error annotations for summarization; HADES - HAllucination DEtection dataset for QA; HalluEditBench - hallucinations across multiple domains; HalluLens - hallucination-focused benchmark; HALOGEN - multi-domain hallucination benchmark; HaluEval - large-scale hallucination evaluation benchmark; HaluEval 2.0 - enhanced version with 8770 questions; HaluEval-Wild - real-world user interactions; HDMBench - hallucination detection benchmark; Head-to-Tail - 18,000 question-answer pairs; HotpotQA - multi-hop reasoning dataset; NQ (Natural Questions) - large-scale Google questions with Wikipedia answers; PubMedQA - medical question answering benchmark; RAGTruth - dataset for analyzing word-level hallucinations in RAG frameworks; SelfCheckGPT - zero-resource hallucination detection benchmark; TriviaQA - 650,000+ question-answer-evidence triplets; TruthfulQA - benchmark for truthfulness assessment; UHGEval - Chinese-language hallucination evaluation benchmark; FEVER - 185,445-claim fact extraction and verification dataset; LegalBench - legal reasoning and interpretation; HumanEval - code generation benchmark; HaluEval; HaluEval 2.0; HaluEval-Wild; ANAH (Analytically Annotated Hallucination dataset); TruthfulQA; FEVER (Fact Extraction and Verification); HotpotQA; Natural Questions (NQ); TriviaQA; BoolQ; PubMedQA; DiaHalu; HADES (HAllucination DEtection dataset); FACTOR (Factuality Assessment Corpus for Text and Reasoning); FRANK benchmark; FELM (Factuality Evaluation of Large Language Models); FEWL (Factuality Evaluation Without Labels); HalluEditBench; HalluLens; HALOGEN (Hallucinations of Generative Models); HDMBench; Head-to-Tail; UHGEval; SciBench; LegalBench; HumanEval (Code Generation); Codex; RAGTruthCode: https://osf.io/zbdy2 (OSF registration for the systematic review); Custom Retrieval-Augmented Generation (RAG) system based on Mistral language model (mentioned but no specific repository provided)Extracted from: pdfAgreement 48%

Explore related topics

Related papers