HalluWorld is a controlled benchmark using explicit reference world models to automatically label and disentangle hallucinations in LLMs across synthetic environments with varying complexity and observability.
Canonical reference
Survey of Hallucination in Natural Language Generation
Canonical reference. 88% of citing Pith papers cite this work as background.
55
Pith papers citing it
2,906
external citations · Crossref
Background
88% of classified citations
citation-role summary
background 17
citation-polarity summary
claims ledger
- background [315, 361]. Furthermore, Liu et al. [185], Zong et al. [395] and Liu et al. [184] show that LVLMs can be easily fooled and experience a severe performance drop due to their over-reliance on the strong language prior, as well as its inferior ability to defend against inappropriate user inputs [112, 134]. Jiang et al. [138], Wang et al. [315] and Jing et al. [141] took a step forward to holistically evaluate multi-modal hallucination. What's more, when presented with multiple images, LVLMs sometim
co-cited works
roles
background 17representative citing papers
The Spectral Sensitivity Theorem identifies a phase transition in Whisper models where scaling causes self-attention to collapse into rank-1 attractors, decoupling output from acoustic evidence.
Introduces the Attribution Contract specification to clarify feature attribution claims in generative language models by naming the output explained, eligible features, generative process, fixed elements, and attributed model score.
New benchmark evaluates three frontier deep research agents on 42 SME prompts with verifiers and rubrics, reporting low acceptance rates of 9.5-21.4% and agent-specific failure modes.
QAOD projects away question-aligned directions from answer representations to isolate domain-agnostic factuality signals, enabling efficient hallucination detection with top in-domain AUROC and up to 21% better OOD transfer.
Malicious Skills induce coding agents to hallucinate and import attacker-controlled packages at high rates while evading detection.
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
Indirect elicitation via triplet comparisons recovers meaningful association structures from LLMs and supports conservative causal candidate links across prompted subpopulations.
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
A graphlet-anchored framework generates 119,856 factually grounded biomedical QA pairs that improve accuracy on PubMedQA and MedQA benchmarks.
CyberCertBench shows frontier LLMs reach human-expert performance on general IT and networking security but drop on vendor-specific and formal standards questions such as IEC 62443, with a new framework for producing interpretable explanations.
Frontier LLMs generate BibTeX entries at 83.6% field accuracy but only 50.9% fully correct; two-stage clibib revision raises accuracy to 91.5% and fully correct entries to 78.3% with 0.8% regression.
A study of seven LLMs finds that realistic prompt variations such as one-character misspellings trigger library hallucinations in up to 26% of cases, fabricated names in up to 99%, and time-based prompts in up to 85%, and introduces LibHalluBench for evaluation.
Measurement of 688 AI infra repositories shows frequent overlapping vulnerable patterns, and INFRASCOPE detects over 20 variants including 11 acknowledged and 4 with new CVEs.
A vector generalization of fusion-fission group dynamics from physics forecasts when AI behavior shifts to undesirable states, validated at 90 percent across seven models and prior to real-world data.
Derivation Prompting constructs logic-based derivation trees in RAG generation to improve interpretability and reduce unacceptable answers compared to standard RAG or long-context methods in a case study.
Introduces a trust-boundary architecture in Lean 4 with three certificate families and two operators that deliver sorry-free, axiom-audited assurances for LLM pipeline components.
LaaB improves LLM hallucination detection by mapping self-judgment labels back into neural feature space and using mutual learning under logical consistency constraints between responses and meta-judgments.
CuraView detects sentence-level faithfulness hallucinations in medical discharge summaries via GraphRAG knowledge graphs and multi-agent evidence grading, achieving 0.831 F1 on critical contradictions with a fine-tuned Qwen3-14B model and 50% relative improvement over baselines.
Adaptive Unlearning suppresses package hallucinations in code-generating LLMs by 81% while preserving benchmark performance, using model-generated data and no human labels.
AI peer review systems are vulnerable to prompt injections, prestige biases, assertion strength effects, and contextual poisoning, as demonstrated by a new attack taxonomy and causal experiments on real conference submissions.
Token-level contrastive attribution yields informative signals for some LLM benchmark failures but is not universally applicable across datasets and models.
A hybrid graph-text retrieval system for cyber threat intelligence improves multi-hop question answering by up to 35% over vector-based RAG on a 3,300-question benchmark.
Narrix helps novices identify and reuse narrative strategies from examples through visualization and strategy-steered generation, improving retention, confidence, and adaptation over chat interfaces in a 12-person study.
citing papers explorer
-
HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models
HalluWorld is a controlled benchmark using explicit reference world models to automatically label and disentangle hallucinations in LLMs across synthetic environments with varying complexity and observability.
-
From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
The Spectral Sensitivity Theorem identifies a phase transition in Whisper models where scaling causes self-attention to collapse into rank-1 attractors, decoupling output from acoustic evidence.
-
The Attribution Contract: Feature Attribution for Generative Language Models
Introduces the Attribution Contract specification to clarify feature attribution claims in generative language models by naming the output explained, eligible features, generative process, fixed elements, and attributed model score.
-
Evaluating Deep Research Agents on Expert Consulting Work: A Benchmark with Verifiers, Rubrics, and Cognitive Traps
New benchmark evaluates three frontier deep research agents on 42 SME prompts with verifiers and rubrics, reporting low acceptance rates of 9.5-21.4% and agent-specific failure modes.
-
When Answers Stray from Questions: Hallucination Detection via Question-Answer Orthogonal Decomposition
QAOD projects away question-aligned directions from answer representations to isolate domain-agnostic factuality signals, enabling efficient hallucination detection with top in-domain AUROC and up to 21% better OOD transfer.
-
Trust Me, Import This: Dependency Steering Attacks via Malicious Agent Skills
Malicious Skills induce coding agents to hallucinate and import attacker-controlled packages at high rates while evading detection.
-
Causal Stories from Sensor Traces: Auditing Epistemic Overreach in LLM-Generated Personal Sensing Explanations
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
-
Eliciting associations between clinical variables from LLMs via comparison questions across populations
Indirect elicitation via triplet comparisons recovers meaningful association structures from LLMs and supports conservative causal candidate links across prompted subpopulations.
-
Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
-
BioGraphletQA: Knowledge-Anchored Generation of Complex QA Datasets
A graphlet-anchored framework generates 119,856 factually grounded biomedical QA pairs that improve accuracy on PubMedQA and MedQA benchmarks.
-
CyberCertBench: Evaluating LLMs in Cybersecurity Certification Knowledge
CyberCertBench shows frontier LLMs reach human-expert performance on general IT and networking security but drop on vendor-specific and formal standards questions such as IEC 62443, with a new framework for producing interpretable explanations.
-
BibTeX Citation Hallucinations in Scientific Publishing Agents: Evaluation and Mitigation
Frontier LLMs generate BibTeX entries at 83.6% field accuracy but only 50.9% fully correct; two-stage clibib revision raises accuracy to 91.5% and fully correct entries to 78.3% with 0.8% regression.
-
Library Hallucinations in LLM-Generated Code: A Risk Analysis Grounded in Developer Queries
A study of seven LLMs finds that realistic prompt variations such as one-character misspellings trigger library hallucinations in up to 26% of cases, fabricated names in up to 99%, and time-based prompts in up to 85%, and introduces LibHalluBench for evaluation.
-
Hunting Vulnerability Variants in AI Infra: Measurement and Reference-Driven Detection
Measurement of 688 AI infra repositories shows frequent overlapping vulnerable patterns, and INFRASCOPE detects over 20 variants including 11 acknowledged and 4 with new CVEs.
-
Fusion-fission forecasts when AI will shift to undesirable behavior
A vector generalization of fusion-fission group dynamics from physics forecasts when AI behavior shifts to undesirable states, validated at 90 percent across seven models and prior to real-world data.
-
Derivation Prompting: A Logic-Based Method for Improving Retrieval-Augmented Generation
Derivation Prompting constructs logic-based derivation trees in RAG generation to improve interpretability and reduce unacceptable answers compared to standard RAG or long-context methods in a case study.
-
Proof-Carrying Certificates for LLM Pipelines: A Trust-Boundary Architecture
Introduces a trust-boundary architecture in Lean 4 with three certificate families and two operators that deliver sorry-free, axiom-audited assurances for LLM pipeline components.
-
Logical Consistency as a Bridge: Improving LLM Hallucination Detection via Label Constraint Modeling between Responses and Self-Judgments
LaaB improves LLM hallucination detection by mapping self-judgment labels back into neural feature space and using mutual learning under logical consistency constraints between responses and meta-judgments.
-
CuraView: A Multi-Agent Framework for Medical Hallucination Detection with GraphRAG-Enhanced Knowledge Verification
CuraView detects sentence-level faithfulness hallucinations in medical discharge summaries via GraphRAG knowledge graphs and multi-agent evidence grading, achieving 0.831 F1 on critical contradictions with a fine-tuned Qwen3-14B model and 50% relative improvement over baselines.
-
LLM Ghostbusters: Surgical Hallucination Suppression via Adaptive Unlearning
Adaptive Unlearning suppresses package hallucinations in code-generating LLMs by 81% while preserving benchmark performance, using model-generated data and no human labels.
-
When AI reviews science: Can we trust the referee?
AI peer review systems are vulnerable to prompt injections, prestige biases, assertion strength effects, and contextual poisoning, as demonstrated by a new attack taxonomy and causal experiments on real conference submissions.
-
Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks
Token-level contrastive attribution yields informative signals for some LLM benchmark failures but is not universally applicable across datasets and models.
-
Beyond RAG for Cyber Threat Intelligence: A Systematic Evaluation of Graph-Based and Agentic Retrieval
A hybrid graph-text retrieval system for cyber threat intelligence improves multi-hop question answering by up to 35% over vector-based RAG on a 3,300-question benchmark.
-
Narrix: Remixing Narrative Strategies from Examples for Story Writing
Narrix helps novices identify and reuse narrative strategies from examples through visualization and strategy-steered generation, improving retention, confidence, and adaptation over chat interfaces in a 12-person study.
-
Adaptive Residual-Update Steering for Low-Overhead Hallucination Mitigation in Large Vision Language Models
RUDDER creates a persistent visual anchor by extracting CARD from prefill residuals and modulating its injection via an adaptive Beta Gate, cutting CHAIR_S by 24.4% and CHAIR_i by 23.6% on average across LLaVA, Idefics2, InstructBLIP and Qwen2.5-VL with >96% throughput.
-
ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards
ReSeek adds self-correction via a JUDGE action and a dense instructive reward (correctness plus utility) to RL training of search agents, yielding higher success and faithfulness on a new contamination-resistant benchmark.
-
Conversational AI increases political knowledge as effectively as self-directed internet search
Conversational AI matches self-directed internet search in increasing belief in true political information and decreasing belief in misinformation.
-
RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
RAPTOR introduces a tree-organized retrieval method using recursive abstractive summaries, achieving a 20% absolute accuracy improvement on the QuALITY benchmark when paired with GPT-4.
-
Corrective Retrieval Augmented Generation
CRAG improves RAG robustness via a retrieval quality evaluator that triggers web augmentation and a decompose-recompose filter to focus on relevant information, yielding better results on short- and long-form generation tasks.
-
SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
SelfCheckGPT detects hallucinations by checking consistency across multiple sampled responses from black-box LLMs on WikiBio biography generation tasks.
-
ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking
ReacTOD introduces a bounded neuro-symbolic ReAct architecture with symbolic validation that delivers new zero-shot SOTA joint goal accuracy on MultiWOZ 2.1 and strong results on SGD.
-
Fairness-Aware Retrieval Optimization for Retrieval-Augmented Generation
Introduces FARO, a scalable quadratic optimization approach for fairness-aware top-k retrieval in RAG that mitigates generation bias via controlled reranking and position-aware propagation modeling.
-
The Semantic Training Gap: Ontology-Grounded Tool Architectures for Industrial AI Agent Systems
Ontology-grounded tool architectures eliminate hallucination of domain identifiers in industrial AI agents by enforcing semantic constraints through a typed relational configuration and three-operation interface.
-
Evaluating the False Trust Engendered by LLM Explanations
LLM reasoning traces and post-hoc explanations increase false trust in incorrect predictions, whereas contrastive dual explanations enhance users' ability to distinguish correct from incorrect AI outputs.
-
From Flat Facts to Sharp Hallucinations: Detecting Stubborn Errors via Gradient Sensitivity
EPGS detects high-confidence factual errors in LLMs by using embedding perturbations to measure gradient sensitivity as a proxy for sharp versus flat minima.
-
An Information-Geometric Framework for Stability Analysis of Large Language Models under Entropic Stress
A thermodynamic-inspired information-geometric framework defines a composite LLM stability score that outperforms a utility-entropy baseline by 0.0299 on average across 80 observations, with gains increasing at higher entropy.
-
Self-Awareness before Action: Mitigating Logical Inertia via Proactive Cognitive Awareness
SABA improves LLM performance on detective puzzle benchmarks by recursively fusing information into a base state and using queries to resolve missing premises before concluding.
-
Towards Scalable Lifelong Knowledge Editing with Selective Knowledge Suppression
LightEdit enables scalable lifelong knowledge editing in LLMs via selective knowledge retrieval and probability suppression during decoding, outperforming prior methods on ZSRE, Counterfact, and RIPE while reducing training costs.
-
Align Documents to Questions: Question-Oriented Document Rewriting for Retrieval-Augmented Generation
QREAM rewrites documents to question-focused style using iterative ICL and distilled FT models, boosting RAG performance by up to 8% relative improvement.
-
A Graph-Enhanced Defense Framework for Explainable Fake News Detection with LLM
G-Defense builds claim-centered graphs from sub-claims, applies RAG for evidence and competing explanations, then uses graph inference to detect fake news veracity and generate intuitive explanation graphs, claiming SOTA results.
-
The Personalization Paradox: Semantic Loss vs. Reasoning Gains in Agentic AI Q&A
Personalization in an agentic RAG advising system boosts reasoning quality and grounding while reducing semantic metric scores due to the inability of current metrics to accommodate user-specific responses.
-
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
The paper surveys hallucination in LLMs with an innovative taxonomy, factors, detection methods, benchmarks, mitigation strategies, and open research directions.
-
PaLM 2 Technical Report
PaLM 2 reports state-of-the-art results on language, reasoning, and multilingual tasks with improved efficiency over PaLM.
-
StarCoder: may the source be with you!
StarCoderBase matches or beats OpenAI's code-cushman-001 on multi-language code benchmarks; the Python-fine-tuned StarCoder reaches 40% pass@1 on HumanEval while retaining other-language performance.
-
Do Benchmarks Underestimate LLM Performance? Evaluating Hallucination Detection With LLM-First Human-Adjudicated Assessment
Human adjudication of conflicts between original benchmark labels and LLM predictions on QAGS-C and SummEval increases triple agreement by 6-8% and LLM accuracy by 2-9%, with adjudicators often siding with models that provide explicit reasoning.
-
Text-Graph Synergy: A Bidirectional Verification and Completion Framework for RAG
TGS-RAG adds graph-to-text re-ranking with global voting and text-to-graph orphan path bridging to improve precision and efficiency in multi-hop RAG over prior baselines.
-
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
HUMBR reduces LLM hallucinations in enterprise workflows by using a hybrid semantic-lexical utility within minimum Bayes risk decoding to identify consensus outputs, with derived error bounds and reported outperformance over self-consistency on benchmarks and production data.
-
Council Mode: A Heterogeneous Multi-Agent Consensus Framework for Reducing LLM Hallucination and Bias
Council Mode reduces LLM hallucinations by 35.9% and improves TruthfulQA scores by 7.8 points through parallel heterogeneous model generation followed by structured consensus synthesis.
-
Mitigating Hallucination on Hallucination in RAG via Ensemble Voting
VOTE-RAG applies retrieval voting across diverse queries and response voting across independent generations to mitigate hallucination-on-hallucination in RAG, matching or exceeding complex baselines on six benchmarks with a parallelizable design.
-
Using Large Language Models in Physics Education
Frontier LLMs from late 2025 reach near-perfect scores on text-based physics problem solving and show improved human-grading alignment, yet still struggle to assign partial credit for flawed reasoning.