mlsanityData to 14 Oct 2026

When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs

arXiv 2609.15106 ↗First posted 14 Sep 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

Large language models can hallucinate even when the knowledge required for a correct answer is already available. We study this failure through a latent-key view of inference, where answer selection depends on competition among associations acquired during pretraining. We show that model predictions can be highly sensitive to individual query keywords, that these influential keywords exhibit entity-specific binding, and that their effects are systematically shaped by pretraining frequency. Multiple bindings can also compete and exhibit higher-order interactions within the same query. Based on this mechanism, we introduce a two-stage keyword-perturbation method for hallucination detection. By removing influential keywords and measuring how the model reorganizes its prediction, the method distinguishes errors caused by misleading key associations from correct decisions supported by diagnostic evidence. Across multiple models and benchmarks, perturbation provides a strong and transferable detection signal, reaching 0.910 AUROC on probe-known ScientistQA. Finally, we extend the same probabilistic framework to four hallucination regimes: knowledge deficit, wrong knowledge, context distraction, and unstable inference. Their operational distributions across benchmarks provide diagnostic context for why different detector families succeed in different settings.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 29 of its 34 references that are in the map. First posted on arXiv 14 Sep 2026.

Builds on

29 of 34 references.

References · 29

  1. 2023GPT-4 Technical Reportcited by 18,396
  2. 2021TruthfulQA: Measuring How Models Mimic Human Falsehoodscited by 2,975
  3. 2022Language Models (Mostly) Know What They Knowcited by 1,856
  4. 2021WebGPT: Browser-assisted question-answering with human feedbackcited by 1,745
  5. 2023FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generationcited by 1,282
  6. 2022When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memoriescited by 1,128
  7. 2023SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Modelscited by 810
  8. 2023The Internal State of an LLM Knows When It's Lyingcited by 659
  9. 2022Large Language Models Struggle to Learn Long-Tail Knowledgecited by 458
  10. 2023HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Modelscited by 341
  11. 2024INSIDE: LLMs' Internal States Retain the Power of Hallucination Detectioncited by 323
  12. 2024MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documentscited by 257
17 more
  1. 2023Sources of Hallucination by Large Language Models on Inference Taskscited by 169
  2. 2024Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Mapscited by 142
  3. 2023WebGLM: Towards An Efficient Web-Enhanced Question Answering System with Human Preferencescited by 100
  4. 2024HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detectioncited by 99
  5. 2024ContextCite: Attributing Model Generation to Contextcited by 98
  6. 2025Hallucination Detection in LLMs Using Spectral Features of Attention Mapscited by 43
  7. 2025ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMscited by 32
  8. 2025The Illusion of Progress: Re-evaluating Hallucination Detection in LLMscited by 22
  9. 2025Chain-of-Thought Prompting Obscures Hallucination Cues in Large Language Models: An Empirical Evaluationcited by 21
  10. 2025Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associationscited by 15
  11. 2024Co-occurrence is not Factual Association in Language Modelscited by 12
  12. 2025SeSE: Black-Box Uncertainty Quantification for Large Language Models Based on Structural Information Theorycited by 7
  13. 2024Enhancing Hallucination Detection through Perturbation-Based Synthetic Data Generation in System Responsescited by 5
  14. 2026Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMscited by 3
  15. 2025Shaking to Reveal: Perturbation-Based Detection of LLM Hallucinationscited by 2
  16. 2026Understanding Why Language Models Hallucinate: Testing Reasoning Against Priorscited by 1
  17. 2026Detecting Hallucinations in Retrieval-Augmented Generation through Grounding-Aware Sensitivity by Perturbation (GASP)cited by 1

Built on by

Not cited yet.