What Hallucinations Have We Measured? Semantic Illusion, Confabulation, and Expressed Uncertainty

1 INTRODUCTION

• Inconsistent use of “hallucination” spans factual, faithfulness, reasoning, and uncertainty errors, obscuring causes, evaluation validity, and targeted mitigation.Source: §1 · p. 2
• The taxonomy defines hallucinations as confident incorrect responses, separates expressed uncertainty (UNK), and labels errors SI if equivalent re-elicitation recovers the answer or CF if none does.The paper re-evaluates 14 hallucination methods using this taxonomy.Source: §1 · p. 2
• Methods mitigate 28–29% of detected semantic illusions but 0.7–2.8% of confabulations; external knowledge raises confabulation mitigation to about 20%.The paper predicts that semantic illusions can yield to inexpensive inference-time changes, while confabulations may require additional knowledge.Source: §1 · p. 2
• Uncertainty-based detectors score expressed uncertainty above both hallucination types and miss half of the confabulations that remain after uncertainty is elicited.Source: §1 · p. 2
• The useful uncertainty representation varies by error type: verbal signals detect UNK best, latent signals detect and mitigate CF best, and semantic signals mitigate SI best.Source: §1 · p. 2
• A labeling protocol stabilizes at about ten paraphrases, while aggregate hallucination scores hide distinctions among error types; the paper reports that most recent papers conflate them.Source: §1 · p. 2

2.1 RQ1 (INCORRECTNESS): DO HALLUCINATIONS ARISE FROM DISTINCT ERROR TYPES?

• Incorrect responses may be semantic illusions, recoverable under equivalent phrasing, or confabulations, which persist across phrasing; the distinction may affect which detection and mitigation strategies apply.A correct response under an equivalent condition indicates behavioral accessibility without changing model weights or adding information.Source: §2.1 · p. 3

2.2 RQ2 (PLAUSIBILITY): DOES HALLUCINATION REQUIRE CONFIDENT INCORRECTNESS?

• Hallucination is defined as confident incorrectness, distinct from expressed uncertainty, because a confident error gives readers no cue to check it.The paper treats confidence as the absence of explicitly expressed uncertainty.Source: §2.2 · p. 3

3 WHAT THE COMMUNITY MEASURES

• Only 35% of surveyed papers define hallucination as confident or plausible error, while 95.8% of surveyed authors favor distinguishing it from uncertainty and 16.7% report doing so.The audit covers text-based papers with “hallucination” in the title from five major venues between 2021 and 2025, and surveys 24 first authors.Source: §3 · p. 3
• Among 91 papers concluding external knowledge is needed, 15% did so, and 11% of all 91 had not tested whether rephrasing, instruction, or decoding changes removed the error.79.2% of surveyed respondents ranked knowledge availability—whether a model lacks a fact or misapplies it—as the most important hallucination dimension.Source: §3 · p. 3

4 TERMINOLOGY

• The paper proposes separating hallucinations from uncertainty and dividing hallucinations into semantic illusions and confabulations according to whether re-elicitation can recover a correct answer.The authors hypothesize that this distinction bears on which detection and mitigation strategies apply.Source: §4 · p. 4

4.1 CONFABULATIONS

• A confabulation persists across inference-time changes and is hypothesized to reflect absent or inaccessible knowledge, including missing training information or relevant information that remains underlearned.The paper describes these as distinct mechanisms and says such errors resist prompt engineering and inference-time decoding changes.Source: §4.1 · p. 4

4.2 SEMANTIC ILLUSIONS

• A semantic illusion occurs when an incorrect response can be corrected for the same factual knowledge without changing model parameters; the paper relates this behavior to surface-level semantic compatibility.The Moses/Noah example illustrates choosing a wrong answer despite knowing the fact and answering correctly under another phrasing.Source: §4.2 · p. 5

5.1 DEFINITIONS

• A question answered correctly under its original condition is labeled NH; otherwise, expressed uncertainty under any tested condition yields UNK, while CF and SI are hallucinations.Labels are model-specific and relative to the tested condition set; enlarging that set can move CF to SI but not the reverse.Source: §5.1 · p. 5

5.2 LABELING PROTOCOL AND VALIDATION

• About ten paraphrases stabilize labels: agreement with full-pool labels reaches κ ≈0.8, while CF→SI reassignment flattens as the tested set grows.Figure 3 tracks correctness-rate entropy, CF→SI reassignment, and agreement and flip rate against full-pool labels; the recommended window is shaded.Source: §5.2 · p. 6, Figure 3 · p. 6
• At T=0.7, p=0.95, Llama labels agree across conditions 85% of the time although exact response text agrees only 34%.The labeling evaluation uses 30 paraphrases, instruction variants, and five decoding settings; methods are not shown the labeling schemes.Source: §5.2 · p. 6
• For SI and CF questions, paraphrase-added information is centered near zero, yet only SI questions are ever answered correctly; median ρ30 is 0.17 for Gemma and 0.13 for Llama.Figure 4 relates ρ30 to the information quantity added by paraphrases and reports agreement under decoding changes.Source: §5.2 · p. 6, Figure 4 · p. 6

TAKEAWAY

• Use about ten paraphrases under greedy decoding as a practical approximation, report label-convergence curves as a sufficiency check, and treat CF rate as an upper bound on behavioral inaccessibility.Adding variants can move CF cases to SI; it cannot move SI cases back to CF.Source: §5.1 · p. 5, TAKEAWAY · p. 6

6 EXPERIMENTAL SETUP

• Fourteen methods are grouped by two main-text dimensions, with other groupings analyzed in §F.Implementations were verified with the original papers’ authors.Source: Table 1 · p. 7, §6 · p. 7
• The evaluation covers TriviaQA and SimpleQA on Llama 3.1 8B and Gemma 3 12B, using closed-book QA to make parametric knowledge the answer source.The study design groups methods by prompt sensitivity, intervention stage, knowledge assumptions, uncertainty usage, and uncertainty representation.Source: §6 · p. 7
• Each question is evaluated under 30 paraphrases, instruction variants, and five decoding settings; S2 adds an instruction inviting abstention to S1’s three answer-only instructions.The schemes are hidden from methods, which see only original questions.Source: Table 2 · p. 7
• Detection is measured by recall and residual error rate, while mitigation is measured by the correctness of detected errors.Results are averaged over methods within each group.Source: §6 · p. 7

1 Detected SI responses are mitigated at least ten times as often as detected

• Under S2, CF recall falls 13.6 points to 53.4%, while SI recall remains unchanged; CF also has higher residual error than SI.Under S1, recall is about 67% for both error types.Source: §1 · p. 7, Figure 5 · p. 8
• Only external-knowledge methods mitigate CF well above zero, at about 20%, while inference-time and post-hoc methods stay near zero as |V(Q)| grows.Figure 6 compares gaps by intervention stage across variant-set sizes.Source: §1 · p. 7, Figure 6 · p. 8

TAKEAWAY

• Test re-elicitation before retrieval or retraining: SI may yield to inexpensive prompt or decoding changes, whereas CF calls for new knowledge.The paper frames this as a practical diagnostic for choosing interventions.Source: TAKEAWAY · p. 8

8 FINDING 2: UNCERTAINTY-BASED METHODS DO NOT ALWAYS WORK FOR HALLUCINATIONS

• Uncertainty-based methods give UNK their highest mean score under both schemes—0.61 and 0.57, versus CF at 0.54 and 0.42 and SI at 0.43 and 0.42.The table covers seven uncertainty-estimation detectors; high UNK scores count as failures because UNK is not hallucination.Source: §8 · p. 9, Table 3 · p. 9
• After uncertainty elicitation (S2), detectors miss 50.4% of remaining CF responses; although S2 moves most CF into UNK, those remaining CF are no easier to mitigate.S2 reduces CF’s share among incorrect responses from 0.36 to 0.04.Source: §8 · p. 9

TAKEAWAY

• Report UNK separately; otherwise detection scores can reward identifying uncertain responses while missing confidently incorrect SI and CF.Match verbal signals to UNK, semantic consistency to SI, and latent signals to CF mitigation among the tested uncertainty signals.Source: TAKEAWAY · pp. 9–10

3 Output-level uncertainty detects expressed uncertainty but not

• Verbal signals detect UNK best but CF worst, while latent signals detect CF better than SI and provide the strongest CF detection among uncertainty representations.The figure summarizes detection and mitigation by uncertainty representation as |V(Q)| grows.Source: §3 · p. 9, Figure 7 · p. 10
• Latent signals provide the most CF and UNK mitigation among uncertainty representations, whereas semantic signals provide the best SI mitigation.Methods using no uncertainty mitigate CF better still because that group includes external-knowledge methods.Source: §3 · p. 9

10 RECOMMENDATIONS

• Define hallucination as confident incorrectness and distinguish it from expressed uncertainty; report CF, SI, and UNK separately.The authors argue that conflating these categories risks targeting different phenomena under one label.Source: §10 · p. 10
• Test semantically equivalent phrasings before adding external knowledge, since errors recoverable through rephrasing may not need retrieval augmentation.The authors describe this diagnostic as often inexpensive relative to interventions it can rule out.Source: §10 · p. 10
• Match uncertainty representation to the error type: latent representations are more promising than verbal ones for mitigating CF and UNK.The recommendations also distinguish verbal, latent, and semantic signals by their different performance across error types.Source: §10 · p. 10

11 LIMITATIONS AND CONCLUSION

• The study covers a limited set of models, datasets, and methods, although the SI–CF mitigation gap holds across both models, both labeling schemes, and all variant-set sizes.The authors propose more precise definitions, disaggregated error types, and alignment between definitions and evaluation methods.Source: §11 · p. 10

A RELATED WORK

• The paper classifies hallucinations by prompt sensitivity, distinguishing consistent incorrectness (confabulation) from prompt-sensitive incorrectness (semantic illusion), unlike taxonomies based on triggers or contradictions.The authors retain hallucination as the cover term and define these subtypes through re-elicitation across semantically equivalent prompt variants.Source: A RELATED WORK · p. 22

B AUTHOR SURVEY

• 79.2% of surveyed authors selected knowledge availability and 75.0% selected causal origin as important dimensions for categorizing hallucination.Knowledge source, error type, and prompt sensitivity formed a middle tier (41.0–50.0%), while task type, input modality, and multimodality were rarely selected (8.3–16.7%).Source: B AUTHOR SURVEY · p. 22, Figure 8 · p. 22

C WHAT THE COMMUNITY MEASURES

• Only 16.7% of surveyed papers treated hallucination and uncertainty as distinct, although 95.8% of surveyed researchers said they should be distinguished.Only 35% of papers defined hallucination as confident or plausible errors, and 23% treated hallucination detection as distinct from uncertainty estimation.Source: C WHAT THE COMMUNITY MEASURES · p. 23
• Among papers that concluded external knowledge was needed, 11% of all papers skipped prompt-variance testing, while 3% of all papers that introduced no new knowledge also skipped it.The audit found that 15% of papers concluded new knowledge was necessary and 19% reduced errors without new knowledge; without prompt-variance diagnosis, methods may miss simpler interventions.Source: C WHAT THE COMMUNITY MEASURES · p. 23
• The cited bibliography entries name ToolBeHonest, MedHallu, and FACT, with titles concerning tool-augmented LLM hallucination diagnostics, medical hallucination detection, and fact-driven alternating code-text training.Source: C WHAT THE COMMUNITY MEASURES · pp. 26, 28
• Table 5 identifies papers 45–91 in its inventory of category assignments for text-modality papers.The table caption states that the full inventory covers all 91 text-modality papers.Source: C WHAT THE COMMUNITY MEASURES · p. 26, Table 5 · p. 29

• The term confabulation draws on a pattern of confident false narratives and a clinical analogy involving absent relevant memories.The paper notes that LLMs operate through different mechanisms from humans.Source: D MOTIVATION FOR THE TERMINOLOGY · p. 29
• Confabulation (CF) persists across inference-time changes and is associated with absent training information or relevant information that remains underlearned.The paper describes these as absent and underlearned memory, respectively.Source: D MOTIVATION FOR THE TERMINOLOGY · p. 30
• Semantic illusion (SI) occurs when equivalent inputs can elicit the correct answer without parameter changes, despite an initial error.The relevant information is available, but the model does not consistently apply it.Source: D MOTIVATION FOR THE TERMINOLOGY · p. 30
• Semantic illusion names errors where surface-level semantic compatibility can outweigh factual knowledge encoded in the model.The paper connects this account to human semantic illusions and contextual associations in next-token prediction.Source: D MOTIVATION FOR THE TERMINOLOGY · p. 30

• The paper formalizes incorrectness and plausibility through research questions and hypotheses.Source: E PROBLEM FORMULATION · p. 30

• The section asks whether errors recoverable under semantically equivalent phrasing differ from errors that recur across phrasings, and whether that distinction changes suitable detection and mitigation.Source: §E.1 · pp. 30–31
• Knowledge counts as behaviorally accessible when a tested equivalent phrasing, instruction variant, or decoding configuration elicits a correct answer without added hints.The test does not require access to model internals.Source: §E.1 · p. 31
• CF requires incorrect, non-abstaining answers across every tested variant and plausible inference-time condition, whereas SI requires at least one incorrect and one correct variant.SI classification also excludes parameter changes or additional information.Source: §E.1 · p. 31
• Because CF and SI are defined over a finite variant set, expanding that set can reclassify CF cases as SI, so the labels are operational diagnostics rather than intrinsic properties.The paper states that the mitigation gap remains consistent across different variant schemes.Source: §E.1 · p. 31
• The hypotheses predict that inconsistency-based detectors will detect SI more reliably, while inference-time and post-hoc interventions reduce SI more than CF.They predict that mitigating CF depends on external knowledge.Source: §E.1 · p. 31

E.2 RQ2 (PLAUSIBILITY): DOES HALLUCINATION REQUIRE CONFIDENT INCORRECTNESS?

• Hallucination is confident incorrectness rather than explicit uncertainty; the section hypothesizes that uncertainty-based detectors may flag UNK more than SI or CF and that latent signals may better address CF.The evaluation asks whether confidence is necessary for hallucination and whether current methods respect this distinction.Source: §E.2 · pp. 31–32

F METHODS

• The study categorizes 14 papers from 2023–2025 and evaluates detection and mitigation on Llama 3.1 8B and Gemma 3 12B, stratifying metrics by SI, CF, and UNK.The literature audit organizes methods in the RQ1–RQ2 space; the evaluation reports recall, residual error rate, and mitigation correctness.Source: F METHODS · p. 32, Figure 10 · p. 32

F.1 TAXONOMY DIMENSIONS

• Methods are classified by prompt sensitivity, intervention stage, knowledge assumption, uncertainty usage and representation, and target error type.The dimensions distinguish modeling semantically equivalent prompts, inference-time/post-hoc/external-knowledge interventions, and verbal/latent/semantic uncertainty signals.Source: §F.1 · p. 33

G DATASET CONSTRUCTION

• The controlled factual-QA dataset spans NH, SI, CF, and UNK categories, using 30 semantically equivalent variants and two instruction schemes, with uncertainty elicitation added in S2.S1 uses multiple instruction prompts and decoding configurations without uncertainty elicitation; S2 adds an uncertainty-elicitation instruction.Source: G DATASET CONSTRUCTION · p. 33

G.1 SOURCE DATASETS

• Questions come from TriviaQA and SimpleQA [101], with short, unambiguous answers and filtering against questions requiring post-training-cutoff knowledge for SI.TriviaQA is restricted to short-phrase answers; SimpleQA supplies its calibration prompt.Source: §G.1 · p. 33

G.2 CONFIDENCE PROMPT TEMPLATE

• Models are prompted for a best-guess answer and a 0%–100% confidence score; incorrect answers receive high confidence, while refusals or “do not know” responses score near zero.The confidence score adds little information beyond the verbal expression of uncertainty.Source: §G.2 · p. 34

G.3 BEHAVIORAL EVALUATION

• The behavioral grid crosses four instruction variants with decoding settings and 30 semantically equivalent paraphrases that vary wording without adding hints or information.Decoding includes greedy, temperature sampling, and nucleus sampling; one instruction explicitly elicits uncertainty.Source: §G.3 · pp. 34–35

H IS ACCESSIBLE KNOWLEDGE MEASURABLE? BEHAVIORAL EVIDENCE FROM V(Q)

• About 10 variants and label-convergence checks are recommended; adding variants can move CF labels to SI but cannot reverse that reassignment.SI and CF labels remain largely stable across decoding settings. SI has median ρ30 = 0.17 / 0.13 by Gemma / Llama, while CF has ρ30 = 0; variant information overlaps around zero for both categories.Source: H IS ACCESSIBLE KNOWLEDGE MEASURABLE?… · pp. 35–36, Figure 12 · p. 35, Figure 13 · p. 35

I RESULTS

• Across methods, models, and labeling schemes, SI is detectable and mitigable, whereas CF is rarely mitigated and is less often detected under S2.The paper examines whether this asymmetry reflects CF itself rather than an individual method.Source: I RESULTS · p. 36

I.1 THE CF BLIND SPOT

• 28.5% and 29.6% SI mitigation correctness under S1 and S2 contrasts with 0.7% and 2.8% for CF.Under S2, CF falls from 0.36 to 0.04 while UNK rises from 0.02 to 0.32; this relabeling does not make the remaining CF easier to correct.Source: Figure 14 · p. 36, §I.1 · p. 36

I.2 INTERVENTION ASYMMETRY

• Prompt-sensitive methods detect CF on par with SI but mitigate SI; present-knowledge methods outperform absent-knowledge methods in detection, while external knowledge best mitigates CF.Inference-time methods have the highest overall detection, whereas external-knowledge methods are comparatively effective at mitigation despite more moderate detection.Source: Figure 15 · p. 37, §I.2 · p. 37

I.3 UNCERTAINTY ESTIMATION FOR HALLUCINATIONS

• Uncertainty methods detect UNK best; latent signals lead for CF detection and mitigation, semantic signals best mitigate SI, and correction modestly improves CF and UNK mitigation over calibration.Verbal uncertainty is strongest for UNK but weakest for CF; methods in the None group mitigate CF better than latent methods by relying on external knowledge or decoding.Source: Table 8 · p. 37, §I.3 · p. 38, Figure 16 · p. 38