Pith. sign in

REVIEW 3 major objections 5 minor 77 references

Source attribution in LLMs is not a fixed ability: it depends on how conversational memory is structured. When in-context retrieval is available, models identify their own outputs almost perfectly; when episodic delay removes that shortcut,

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-07-31 23:29 UTC pith:ZCAG2YVT

load-bearing objection Worth refereeing, but the central "episodic delay" result is probably transcript scanning rather than memory: the encoding context is never cleared. the 3 major comments →

arxiv 2607.23927 v1 pith:ZCAG2YVT submitted 2026-07-27 cs.AI cs.CLcs.CYq-bio.NC

Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory

Saurabh Ranjan , Konstantina Sokratous , Brian Odegaard This is my paper
classification cs.AI cs.CLcs.CYq-bio.NC
keywords reality monitoringsource attributionlarge language modelsconversational memorymetacognitionconfidence calibrationreading hallucinationsepisodic delay
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that an LLM's ability to tell its own outputs from user-provided content—reality monitoring—is not an invariant property of the model but a function of how conversational memory is organized. Under minimal memory demands, models identify self-generated words at ceiling accuracy; once a delay separates encoding from testing, that advantage reverses into a fragile external-item edge. Corrective feedback creates dissociations invisible to standard benchmarks: some models swap their source responses entirely, while others get more accurate as their confidence becomes less diagnostic of correctness. The pattern across six models points to active expert-network size rather than aggregate parameter count, suggesting that scale alone is the wrong lever for improving self-knowledge in deployed systems.

Core claim

The paper adapts the human reality-monitoring word-pair task to large language models, contrasting externally supplied second words with self-generated ones. In single-trial conditions, internally generated items are identified with near-perfect accuracy because the generated word is still in the active context window. When an episodic delay is introduced—all encoding trials completed before a separate test phase—the source advantage reverses: externally perceived items become easier to attribute than imagined ones, with no model reaching ceiling and mean imagined-item accuracy falling below 50% without feedback. Corrective feedback selectively improves imagined-item accuracy but simultaneou

What carries the argument

The central apparatus is a word-pair reality-monitoring task in which an LLM either reproduces a provided second word (external/perceived) or generates a novel one (internal/imagined), then attributes each word's source and rates confidence. Three conversational-memory architectures vary the demand: Single-Turn (all subtasks in one interaction), Trial-Chain (subtasks across sequential turns with history retained), and Episodic-Chain (a delayed test phase after a block of encoding trials). Signal detection measures (d′, criterion c, Type 2 AUC) and the Goodman–Kruskal γ coefficient quantify source discrimination and metacognitive calibration. 'Reading hallucinations'—failures to reproduce an

Load-bearing premise

The reversal and confidence-decoupling effects are interpreted as consequences of episodic delay removing an in-context shortcut, which presupposes that the Episodic-Chain source judgments are actually answered from memory of the encoding phase rather than from shallow in-context cues (recency, position, semantic relatedness, or response bias) while the full transcript remains physically present in the conversation window.

What would settle it

Run the Episodic-Chain condition with the encoding-phase transcript removed from the test context (or replaced by unrelated filler turns). If source accuracy remains high or the external-item advantage persists under that manipulation, the effect is driven by prompt artifacts or response heuristics rather than by episodic memory; if accuracy drops to chance and the reversal disappears, the paper's memory-structure interpretation is confirmed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Existing knowledge-focused and calibration benchmarks cannot detect source-attribution failures, so multi-turn and agentic deployments need provenance-tracking evaluations alongside accuracy metrics.
  • Expressed confidence should not be treated as a calibrated trustworthiness signal in high-stakes settings, because feedback can make confidence less diagnostic of correctness even as accuracy improves.
  • Feedback loops in conversational systems may amplify reading hallucinations rather than reduce them, especially under larger memory loads, complicating the assumption that self-correction is beneficial.
  • Scaling aggregate parameter count is an unreliable path to better self-knowledge; architecture choices such as expert-network size within mixture-of-experts models appear more consequential.
  • Encoding fidelity (reading hallucination rate) is a distinct safety-relevant failure mode that deserves monitoring in its own right, separate from source-attribution accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct causal test of the memory-structure claim would remove the encoding transcript from the test phase: if models still show the external advantage, the effect is due to learned source cues rather than genuine episodic memory, whereas if accuracy collapses to chance, the episode truly is the carrier of the attribution.
  • The human comparison suggests a convergent risk: humans externalize imagined content at higher rates under delay, and LLMs here show a similar externalizing pattern, implying that agentic systems may share the same failure mode when memory demands rise.
  • The authors' reading-hallucination construct could be repurposed as a lightweight proxy metric for input-fidelity monitoring in real deployments—detecting when a model silently rewrites a user-provided fact before it ever reaches a downstream reasoner.
  • The confidence–accuracy decoupling under feedback hints that training-time reward signals on correctness alone may actively un-calibrate source memory; testing alternative feedback forms (explanatory, delayed, or confidence-targeted) could reveal a design space that preserves calibration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper adapts the human reality-monitoring word-pair paradigm to six instruction-tuned LLMs (Gemma 3 12B/27B with and without QAT, Llama 3.3 70B, Llama 4 16x17B). Experiment 1 compares Single-Turn and Trial-Chain implementations of a four-component task (word completion, relatedness rating, source judgment, confidence). Internal items are at ceiling; external-item accuracy varies, with reading hallucinations as the strongest predictor of source errors. Experiment 2 uses an Episodic-Chain design crossing set size (20 vs 40) with corrective feedback, reporting a reversal to an external-item advantage, degraded metacognitive sensitivity, feedback-induced response inversion or confidence-accuracy decoupling, and model-specific patterns. The authors conclude that source attribution is not an invariant property of a model but a function of how conversational memory is structured, with implications for agentic AI safety.

Significance. If the episodic results were valid measures of memory-based source attribution, the paper would make a useful contribution: it introduces a behavioral paradigm for studying LLM source monitoring, reports effects replicated across six models, and proposes a safety-relevant dissociation between source-discrimination accuracy and metacognitive calibration. Strengths include open code/data, per-trace random-effects clustering, disclosed sensitivity analyses treating model as a random effect, and explicit caveats about ceiling effects and SDT interpretability. However, the central interpretation is compromised by the Episodic-Chain design, in which encoding turns remain in the context window at test and source can be read directly from the transcript structure. The paper's main claims are therefore currently ahead of the data; with a true-memory control or a substantial reframing as context-structure effects, the empirical findings could still be valuable.

major comments (3)
  1. [§5.3.2, S.PC–S.PD, Fig. 2c] The Episodic-Chain condition is not a memory manipulation: the encoding turns are never removed from the conversation before test. The test prompt tells the model that Test_Word appeared as word_1 in a previous pair, and all previous human/assistant turns remain in the context. For external trials the user's encoding turn contains word_2; for internal trials that user turn contains a blank and word_2 appears only in the assistant response. The source is therefore directly readable from the transcript by locating the matching encoding turn. The claim that episodic delay removes the within-turn retrieval shortcut (Discussion) is unsupported; the observed reversal and confidence decoupling could reflect a shift in which visible transcript cue is used. A control that removes or masks the encoding phase before test is needed, or the claims should be reframed as context-structure effects. This
  2. [§3 and Conclusion] The conclusion that the pattern implicates active rather than aggregate parameter count is not supported by the design. Only one MoE architecture (Llama4 16x17B) is included, and active versus aggregate parameters are not independently varied; the six models differ in family, training data, instruction tuning, and inference settings. The sentence 'This pattern implicated individual expert-network size...' overstates the evidence. At most this is a hypothesis that requires a matched dense-versus-MoE comparison or a within-family manipulation of active parameters.
  3. [§2.2.1, §2.2.2] Because encoding turns remain visible in Episodic-Chain, the reading-hallucination analysis does not uniquely identify an upstream encoding bottleneck. A reading hallucination is scored as a mismatch between the model's response and the supplied word_2, but the supplied word_2 remains present in the user turn; with transcript access the source is recoverable independently of the model's own reproduction. Thus the conditional observation that hallucination-free trials are not near ceiling (e.g., no feedback 79.1% versus 67.9%) does not demonstrate that source monitoring itself degrades under memory load; it is equally consistent with the model failing to use the visible transcript consistently. The upstream-bottleneck interpretation needs to be re-derived under the transcript-parsing account.
minor comments (5)
  1. [§5.2, Ref. [31]] The human RM dataset is cited as blinded, so the claim of direct human-LLM comparison cannot be checked. Please provide the full reference or a public data description, and ideally the human summary statistics used for the qualitative comparison.
  2. [Fig. 3d, Fig. 5] The figures show Fisher Z-transformed gamma values, but the text sometimes cites raw gamma values without noting the transformation. Add a note in the text and captions stating which quantity is displayed and which is cited.
  3. [§5.1] For Llama4 16x17B, define what is meant by active versus aggregate parameter count, and give the approximate active parameters per forward pass. The term individual expert-network size is used without a precise definition.
  4. [§5.4] The claim that zero trials were excluded due to parse failures across 75,456 LLM responses is surprising. A per-model and per-condition table of parse counts and any retry logic would aid reproducibility.
  5. [S.PE] The imagined-trial feedback refers to a word not found in an English dictionary, but the dictionary source is unspecified. Specify the dictionary or wordlist used, or remove this criterion from the feedback definition.

Circularity Check

0 steps flagged

No significant circularity: the central effects are empirical, and the one acknowledged task-transparency artifact (Exp1 internal ceiling) is explicitly flagged rather than presented as a prediction.

full rationale

The paper makes no first-principles derivation whose outputs are equivalent to its inputs by construction. Experiment 1's ceiling for internal items is a designed transparency property: 'the generated word remained in the active context window at the time of judgment, making imagined-item identification trivially accessible rather than requiring inferential source monitoring.' The authors explicitly identify this as an artifact and do not use it as evidence for reality-monitoring capacity. Experiment 2's reversal, confidence decoupling, and d'/Type-2 dissociation are empirical outcomes from 72,000 test-phase observations, not fitted parameters renamed as predictions. The GLM's 'predicted accuracy' conditional on no reading hallucination is a within-sample conditional estimate, not a held-out prediction, and it is not load-bearing for the central claim. The Episodic-Chain design leaves encoding turns physically present in the context window, so source labels are visible in transcript structure; this raises a legitimate construct-validity threat (transcript parsing rather than memory), but that is a confounding interpretation, not a circular derivation: the reported reversal is measured, not entailed by the design definition. Reference [31] is blinded and may be self-authored, but it supplies human-comparison materials and a qualitative baseline only; the core LLM findings do not depend on it. No uniqueness theorem, ansatz, or external benchmark is imported from the authors' prior work to force the conclusion. Therefore no circular step meets the required evidence bar.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper introduces no new theoretical entities, forces, or free parameters fitted to manufacture the central claim. The 'trace' is an analysis unit (a complete encoding-to-test sequence), not a theoretical construct. The principal unverified inputs are the construct validity of the word-pair RM task for LLMs, the representativeness of six models run under Ollama defaults, and the blinded human-comparison dataset.

axioms (3)
  • domain assumption The word-pair source-attribution task is a valid operationalization of human reality monitoring when applied to LLMs.
    The entire paper interprets LLM JSON judgments ('internal' vs 'external') as source monitoring; construct validity is assumed based on analogy to human RM paradigms (Methods 5.3, Figure 1), not independently established for LLMs.
  • domain assumption The six tested instruction-tuned models run via Ollama at temperature 0 with default context lengths are representative enough to attribute behavioral differences to architecture.
    Methods 5.1 fixes temperature at 0 but leaves top-p and context length at Ollama defaults per model; context-window differences could affect memory-dependent results, and the paper itself limits generalization to the architecture sample.
  • domain assumption The human RM baseline is valid and comparable via the prior human dataset [31].
    The human comparison and 'same paradigm and materials' claim depend on reference [31], which is currently 'Anonymous. [blinded for review]' in this preprint; the comparison cannot be independently verified.

pith-pipeline@v1.3.0-alltime-deepseek · 32672 in / 11419 out tokens · 131061 ms · 2026-07-31T23:29:31.002384+00:00 · methodology

0 comments
read the original abstract

A conversational AI that cannot tell its own output from what a user said will treat its own mistakes as user-provided facts. In humans, this capacity is called reality monitoring, and its failures are linked to hallucinations, delusions, and confabulation, yet whether LLMs possess it remains untested. Here we show, across two experiments and six LLMs, that source attribution depends on how conversational memory is structured: ceiling accuracy for self-generated content under minimal memory demands reverses to a fragile external-item advantage once episodic delay removes that shortcut. Feedback exposes two failures: in some models, internal and external judgments swap; in others, accuracy improves while confidence decouples from correctness, dissociations invisible to existing benchmarks. Across models, this pattern implicates active, not aggregate, parameter count. This suggests that as AI systems take on autonomous, multi-turn roles, evaluating what they know is not enough: tracking where that knowledge came from may matter equally.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

77 extracted references · 17 linked inside Pith

  1. [1]

    Brown, T. B.et al. Language models are few-shot learners. InAdvances in Neural Information Processing Systems, Vol. 33, 1877–1901 (Curran Associates, Inc., 2020)

    1901

  2. [2]

    Preprint, arXiv:2206.07682

    Wei, J.et al.Emergent abilities of large language models (2022). Preprint, arXiv:2206.07682

    Pith/arXiv arXiv 2022

  3. [3]

    Preprint, arXiv:2406.11813

    Chang, H.et al.How do large language models acquire factual knowledge during pretraining? (2024). Preprint, arXiv:2406.11813

    Pith/arXiv arXiv 2024

  4. [4]

    Preprint, arXiv:1909.01066

    Petroni, F.et al.Language models as knowledge bases? (2019). Preprint, arXiv:1909.01066

    Pith/arXiv arXiv 2019

  5. [5]

    Preprint, arXiv:2308.09124

    Hernandez, E.et al.Linearity of relation decoding in transformer language models (2024). Preprint, arXiv:2308.09124

    Pith/arXiv arXiv 2024

  6. [6]

    Preprint, arXiv:2210.13382

    Li, K.et al.Emergent world representations: Exploring a sequence model trained on a synthetic task (2024). Preprint, arXiv:2210.13382

    Pith/arXiv arXiv 2024

  7. [7]

    Emergent introspective awareness in large language models (2025)

    Lindsey, J. Emergent introspective awareness in large language models (2025). URL https://transformer-circuits.pub/2025/introspection/index.html. Anthropic Interpretability Research

    2025

  8. [8]

    J.et al.Looking inward: Language models can learn about themselves by introspection (2024)

    Binder, F. J.et al.Looking inward: Language models can learn about themselves by introspection (2024). Preprint, arXiv:2410.13787

    Pith/arXiv arXiv 2024

  9. [9]

    Introspective capabilities in large language models.Journal of Consciousness Studies30, 143–153 (2023)

    Long, R. Introspective capabilities in large language models.Journal of Consciousness Studies30, 143–153 (2023). 46

    2023

  10. [10]

    Comsa, I. M. & Shanahan, M. Does it make sense to speak of introspection in large language models? (2025). Preprint, arXiv:2506.05068

    Pith/arXiv arXiv 2025

  11. [11]

    & Lu, C.From imitation to introspection: Prob- ing self-consciousness in language models

    Chen, S., Yu, S., Zhao, S. & Lu, C.From imitation to introspection: Prob- ing self-consciousness in language models. InFindings of the Association for Computational Linguistics: ACL 2025, 7553–7583 (2025)

    2025

  12. [12]

    & Evans, O.Towards a situational awareness benchmark for LLMs

    Laine, R., Meinke, A. & Evans, O.Towards a situational awareness benchmark for LLMs. InNeurIPS 2023 Workshop on Socially Responsible Language Modelling Research (SoLaR)(2023)

    2023

  13. [13]

    Nguyen, J., Hoang, K., Attubato, C. L. & Hofst¨ atter, F. Probing and steering evaluation awareness of language models (2025). Preprint, arXiv:2507.01786

    Pith/arXiv arXiv 2025

  14. [14]

    & Hobbhahn, M

    Needham, J., Edkins, G., Pimpale, G., Bartsch, H. & Hobbhahn, M. Large language models often know when they are being evaluated (2025). Preprint, arXiv:2505.23836

    Pith/arXiv arXiv 2025

  15. [15]

    Claude Sonnet 3.7 (often) knows when it’s in align- ment evaluations (2025)

    Apollo Research. Claude Sonnet 3.7 (often) knows when it’s in align- ment evaluations (2025). URL https://www.apolloresearch.ai/science/ claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/. Apollo Research Blog

    2025

  16. [16]

    The Claude 3 model family: Opus, Sonnet, Haiku

    Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku. Tech. Rep., Anthropic (2024)

    2024

  17. [17]

    Singhal, K.et al.Large language models encode clinical knowledge.Nature620, 172–180 (2023)

    2023

  18. [18]

    H., Hickman, K

    Choi, J. H., Hickman, K. E., Monahan, A. B. & Schwarcz, D. ChatGPT goes to law school.Journal of Legal Education71, 387–400 (2023)

    2023

  19. [19]

    A., MacKnight, R., Kline, B

    Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models.Nature624, 570–578 (2023)

    2023

  20. [20]

    Escalation risks from language models in military and diplo- matic decision-making

    Rivera, J.-P.et al. Escalation risks from language models in military and diplo- matic decision-making. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 836–848 (2024)

    2024

  21. [21]

    Johnson, M. K. & Raye, C. L. Reality monitoring.Psychological Review88, 67–85 (1981)

    1981

  22. [22]

    Johnson, M. K. Reality monitoring: An experimental phenomenological approach. Journal of Experimental Psychology: General117, 390–394 (1988)

    1988

  23. [23]

    R., Bond, R., Gibbard, E., Johnson, M

    Garrison, J. R., Bond, R., Gibbard, E., Johnson, M. K. & Simons, J. S. Monitoring what is real: The effects of modality and action on accuracy and type of reality monitoring error.Cortex87, 108–117 (2017). 47

    2017

  24. [24]

    S., Garrison, J

    Simons, J. S., Garrison, J. R. & Johnson, M. K. Brain mechanisms of reality monitoring.Trends in Cognitive Sciences21, 462–473 (2017)

    2017

  25. [25]

    S., Ritchey, M

    Simons, J. S., Ritchey, M. & Fernyhough, C. Brain mechanisms underlying the subjective experience of remembering.Annual Review of Psychology73, 159–186 (2022)

    2022

  26. [26]

    S., Henson, R

    Simons, J. S., Henson, R. N. A., Gilbert, S. J. & Fletcher, P. C. Separable forms of reality monitoring supported by anterior prefrontal cortex.Journal of Cognitive Neuroscience20, 447–457 (2008)

    2008

  27. [27]

    & Fleming, S

    Dijkstra, N., Kok, P. & Fleming, S. M. Perceptual reality monitoring: Neural mechanisms dissociating imagination from reality.Neuroscience & Biobehavioral Reviews135, 104557 (2022)

    2022

  28. [28]

    & de Haan, E

    Aleman, A. & de Haan, E. H. F. Fantasy proneness, mental imagery and reality monitoring.Personality and Individual Differences36, 1747–1754 (2004)

    2004

  29. [29]

    & Dudley, R

    Aynsworth, C., Nemat, N., Collerton, D., Smailes, D. & Dudley, R. Reality monitoring performance and the role of visual imagery in visual hallucinations. Behaviour Research and Therapy97, 115–122 (2017)

    2017

  30. [30]

    P., Baker, G

    Bentall, R. P., Baker, G. A. & Havers, S. Reality monitoring and psychotic hallucinations.British Journal of Clinical Psychology30, 213–222 (1991)

    1991

  31. [31]

    [blinded for review] (2024)

    Anonymous. [blinded for review] (2024). Reference details omitted for double-blind review

    2024

  32. [32]

    The signs and symptoms of schizophrenia.Comprehensive Psychiatry 26, 103–112 (1985)

    Alpert, M. The signs and symptoms of schizophrenia.Comprehensive Psychiatry 26, 103–112 (1985)

    1985

  33. [33]

    Aleman, A., B¨ ocker, K. B. E., Hijman, R., de Haan, E. H. F. & Kahn, R. S. Cognitive basis of hallucinations in schizophrenia: role of top-down information processing.Schizophrenia Research64, 175–185 (2003)

    2003

  34. [34]

    J., Gorman, J

    Br´ ebion, G., Smith, M. J., Gorman, J. M. & Amador, X. Discrimination accuracy and decision biases in different types of reality monitoring in schizophrenia.Journal of Nervous and Mental Disease185, 247–253 (1997)

    1997

  35. [35]

    M., Amador, X., Malaspina, D

    Br´ ebion, G., Gorman, J. M., Amador, X., Malaspina, D. & Sharif, Z. Source monitoring impairments in schizophrenia: Characterisation and associations with positive and negative symptomatology.Psychiatry Research112, 27–39 (2002)

    2002

  36. [36]

    Brunelin, J.et al.Source monitoring deficits in hallucinating compared to non- hallucinating patients with schizophrenia.European Psychiatry21, 259–261 (2006). 48

    2006

  37. [37]

    Brunelin, J.et al.Impaired verbal source monitoring in schizophrenia: An intermediate trait vulnerability marker?Schizophrenia Research89, 287–292 (2007)

    2007

  38. [38]

    Thoresen, C.et al.Frontotemporal hypoactivity during a reality monitoring paradigm is associated with delusions in patients with schizophrenia spectrum disorders.Cognitive Neuropsychiatry19, 97–115 (2014)

    2014

  39. [39]

    R.et al.Illusions and delusions: relating experimentally-induced false memories to anomalous experiences and ideas.Frontiers in Behavioral Neuroscience3, 53 (2009)

    Corlett, P. R.et al.Illusions and delusions: relating experimentally-induced false memories to anomalous experiences and ideas.Frontiers in Behavioral Neuroscience3, 53 (2009)

    2009

  40. [40]

    F., Staff, R

    Venneri, A., Shanks, M. F., Staff, R. T. & Della Sala, S. Nurturing syn- drome: A form of pathological bereavement with delusions in Alzheimer’s disease. Neuropsychologia38, 213–224 (2000)

    2000

  41. [41]

    Benedetti, F.et al.Reality monitoring and paranoia in remitted delusional depression.Clinical Neuropsychiatry: Journal of Treatment Evaluation2, 199–205 (2005)

    2005

  42. [42]

    Koller, W. N. & Cannon, T. D. Paranoia is associated with impaired novelty detection and overconfidence in recognition memory judgments.Journal of Abnormal Psychology130, 273–285 (2021)

    2021

  43. [43]

    A., Leong, Y

    Rossi-Goldthorpe, R. A., Leong, Y. C., Leptourgos, P. & Corlett, P. R. Paranoia, self-deception and overconfidence.PLOS Computational Biology17, e1009453 (2021)

    2021

  44. [44]

    R.et al.Hallucinations and strong priors.Trends in Cognitive Sciences 23, 114–127 (2019)

    Corlett, P. R.et al.Hallucinations and strong priors.Trends in Cognitive Sciences 23, 114–127 (2019)

    2019

  45. [45]

    R., Canavan, S

    Corlett, P. R., Canavan, S. V., Nahum, L., Appah, F. & Morgan, P. T. Dreams, reality and memory: confabulations in lucid dreamers implicate reality-monitoring dysfunction in dream consciousness.Cognitive Neuropsychiatry19, 540–553 (2014)

    2014

  46. [46]

    & Kopelman, M

    Dalla Barba, G. & Kopelman, M. Confabulations and related disorders: We’ve come a long way, but there is still a lot to do!Cortex87, 1–4 (2017)

    2017

  47. [47]

    Spontaneous confabulation, reality monitoring, and the limbic system — a review.Brain Research Reviews36, 150–160 (2001)

    Schnider, A. Spontaneous confabulation, reality monitoring, and the limbic system — a review.Brain Research Reviews36, 150–160 (2001)

    2001

  48. [48]

    Fotopoulou, A., Conway, M. A. & Solms, M. Confabulation: Motivated reality monitoring.Neuropsychologia45, 2180–2190 (2007)

    2007

  49. [49]

    & Luqman, H

    Alansari, A. & Luqman, H. Large language models hallucination: A comprehensive survey (2025). Preprint, arXiv:2510.06265. 49

    arXiv 2025

  50. [50]

    Huang, L.et al.A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems43, 42:1–42:55 (2025)

    2025

  51. [51]

    Higher-order representation in AI.Philosophy and the Mind Sciences7 (2026)

    Butlin, P. Higher-order representation in AI.Philosophy and the Mind Sciences7 (2026)

    2026

  52. [52]

    Wang, J. Z. MIRROR: A hierarchical benchmark for metacognitive calibration in large language models (2026). arXiv:2604.19809

    Pith/arXiv arXiv 2026

  53. [53]

    The metacognitive monitoring battery: A cross-domain benchmark for LLM self-monitoring (2026)

    Cacioli, J.-P. The metacognitive monitoring battery: A cross-domain benchmark for LLM self-monitoring (2026). arXiv:2604.15702

    Pith/arXiv arXiv 2026

  54. [54]

    InProceedings of the International Conference on Learning Representations(2026)

    Ackerman, C.Evidence for limited metacognition in LLMs. InProceedings of the International Conference on Learning Representations(2026). arXiv:2509.21545

    Pith/arXiv arXiv 2026

  55. [55]

    arXiv:2410.08133

    Pink, M.et al.Assessing episodic memory in LLMs with sequence order recall tasks (2024). arXiv:2410.08133

    Pith/arXiv arXiv 2024

  56. [56]

    Fleming, S. M. & Lau, H. C. How to measure metacognition.Frontiers in Human Neuroscience8, 443 (2014)

    2014

  57. [57]

    Christensen, R. H. B.ordinal: Regression Models for Ordinal Data(2023). URL https://CRAN.R-project.org/package=ordinal. R package version 2023.12-4

    2023

  58. [58]

    J., Podd, J

    Galvin, S. J., Podd, J. V., Drga, V. & Whitmore, J. Type 2 tasks in the theory of signal detectability: Discrimination between correct and incorrect decisions. Psychonomic Bulletin & Review10, 843–876 (2003)

    2003

  59. [59]

    F.et al.Lost in the middle: How language models use long contexts

    Liu, N. F.et al.Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics12, 157–173 (2024)

    2024

  60. [60]

    Kumaran, D.et al.Competing biases underlie overconfidence and underconfidence in LLMs.Nature Machine Intelligence(2026)

    2026

  61. [61]

    & Phillips, L

    Lichtenstein, S., Fischhoff, B. & Phillips, L. D. inCalibration of probabilities: The state of the art to 1980InJudgment under Uncertainty: Heuristics and Biases (eds Kahneman, D., Slovic, P. & Tversky, A.) 306–334 (Cambridge University Press, 1982)

    1982

  62. [62]

    & Valdenegro-Toro, M.Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models

    Groot, T. & Valdenegro-Toro, M.Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models. InProceedings of the 4th Workshop on Trustworthy Natural Language Processing (TrustNLP 2024), 145–171 (Association for Computational Linguistics, Mexico City, Mexico, 2024)

    2024

  63. [63]

    arXiv:2207.05221

    Kadavath, S.et al.Language models (mostly) know what they know (2022). arXiv:2207.05221. 50

    Pith/arXiv arXiv 2022

  64. [64]

    Henkel, L. A. & Franklin, N. Reality monitoring of physically similar and conceptually related objects.Memory & Cognition26, 659–673 (1998)

    1998

  65. [65]

    Roediger, H. L. & McDermott, K. B. Creating false memories: Remembering words not presented in lists.Journal of Experimental Psychology: Learning, Memory, and Cognition21, 803–814 (1995)

    1995

  66. [66]

    Alter, A. L. & Oppenheimer, D. M. Uniting the tribes of fluency to form a metacognitive nation.Personality and Social Psychology Review13, 219–235 (2009)

    2009

  67. [67]

    N., Lynch, A., Heimersheim, S

    Conmy, A., Mavor-Parker, A. N., Lynch, A., Heimersheim, S. & Garriga-Alonso, A.Towards automated circuit discovery for mechanistic interpretability. In Advances in Neural Information Processing Systems, Vol. 36, 16318–16352 (Curran Associates, Inc., 2023). arXiv:2304.14997

    Pith/arXiv arXiv 2023

  68. [68]

    Zheng, Z.et al.Attention heads of large language models.Patterns6, 101176 (2025)

    2025

  69. [69]

    & Gavves, E

    Bereska, L. & Gavves, E. Mechanistic interpretability for AI safety — a review. Transactions on Machine Learning Research(2024). URL https://openreview. net/forum?id=ePUVetPKu6

    2024

  70. [70]

    arXiv:2603.18353

    Basu, S.et al.Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations (2026). arXiv:2603.18353

    arXiv 2026

  71. [71]

    LessWrong (2025)

    Nanda, N.et al.A pragmatic vision for interpretability. LessWrong (2025). Google DeepMind Mechanistic Interpretability Team. https://www.lesswrong.com/posts/ StENzDcD3kpfGJssR/

    2025

  72. [72]

    & Tafjord, O

    Kim, B., Hewitt, J., Nanda, N., Fiedel, N. & Tafjord, O. Because we have LLMs, we can and should pursue agentic interpretability (2025). arXiv:2506.12152

    Pith/arXiv arXiv 2025

  73. [73]

    & Hruschka, E

    Pezeshkpour, P. & Hruschka, E. Large language models sensitivity to the order of options in multiple-choice questions (2023). arXiv:2308.11483

    Pith/arXiv arXiv 2023

  74. [74]

    & Schielzeth, H

    Nakagawa, S. & Schielzeth, H. A general and simple method for obtaining R2 from generalized linear mixed-effects models.Methods in Ecology and Evolution 4, 133–142 (2013)

    2013

  75. [75]

    Pedregosa, F.et al.Scikit-learn: Machine learning in Python.Journal of Machine Learning Research12, 2825–2830 (2011)

    2011

  76. [76]

    Hautus, M. J. Corrections for extreme proportions and their biasing effects on estimated values of d′.Behavior Research Methods, Instruments, & Computers 27, 46–51 (1995). 51

    1995

  77. [77]

    Macmillan, N. A. & Creelman, C. D.Detection Theory: A User’s Guide2 edn (Lawrence Erlbaum Associates, Mahwah, NJ, 2005). 52

    2005