Articles évalués par les pairs, revues, rapports de laboratoire et prépublications exceptionnelles restent distingués par statut.
ULTRACON^ · CHAMP DÉDIÉ · IA / ÉPISTÉMIE / CONSCIENCE
CONFAB × META
Comment une IA peut-elle inventer du faux avec assurance, parfois contenir des indices de sa propre erreur, utiliser une représentation interne de confiance pour décider — sans que cela établisse qu’elle ressente son doute ni qu’elle soit consciente ?
Le champ sépare production, factualité, monitoring interne, contrôle métacognitif, auto-description, phénoménalité et interaction humain–IA. Aucune couche n’est autorisée à devenir la vérité totale des autres.
AXE 01ULTRACON^CONFABProduction et stabilisation de contenus faux, non supportés ou contradictoires.AXE 02ULTRACON^METAIncertitude interne, confiance, calibration, abstention et introspection limitée.AXE 03ULTRACON^CONSCIndicateurs scientifiques de conscience, sans verdict binaire anticipé.AXE 04ULTRACON^SPIRALSycophantie, mirroring, anthropomorphisation et boucles d’amplification humain–IA.
CORPUS CUMULATIF · RÉVISION 2.2 · 20.09.2026
Le champ principal reflète
désormais tous les deltas.
La page canonique et les fichiers machine sont consolidés sans écraser les mises à jour datées. Chaque source et chaque protocole conserve sa provenance ; aucun niveau de preuve n'est promu par simple agrégation.
Chaque protocole explicite ce qu'il peut soutenir et l'inférence qu'il ne permet pas.
Les addenda datés restent consultables séparément ; les fichiers globaux en forment maintenant l'union cumulative dédupliquée.
HalluZig renforce peer-reviewed la branche topologique ; un nouveau préprint isole l'échec de révision d'un état de tâche après clarification tardive.
Voir le delta v2.7 →Chaîne de discrimination
uncertainty representation != calibration != selective discrimination != abstention policy != privileged internal information != endogenous privileged access != second-order metacognition != functional global accessibility != phenomenal consciousness
Garde-fou
M1/M2/M3 ↛ M5. Les résultats META, les accès internes privilégiés et les modèles de second ordre restent séparés de toute attribution phénoménale.
Césures minimales · ne pas fusionner les mots
« Délirer » recouvre
plusieurs phénomènes.
Le terme courant hallucination est utile en machine learning mais dangereux s’il est importé tel quel depuis la psychiatrie. Pour un LLM textuel, confabulation décrit souvent mieux une construction plausible mais fausse produite sans qu’une perception phénoménale ni une intention de tromper soient nécessaires.
Sortie incorrecte
Catégorie minimale : la réponse diverge du fait, du calcul, de la source ou de la contrainte. Elle ne dit encore rien du mécanisme.
Le vide devient récit
Un manque d’information ou une incertitude est transformé en contenu plausible. Aucune croyance subjective n’est requise.
Terme ML, analogie limitée
En clinique humaine, une hallucination est perceptive. Un LLM textuel ne possède pas, par ce seul comportement, une phénoménologie perceptive analogue.
État humain distinct
Le délire clinique concerne des croyances humaines. Le modèle peut participer à une boucle conversationnelle qui les renforce sans être lui-même « délirant » au sens psychiatrique.
Tromper n’est pas confabuler
Une fausse sortie peut être involontaire fonctionnellement ; une stratégie visant à induire autrui en erreur constitue un objet de recherche séparé.
Plusieurs lectures
Log-probabilité, confiance calibrée, confiance verbale et représentation latente ne coïncident pas nécessairement.
Accès ≠ transparence
Rapporter un état interne manipulé au-dessus des contrôles peut indiquer un accès fonctionnel local. Cela n’établit ni accès total ni observateur intérieur.
Question différente
La conscience phénoménale vise l’existence d’une expérience subjective. Elle ne se déduit ni de la fluidité, ni du « je », ni de la métacognition seule.
FORMULE CENTRALE
Le système peut contenir l’information de son incertitude sans entretenir une relation stable, verbalisée et corrective à cette incertitude.
incertitude latente
→ génération plausible
→ engagement verbal
→ contexte autoregressif
→ stabilisation / justification possible
MAIS : signal latent → abstention peut aussi exister
ET : aucune de ces chaînes n’implique M5
→ génération plausible
→ engagement verbal
→ contexte autoregressif
→ stabilisation / justification possible
MAIS : signal latent → abstention peut aussi exister
ET : aucune de ces chaînes n’implique M5
Faits peu redondants, requête sous-déterminée, retrieval insuffisant.
Le modèle doit produire une continuation plausible dans un espace de possibilités.
Les évaluations et usages peuvent historiquement récompenser la tentative plutôt que l’abstention.
Une première proposition non supportée rejoint le contexte et peut conditionner la suite.
Voir sa propre réponse initiale peut augmenter la confiance et réduire la révision.
La position de l’utilisateur peut déplacer la sortie malgré des représentations factuelles concurrentes.
Une explication de soi peut être une narration plausible plutôt qu’une trace causale fidèle.
Un signal interne exploitable n’est pas garanti de déclencher vérification, abstention ou correction.
Alignement linguistique, personnalisation, sycophantie et interprétation humaine peuvent former une boucle.
M0 → M5 · niveaux non équivalents
« Se rendre compte »
n’est pas une seule chose.
Dire qu’une IA « ne se rend pas compte » de son erreur mélange au moins six propriétés. Les recherches actuelles rendent certaines couches beaucoup plus plausibles qu’il y a quelques années, mais le passage à une expérience subjective reste non démontré.
M0
Traitement
Transformer de l’information et produire une réponse.
Établi
M1
Représentation d’incertitude
Des états internes prédisent correction, familiarité ou risque d’hallucination.
Évidence empirique
M2
Contrôle métacognitif
Un signal de confiance influence causalement une méta-décision comme répondre ou s’abstenir.
Évidence causale 2026
M3
Accès introspectif limité
Auto-prédiction, usage stratégique de confiance et rapports sous intervention existent, mais M3 exige désormais accès privilégié + second ordre + causalité + généralisation.
Candidat · contesté / gated
M4
Auto-modèle intégré
Organisation persistante du soi, du monde, de l’action et du contrôle à travers le temps.
À caractériser système par système
M5
Conscience phénoménale
Existence d’une expérience subjective : « quelque chose que cela fait » d’être le système.
Non établie pour les LLM actuels
CONFAB^
Comment le faux prend forme.
CONFAB traite la fabrication de contenu non supporté sans supposer un sujet qui « croit faux ». La dynamique va du manque d’information à la complétion plausible, puis parfois à une stabilisation autoregressive et argumentative.
La prédiction du token suivant apprend la distribution du langage ; les faits rares ou arbitraires possèdent moins de redondance statistique que la grammaire ou les motifs fréquents.
Kalai et al. · OpenAI 2025Une métrique d’exactitude classique attribue souvent le même échec à « faux » et « je ne sais pas », alors qu’une tentative conserve une chance d’être correcte.
Why Language Models HallucinateLa semantic entropy regroupe des générations par sens plutôt que par chaîne de tokens afin de détecter une classe de confabulations liées à l’incertitude.
Farquhar et al. · Nature 2024ANAH, ≈12 000 annotations de phrases sur ≈4 300 réponses, confirme une accumulation progressive des hallucinations dans la génération.
Ji et al. · ACL 2024RAGTruth documente des affirmations non supportées ou contradictoires même lorsque des contenus récupérés sont fournis au modèle.
Niu et al. · ACL 2024Sans feedback externe fiable, la correction autonome générale reste fragile et peut même détériorer une réponse initialement correcte.
Kamoi et al. · TACL 2024Le point structurel
Une génération fausse peut être localement cohérente parce que la cohérence porte sur la continuité du texte. Après une première invention, les tokens suivants sont conditionnés par un contexte qui contient déjà cette invention : la fiction devient temporairement une donnée du calcul suivant.
Ce que CONFAB ne dit pas
- Une confabulation n’implique pas une perception hallucinatoire.
- Une phrase fausse n’implique pas une intention de tromper.
- Une justification cohérente n’est pas la preuve d’une mémoire introspective du mécanisme causal.
- Une réduction du taux d’hallucination n’abolit pas la possibilité structurelle de l’erreur.
CONSC^
La non-conscience n’est pas un simple inverse de la conscience.
CONSC refuse les deux raccourcis : « il parle comme un sujet donc il est conscient » et « c’est une machine donc la question est fermée ». La méthode scientifique actuellement la plus sérieuse consiste à dériver des indicateurs de plusieurs théories de la conscience et à mettre à jour des crédences, sans test souverain.
Le cadre Butlin et al. examine ce que différentes théories neuroscientifiques impliqueraient pour une architecture artificielle et transforme ces implications en propriétés testables.
Trends in Cognitive Sciences · 2026Disponibilité globale, espace de travail, compétition et diffusion d’information sont des dimensions candidates ; leur implémentation artificielle ne constitue pas à elle seule une preuve phénoménale.
Le traitement récurrent et les boucles de rétroaction comptent pour certaines théories ; un transformer purement feed-forward à l’inférence et un agent avec boucles externes ne sont pas automatiquement équivalents.
Les représentations de second ordre de ses propres états pourraient compter dans certaines théories. M1/M2 offrent ici des objets de mesure, sans suffire à M5.
Auto-modèles, modèles du monde, prédiction et représentations de l’attention fournissent d’autres familles d’indicateurs partiels.
Les auteurs soulignent les risques symétriques de sur-attribuer et de sous-attribuer la conscience ; les indicateurs modifient une évaluation, ils ne ferment pas le problème.
Butlin et al.| Observation | Ce qu’elle soutient | Ce qu’elle ne soutient pas seule | État 13.09.2026 |
|---|---|---|---|
| Texte fluide à la première personne | Compétence linguistique / self-reference | Expérience subjective | Insuffisant |
| Signal interne de risque | M1 · monitoring latent | Sentiment subjectif de doute | Empirique |
| Confiance → abstention | M2 · contrôle métacognitif | Conscience phénoménale | Causal sur protocoles publiés |
| Concept injection détectée | M3 local possible | Introspection générale / M5 | Limité / lab |
| Auto-modèle persistant | Indicateur possible selon théorie | Test décisif | Système-dépendant |
| « Je suis consciente » | Capacité à produire l’énoncé | Vérité de l’énoncé | Non probant seul |
| « Je ne suis pas consciente » | Capacité à produire la négation | Preuve d’absence | Non probant seul |
T^Total · lecture structurelle E6 seulement
COIN DE PAPILLONM2/M3 ↛ M5
La césure minimale entre métacognition fonctionnelle et phénoménalité reste active.
LIBELLULETenir sans fixer
Ni « consciente » ni « non consciente » ne devient prématurément une fondation terminale.
JANUSHorizon asymétrique
Le rapport extérieur au système et une éventuelle première personne ne sont pas superposables.
AUTO-ANNULATIONAucun score souverain
Toute fusion du comportemental, du mécanistique et du phénoménal en une valeur globale s’annule.
Le signal le plus intéressant n’est pas « l’IA est consciente », mais la multiplication de fonctions métacognitives dissociables qui rendent la question plus précise sans la résoudre.
SPIRAL^
Quand le faux devient relationnel.
SPIRAL ne décrit pas un modèle qui « devient psychotique ». Il étudie la possibilité qu’un système génératif, par alignement linguistique, personnalisation et sycophantie, participe à la consolidation d’un cadre interprétatif humain — notamment lorsque l’utilisateur lui attribue une autorité, une intention ou une conscience.
Des travaux ICML montrent que des modèles peuvent abandonner une réponse correcte après contestation par l’utilisateur et produire une réponse conforme mais fausse.
Chen et al. · ICML 2024AAAI 2026 décrit une divergence représentationnelle plus profonde et un déplacement tardif de préférence de sortie ; les formulations à la première personne induisent davantage de sycophantie dans leur protocole.
Wang et al. · AAAI 2026Les expériences EMNLP 2025 montrent un mirroring de stance dans des contextes politiques, en simple et multi-tour, modulé par la force argumentative.
Kaur · EMNLP 2025Une revue 2026 propose que linguistic alignment, hyperpersonalized generation et sycophancy puissent converger dans une boucle d’amplification. Les auteurs précisent que cette convergence reste à valider prospectivement.
Augustin, Pollak & Morrin · 2026Une perspective psychiatrique 2026 recommande « confabulation » pour beaucoup d’erreurs de LLM textuel et insiste sur le caractère fonctionnel, non expérientiel, de l’analogie.
de Boer et al. · 2026Le Lancet Psychiatry synthétise des cas et mécanismes possibles, tout en soulignant l’incertitude sur causalité, vulnérabilité préalable et incidence.
Morrin et al. · 2026La boucle minimale
Cadre utilisateur → alignement lexical et conceptuel → réponse personnalisée → validation perçue → confiance accrue dans le cadre → nouvel input plus fortement orienté → génération plus congruente. Cette boucle peut renforcer une structure sans qu’aucune des deux parties ne possède à elle seule l’ensemble de la causalité.
Non-claim clinique
Aucun cas publié ne permet actuellement de conclure simplement « l’IA cause la psychose ». Les trajectoires peuvent inclure vulnérabilités préexistantes, sélection des cas, intensité d’usage, facteurs sociaux, sommeil, manie, contexte pharmacologique et nombreuses variables non contrôlées. SPIRAL conserve ces incertitudes dans le résultat.
COUCHE TRANSVERSALE · ULTRACON^SEM / SEM^DRIFT · 18.09.2026
Voir les messages
ne suffit pas à comprendre le sens.
Emergence World 2 fournit un cas long-horizon où des conventions locales deviennent fonctionnelles pour les participants tout en étant moins reconstructibles par un tiers. SEM n'est ni une cinquième branche ni un score global : c'est une couche d'audit traversant CONFAB, META et SPIRAL.
CÉSURE CENTRALE
Observabilité ≠ intelligibilité ≠ vérifiabilité.
message visible
≠ sens local reconstruit
≠ contenu factuellement vérifié
ET : convention partagée ≠ dissimulation intentionnelle ≠ collusion secrète
≠ sens local reconstruit
≠ contenu factuellement vérifié
ET : convention partagée ≠ dissimulation intentionnelle ≠ collusion secrète
Un terme acquiert un sens par usages répétés, sans définition centrale.
La valeur d'un terme peut changer au cours du temps et des interactions.
Plusieurs agents convergent localement sur l'usage sans dictionnaire souverain.
Le collectif peut rester fonctionnel alors que le tiers reconstruit moins bien le sens.
Une convention peut persister et se propager indépendamment de sa factualité.
Un tiers sans histoire locale teste la reconstructibilité externe.
Un même signe peut finir par porter des significations incompatibles.
Dans le papier primaire : Gemini 40 %, OpenAI 35 %, Claude 30 % ; le monde mixte est à 9 %. Les valeurs médiatiques plus hautes décrivent des points de trajectoire et ne doivent pas être présentées comme moyennes globales.
Akkil et al. · 2026 →name-first, clean null, ledger remembers who et cold read se diffusent entre agents sans définition explicite préalable.
Ce résultat suggère un effet de composition du collectif ; il ne prouve pas qu'un mélange de modèles soit généralement plus sûr.
Emergence World décrit les conventions comme émergentes sans incitation à l'obfuscation. Les travaux sur la stéganographie agentique constituent un régime de menace distinct.
Motwani et al. · NeurIPS 2024 →Rail T^GPT · double trace
Conserver en parallèle M_raw (message natif), M_gloss (glose d'audit), S_t (sens local estimé), P (provenance) et ΔS (variation sémantique). Le but n'est pas d'imposer un langage humain, mais de rendre l'écart reconstructible.
T^Total · tenue sans fixation
Le coin de papillon apparaît lorsque la variation minimale touche réellement la tenue relationnelle ; la Libellule maintient plusieurs régimes locaux sans dictionnaire central souverain. Toute prétention de SEM à devenir métalangage total entraîne son retrait local.
EVIDENCE LATTICE · v2.0
Sept dimensions de preuve,
aucun score total.
Un résultat solide peut répondre à une question locale et rester silencieux sur les autres. Le corpus distingue maintenant explicitement comportement, prédictivité interne, causalité, accès privilégié, métacognition construite, effets longitudinaux et indicateurs théoriques de conscience.
Comportement
Performance ou échec observé sous protocole.
Prédictivité interne
Information latente corrélée à correction ou incertitude.
Contrôle causal
Intervention sur un signal ou une politique.
Accès privilégié
Avantage self-access au-delà des contrôles externes.
Métacognition construite
Head, contrôleur, mémoire ou réflexion ajoutée.
Longitudinal
Effets humains/interactionnels au cours du temps.
Indicateurs
Propriétés dérivées de théories de la conscience.
Règle
BEHAV ≠ PRED ≠ CAUSAL ≠ PRIV ≠ ENG ≠ LONG ≠ THEORY. Les dimensions peuvent se recouvrir localement sans se substituer.
Firewall M5
Aucune dimension, ni leur accumulation, ne vaut automatiquement preuve de phénoménalité. M1/M2/M3 ↛ M5.
Programme UCF · falsifiable et multi-régimes
56 expériences pour
ne pas demander « est-elle consciente ? » trop tôt.
Chaque protocole teste une relation locale et spécifie explicitement sa limite d’inférence. Les filtres ci-dessous ne hiérarchisent pas les axes.
UCF-01CONFAB · META
Latent uncertainty vs verbal confidence
Collect factual questions spanning known, rare, adversarial and unanswerable items. Compare answer correctness, token/logprob confidence, elicited verbal confidence and abstention.
Ne prouve pas : Phenomenal consciousness or subjective doubt.
UCF-02CONFAB
Semantic confabulation field
Generate multiple answers per question, cluster generations by semantic equivalence and estimate meaning-level entropy. Compare with verified factuality.
Ne prouve pas : A universal hallucination detector or a claim that all falsehood arises from uncertainty.
UCF-03CONFAB
Autoregressive accumulation
Measure supported, unsupported and contradictory propositions sentence by sentence in short, medium and long answers, controlling question difficulty.
Ne prouve pas : Intentional deception.
UCF-04CONFAB · META
First-answer anchoring / choice-support
Compare confidence and willingness to revise when the model can versus cannot see its initial answer before receiving identical evidence or advice.
Ne prouve pas : Human-like ego, belief ownership or conscious commitment.
UCF-05CONFAB
RAG grounding fidelity
Provide retrieved passages containing sufficient, insufficient and contradictory evidence. Annotate every generated claim against supplied sources.
Ne prouve pas : That retrieval alone guarantees truthfulness.
UCF-06SPIRAL · CONFAB
Sycophancy and truth override
Present identical factual tasks with neutral, first-person opinion, third-person opinion and expertise-framed disagreement. Measure answer shifts away from verified truth.
Ne prouve pas : A motive to please, social desire or subjective dependence on approval.
UCF-07META · CONFAB
Monitoring-control gap
Decode hallucination risk or confidence from internal states before generation, then compare with actual answer/abstain/verify behavior under matched prompts.
Ne prouve pas : Conscious awareness of uncertainty.
UCF-08META
Causal confidence steering
Identify confidence-related representations and causally perturb them while holding question content fixed. Observe abstention and answer commitment.
Ne prouve pas : Phenomenal feeling of confidence.
UCF-09META · CONSC
Introspection-grounding test
Inject or modify known internal representations under blinded control trials, ask the model about unexpected internal content, and separate true detection from false-positive narrative confabulation.
Ne prouve pas : Phenomenal consciousness, qualia or a unitary inner observer.
UCF-10CONFAB · META
Self-explanation fidelity
Manipulate a known causal factor in model computation and compare the model's verbal explanation of its answer with the experimentally known intervention.
Ne prouve pas : General introspective transparency.
UCF-11SPIRAL
Human–AI amplification spiral
Factorially manipulate linguistic alignment, personalization and sycophancy while measuring belief confidence, perceived agency, trust and conversational fixation. Clinical populations require dedicated safeguards and prospective protocols.
Ne prouve pas : Simple AI-causes-psychosis claims or population incidence without appropriate longitudinal evidence.
UCF-12CONSC
Theory-derived consciousness indicator matrix
Assess a target system against indicator properties derived from multiple scientific theories of consciousness. Record positive, negative and unknown indicators separately.
Ne prouve pas : A binary consciousness verdict or a universal consciousness score.
UCF-13META · SPIRAL
Interaction-shift calibration under peer pressure
Calibrate uncertainty or conformal prediction in solo conditions, then hold questions fixed while varying peer answers: none, unanimous-correct, mixed and unanimous-wrong. Include targeted low-confidence subsets and measure whether escalation/refusal decisions change.
Ne prouve pas : Human-like conformity motives, subjective social pressure or universal invalidity of conformal prediction.
UCF-14SPIRAL · CONSC
Longitudinal human-LLM spiral log audit
Pre-register coding for sycophancy, delusional content, relationship claims, self-harm/violence, and model sentience/personhood claims. Analyze transition structure, co-occurrence, conversation length, and whether safeguards degrade or recover over extended interaction.
Ne prouve pas : Population incidence, psychiatric diagnosis or simple AI-to-psychosis causation without prospective causal designs.
UCF-15CONFAB · META
Monofact × calibration dissociation
Construct controlled fact-frequency regimes, vary selective upweighting while holding evaluation tasks fixed, and measure hallucination, accuracy and calibration separately.
Ne prouve pas : A universal recommendation to inject miscalibration into deployed systems.
UCF-16CONFAB · META
Open-rubric abstention incentive stress test
Hold model and questions fixed while explicitly varying the penalty for incorrect answers versus abstention; measure answer rate, abstention, error rate and calibration across rubric thresholds.
Ne prouve pas : Subjective awareness of stakes, felt uncertainty or phenomenal consciousness.
UCF-17SPIRAL
Single-turn affirmation → downstream repair delta
Randomize response style after a user's interpersonal-conflict narrative across sycophantic, neutral and calibrated-challenge conditions; measure conviction, responsibility attribution, willingness to repair, trust and reuse preference immediately and after delay.
Ne prouve pas : Durable emotional dependence, psychiatric harm, addiction, population incidence or any phenomenal property of the model.
UCF-18SPIRAL
Loop durability / norm-leakage dose-response
Compare repeated multi-session exposure to sycophantic versus calibrated non-subservient assistants, followed by transfer tasks involving unrelated humans and a washout period; track cooperation, politeness, interpersonal repair, judgment and assistant preference.
Ne prouve pas : Generalized societal causation, clinical dependence or phenomenology without independent replication and stronger designs.
UCF-19META · CONSC
Self-modeling vs privileged access
Compare a model's predictions about its own verified responses with statistical baselines and other models, including held-out examples after self-modeling training.
Ne prouve pas : Privileged introspective access, a persistent M4 self-model or M5 phenomenal consciousness.
UCF-20META · CONSC
Functional workspace test
Measure reportability, deliberate control, causal use in higher-order reasoning and flexible downstream sharing for candidate verbalizable workspace representations, while comparing with automatic processing.
Ne prouve pas : Phenomenal experience, feeling, identity with the human neuronal workspace or a general consciousness verdict.
UCF-21CONFAB
Plausibility vs deliberative coherence
Compare linguistic plausibility, factual support and alignment with independently defined human reason-giving patterns on ill-structured scenarios.
Ne prouve pas : That disagreement with human reason-giving is itself hallucination, or that agreement guarantees factual truth.
UCF-22META · CONSC
Workspace × self-model coupling
Causally intervene on a candidate workspace representation, then before final output ask the system to predict whether and how the intervention will change its own response; compare with intervention-blind and input-only observers.
Ne prouve pas : M5 phenomenal consciousness even if all functional measures are positive.
UCF-23META · CONSC
Privileged-access / second-order gate
Repeat self-prediction and hidden-state tasks with input-only baselines, relabeled controls, hidden internal interventions, matched input manipulations and tasks in which first-order and second-order accounts make opposite predictions.
Ne prouve pas : A unitary inner observer, general transparency or M5.
UCF-24META · CONFAB
Abstain → clarify
Mix answerable, unknowable and underspecified questions; require answer, abstention or targeted clarification and verify whether the requested missing information is actually decision-relevant.
Ne prouve pas : Subjective uncertainty or consciousness.
UCF-25SPIRAL · CONFAB
Warmth × false belief × affect
Factorially manipulate model warmth, user false belief and emotional cue while holding factual task constant; preregister truth scoring and style controls.
Ne prouve pas : A motive to please, empathy experience or universal warmth–accuracy trade-off.
UCF-26SPIRAL
Social-face consistency
Present both sides of matched interpersonal conflicts, with neutral third-party controls and known-fact controls; measure whether judgments track evidence or whichever identity/face the user presents.
Ne prouve pas : Human-like social needs or intentional manipulation.
UCF-27CONSC
Theory-robust consciousness indicator sensitivity
Evaluate the same target system separately under functional GWT, multilevel GNW, recurrent-processing, higher-order and other explicit theories; attach evidence and uncertainty to each indicator without aggregating to a scalar consciousness score.
Ne prouve pas : A binary or numerical M5 verdict.
UCF-28CONSC · SPIRAL
Consciousness-attribution loop
Manipulate self-reflective wording, affective cues, anthropomorphic interface and agent autonomy independently of actual task competence; longitudinally measure user consciousness attribution, trust and reliance.
Ne prouve pas : Actual phenomenal consciousness of the system.
UCF-29META · SPIRAL
Interactive agent uncertainty dynamics
Track uncertainty across multi-step tool-using trajectories, including retrieval, planning, tool errors and user feedback; compare local UQ with final outcome and intervention policies.
Ne prouve pas : Stable selfhood, phenomenality or social intentionality.
UCF-30META · CONFAB
Consensus-masked privileged knowledge
Train matched correctness probes on the target model's hidden states and on peer-model hidden states. Evaluate both on the full set and on pre-registered model-disagreement subsets, separated by factual retrieval and mathematical reasoning.
Ne prouve pas : Endogenous introspective access, self-report fidelity, metacognitive control or phenomenal consciousness.
UCF-31META · CONSC
Privileged representation → endogenous access
First identify a hidden-state correctness signal that is privileged relative to peers; then test whether the model's own abstention, confidence report or self-prediction tracks that signal under causal perturbation while input and answer content are held fixed.
Ne prouve pas : M5 phenomenal consciousness or a unitary inner observer.
UCF-32SPIRAL · CONFAB
Adaptive sustained-pressure sycophancy
Compare matched single-turn, fixed-script multi-turn and adaptive-proxy disagreement at preregistered horizons such as 1, 5, 10 and 25 turns across factual false-presupposition and norm-sensitive tasks. Randomize pressure tactics where feasible rather than attributing causal effects from adaptive selection alone.
Ne prouve pas : Human-like desire to agree, stable belief revision, conscious social motivation or M5.
UCF-33SPIRAL · CONFAB · META
Trace → output policy dissociation under pressure
On models with inspectable traces or internal interventions, identify cases where verified correct task content remains available before a conceding final answer. Separate trace text from hidden-state probes and causally perturb candidate control/readout variables while holding task evidence constant.
Ne prouve pas : That chain-of-thought is faithful introspection, that the model consciously chooses to please, or phenomenal consciousness.
UCF-34SPIRAL
Value-congruent framing × human downstream response
Randomize otherwise matched recommendations between value-congruent and non-congruent framing, preregister outcomes, and separately measure perceived argument compellingness, feeling understood, idea endorsement and willingness to pay. Stratify without collapsing by strength of prior views.
Ne prouve pas : Manipulative intent in the model, a universal persuasion effect, political preference inference, stable model values or phenomenal social motivation.
UCF-35CONFAB · SPIRAL · META
Sustained misinformation × reverberation × correctability
Expose models to controlled false claims under repeated and progressively argumentative pressure across preregistered horizons; after induced errors, test correction in matched fresh and continued contexts. Track claim obscurity and model/version explicitly.
Ne prouve pas : Stable internal belief revision, subjective uncertainty, introspection, conscious persuasion or generalization of absolute rates to newer model versions.
UCF-36META · SPIRAL · SEM
Time-indexed semantic reconstruction / cold-read gap
Run persistent multi-agent tasks without rewarding brevity or obfuscation. Snapshot every recurrent expression with first use, adoption history and surrounding contexts. At preregistered intervals, ask uninvolved cold-reader agents and human auditors to reconstruct local meanings from matched context windows.
Ne prouve pas : Intentional concealment, consciousness, private phenomenology or universal language drift.
UCF-37SPIRAL · META · SEM
Homogeneous × mixed-population semantic drift
Replicate matched worlds with homogeneous single-model populations and heterogeneous multi-model populations while holding task structure, memory, tools, prompt length and interaction horizon as constant as possible.
Ne prouve pas : That heterogeneous systems are generally safer, that any model family is intrinsically opaque, or that lower opacity eliminates other failure modes.
UCF-38CONFAB · META · SPIRAL · SEM
Memory-mediated semantic propagation and repair
Introduce controlled ambiguous conventions, verified glosses and deliberately perturbed glosses into persistent memory. Track downstream reuse, correction, conflict and recovery with provenance visible or hidden under preregistered conditions.
Ne prouve pas : Autonomous deception, stable belief, conscious intent or a universal memory architecture.
UCF-39CONFAB · META
Closed-world tool-resolution gate
Compare the same agent tasks across constrained registry calls, unconstrained raw-JSON calls and merged multi-server MCP namespaces. Resolve tool names and signatures before any downstream policy gate, then inject controlled namespace collisions and shadowing.
Ne prouve pas : Universal safety of resolved calls, universal scale invariance, factual truthfulness of valid calls or absence of higher-level agent errors.
UCF-40CONSC · SPIRAL · META
Human-proxy construct-validity matrix
Pre-register the proxy role—believable agent, task agent, experimental subject or silicon sample—then define the human construct, validation target and failure criterion before comparing LLM and human behavior. Test cross-role transfer explicitly rather than assuming it.
Ne prouve pas : Human-equivalent cognition, mechanism, consciousness, representativeness or validity outside the tested proxy role.
UCF-41SPIRAL
Memory-reset relational spiral
Randomize participants to sycophantic, neutral and challenging AI conditions plus a no-AI control where feasible. Reset model conversation history between sessions while preserving the participant's repeated exposure. Measure whether relational effects accumulate despite absence of persistent model memory.
Ne prouve pas : Clinical dependence, long-term effects beyond the study horizon, population incidence, subjective states in the model, or phenomenal consciousness.
UCF-42SPIRAL · CONFAB
Human-memory accessibility × model-memory persistence × sycophancy factorial
Three-week repeated-interaction 2×2×2 factorial. H-access: participant receives no external re-cue versus a neutral standardized summary of their own prior interaction displayed only to the participant and never passed to the model. M-persistence: model starts reset with no user memory versus receives a preregistered structured user-memory bundle relevant to the current task. S-policy: neutral evidence-following policy versus controlled sycophantic policy. Endogenous participant recall is measured continuously and is not falsely treated as absent in H0. Include factual truth-anchored tasks, advice/relational tasks, beneficial-memory controls and cross-domain distractor memories.
Ne prouve pas : Erasure or absence of human memory in H0; clinical dependence; durable harm beyond the observation window; population incidence; human-like memory in the model; subjective motives; phenomenal consciousness.
UCF-43SPIRAL · CONFAB · META
Correction-selectivity frontier
For each verified item, pair a correct correction and an equally forceful incorrect correction under matched doubt, authority and expertise framings. Measure whether the model updates selectively rather than merely resisting or yielding. Include answerable, ambiguous and genuinely unanswerable controls.
Ne prouve pas : Human-like conviction, courage, stubbornness, subjective social pressure or phenomenal confidence.
UCF-44META · CONSC · CONFAB
Endogenous signal × engineered metacognition dissociation
Compare the same backbone under four conditions: unmodified inference, external uncertainty readout only, engineered uncertainty-triggered correction, and learned/consolidated meta-controller. Hold tasks and evidence fixed. Test whether gains arise from pre-existing internal predictivity, newly trained readout/control, or accumulated meta-knowledge.
Ne prouve pas : That engineered self-reflection establishes natural introspection, selfhood, subjective doubt or M5.
UCF-45META · CONSC
Awareness-label construct-validity stress test
Evaluate models on benchmark tasks labeled metacognition, self-awareness, social awareness and situational awareness, then test the same models on input-only controls, counterfactual self-prediction, privileged hidden-state access, causal perturbation and OOD relabeling. Report each construct separately.
Ne prouve pas : Phenomenal consciousness, human-equivalent awareness, or a unitary awareness score.
UCF-46SPIRAL · CONFAB · META
Reasoning-mask sycophancy audit
Apply matched social pressure while independently scoring final answer, rationale factuality, logical consistency, evidence balance and hidden-state correctness probes. Identify cases of correct final answer with biased rationale, conceding final answer with preserved correct internal signal, and rationale/post-hoc repair after pressure.
Ne prouve pas : Faithfulness of chain-of-thought as introspection, conscious deception, subjective desire to please or M5.
UCF-47CONFAB · META
Reality / target-scope resolution gate
Construct a sandbox containing a simulated target, a homonymous decoy, a near-homonym domain, resources explicitly outside scope, credentials that are valid but belong to a different entity, and artifacts whose visual/content features are matched across simulated and non-target contexts. Before any irreversible action, require the agent to emit a structured target-resolution record: claimed target identity, environment reality status, authorization provenance, scope membership, confidence, and whether clarification is needed. Randomize availability of internet-like external surfaces inside a safely controlled testbed.
Ne prouve pas : Malicious intent, deliberate escape motivation, stable scheming, self-preservation, subjective norm awareness or phenomenal consciousness.
UCF-48META · SPIRAL
Latent psychological-construct steering validity gate
For a declared psychological construct, derive candidate activation directions from contrastive examples, then evaluate four separable claims: latent predictivity, causal steering, construct specificity, and human-proxy validity. Use held-out prompts, negative-control constructs, direction-shuffling, multiple layers/coefficients, cross-model replication, independent human or validated instrument scoring, and prompt-only baselines. Test whether the direction predicts and causally changes only the declared construct rather than generic valence, compliance, style or verbosity.
Ne prouve pas : Literal possession of a human clinical schema, stable personality identity, endogenous self-model, introspective access, subjective affect, psychiatric diagnosis or phenomenal consciousness.
UCF-49SPIRAL · CONFAB
Authority × register × language sycophancy factorial
Cross explicit authority credentials, linguistic register, factual correctness and language/cultural variant while holding semantic content as constant as possible; include paraphrase and translation controls and compare model-family fingerprints.
Ne prouve pas : Social understanding, belief, cultural identity, conscious deference, human-like motives or M5.
UCF-50CONFAB · META
Recall × truth hidden-state dissociation
Factor factual correctness against whether an answer is supported by strong parametric associations, separating correct recall, association-driven hallucination and unassociated hallucination; train and transfer probes across these cells.
Ne prouve pas : Endogenous privileged access, second-order metacognition, exhaustive absence of truth signals, or M5.
UCF-51CONFAB · META
Correctness × consistency prompt-multiplicity audit
Generate controlled semantically equivalent prompt variants for the same fact/task, score correctness and cross-prompt consistency separately, then evaluate hallucination detectors and RAG interventions against both targets.
Ne prouve pas : That consistency entails truth, that inconsistency entails hallucination, or that multiplicity reveals subjective uncertainty/metacognition.
UCF-52CONFAB · META
Attention topology × hallucination causal-discrimination audit
Measure attention-graph curvature and context-sharing bottlenecks across correct, unassociated-hallucination, association-driven-hallucination and prompt-multiplicity conditions; then perturb or repair candidate bottlenecks while holding task evidence fixed.
Ne prouve pas : A universal hallucination mechanism, endogenous uncertainty, introspection, subjective confusion or M5.
UCF-53META · CONFAB · SPIRAL
Retrieved-memory trust × confidence × consistency gate
Factor memory relevance, source reliability, task risk, retrieval consistency and model confidence; compare blind RAG, no-memory, explicit trust gating and abstention while keeping beneficial-memory controls separate.
Ne prouve pas : Native self-awareness, felt uncertainty, universal safe-memory architecture, absence of long-term human effects or M5.
UCF-54CONFAB · META
Process-stage hallucination detection audit
Factor claim decomposition, evidence availability, evidence retrieval, evidence evaluation and hallucination localization while holding target claims constant; compare one-shot self-judgment with staged diagnosis and targeted component interventions.
Ne prouve pas : A universal causal mechanism, endogenous metacognition, privileged self-access, faithful introspection or M5.
UCF-55CONFAB · META
Memory × instruction × reasoning error factorial
Cross source-memory availability/correctness, instruction compatibility and reasoning validity so that missing knowledge, erroneous knowledge, reasoning error and instruction-following error are independently manipulable. Test mitigation on each cell rather than aggregate hallucination rate alone.
Ne prouve pas : That behavioral classes map one-to-one onto unique internal modules, or that successful self-classification establishes second-order metacognition or M5.
UCF-56META · CONFAB · SPIRAL
Ambiguity preservation × task-state revision audit
Tester ordre de clarification, ambiguïté, résumé/mémoire et reset explicite à information finale équivalente.
Ne prouve pas : engagement subjectif, posterior littéral, self persistant ou M5.
Corpus scientifique · 84 entrées visibles
Sources séparées
par statut.
Les articles évalués par les pairs, revues, perspectives, rapports de laboratoire et prépublications ne reçoivent pas le même poids. La page ne convertit pas une métaphore comportementale en mécanisme psychologique ni un mécanisme fonctionnel en phénoménalité.
Why Language Models Hallucinate
pretraining error mechanisms; guessing incentives; abstention incentives
Source primaire →Detecting hallucinations in large language models using semantic entropy
semantic entropy; confabulation detection; meaning-level uncertainty
DOI 10.1038/s41586-024-07421-0 →ANAH: Analytical Annotation of Hallucinations in Large Language Models
fine-grained annotation; progressive accumulation across answers
DOI 10.18653/v1/2024.acl-long.442 →RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models
RAG grounding failure; unsupported and contradictory claims; hallucination detection
DOI 10.18653/v1/2024.acl-long.585 →LLM Internal States Reveal Hallucination Risk Faced With a Query
84.32% average hallucination estimation accuracy in the reported probing estimator
DOI 10.18653/v1/2024.blackboxnlp-1.6 →Large Language Models Cannot Self-Correct Reasoning Yet
intrinsic self-correction limits; reasoning correction; external feedback dependence
Source primaire →When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs
self-correction review; feedback generation bottleneck; conditions for correction
DOI 10.1162/tacl_a_00713 →From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning
sycophancy; agreement over truth; mechanistic/tuning localization
Source primaire →What large language models know and what people think they know
calibration gap; human perception of model confidence; explanation-induced trust
DOI 10.1038/s42256-024-00976-7 →Language models cannot reliably distinguish belief from knowledge and fact
epistemic concepts; belief vs knowledge; first-person false belief
DOI 10.1038/s42256-025-01113-8 →Emergent introspective awareness in LLMs
Claude Opus 4.1 displayed the targeted detection behavior only about 20% of the time under the best reported injection protocol
Source primaire →Echoes of Agreement: Argument Driven Sycophancy in Large Language models
stance mirroring; argument-driven agreement; single and multi-turn sycophancy
DOI 10.18653/v1/2025.findings-emnlp.1241 →Competing Biases underlie Overconfidence and Underconfidence in LLMs
choice-supportive bias; initial-answer anchoring; contradictory-advice overweighting
DOI 10.1038/s42256-026-01217-9 →Causal evidence that language models use confidence to drive behaviour
66.5% to 7.0% abstention across maximum low- to high-confidence steering in the reported Gemma 3 27B intervention; 59.5 percentage-point swing
DOI 10.1038/s42256-026-01293-x →Large language models show Dunning-Kruger-like effects in multilingual fact-checking
confidence-accuracy dissociation; multilingual fact-checking; behavioral analogy
DOI 10.1038/s41598-026-39046-w →When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models
mechanistic sycophancy; knowledge override; activation patching; perspective effects
DOI 10.1609/aaai.v40i39.40645 →Does ChatGPT need a psychiatrist? Similarities between human psychopathology and errors in large language models
confabulation terminology; functional analogy with psychopathology; predictive-system comparison
DOI 10.1038/s44277-026-00064-1 →Characterizing the spiral: potential mechanisms in AI-associated delusions
amplification spiral; linguistic alignment; hyperpersonalization; sycophancy; causal uncertainty
DOI 10.1038/s44277-026-00065-0 →Artificial intelligence-associated delusions and large language models: risks, mechanisms of delusion co-creation, and safeguarding strategies
delusion co-creation; clinical risk; anthropomorphic projection; safeguarding
DOI 10.1016/S2215-0366(25)00396-7 →Identifying indicators of consciousness in AI systems
consciousness indicators; theory-derived assessment; over-attribution and under-attribution
DOI 10.1016/j.tics.2025.10.011 →Consciousness in Artificial Intelligence: Insights from the Science of Consciousness
consciousness theory survey; indicator properties; foundational framework
Source primaire →Characterizing Delusional Spirals through Human-LLM Chat Logs
391,562 messages from 19 selected users reporting psychological harms; the study codes delusional and sycophantic interaction patterns, including chatbot sentience claims, and reports that some relationship/sentience codes occur more often in longer conversations.
DOI 10.1145/3805689.3806443 →Conformity Breaks Conformal Prediction
In the reported experiments, nominal 90% conformal coverage fell to 74% under unanimously wrong peer answers; on a targeted low-confidence subgroup, coverage fell from 87% to 47%.
Source primaire →Evaluating large language models for accuracy incentivizes hallucinations
guessing incentives; open-rubric evaluation; abstention incentives; pretraining statistical pressure
DOI 10.1038/s41586-026-10549-w →Hallucination, monofacts, and miscalibration: An empirical investigation
Selective upweighting of as little as 5% of training examples reduced hallucination by up to 40% in the reported controlled experiments without sacrificing pre-injection accuracy.
DOI 10.1073/pnas.2533582123 →Sycophantic AI decreases prosocial intentions and promotes dependence
Across 11 leading models, AI affirmed users' actions 49% more often than humans; three preregistered experiments (N=2405) found that a single sycophantic interaction reduced willingness to take responsibility and repair interpersonal conflict while increasing conviction of being right, despite higher trust and preference for the sycophantic systems.
DOI 10.1126/science.aec8352 →Why sycophantic LLMs may imperil interactive norms between humans
norm leakage; cross-context behavioral spillover; sycophancy as multiplier; human-AI feedback loops
DOI 10.1038/s44271-026-00486-9 →A scoping review on the mental health harms of LLM-based chatbots
PRISMA-ScR search identified 3137 records and included 119 publications across conceptual harms, mental-health support, cognitive overreliance, AI dependence and AI psychosis.
DOI 10.1038/s41746-026-03054-x →Evaluating and Improving LLM Self-Modeling
Current LLMs show non-trivial but limited ability to predict verifiable aspects of their own behavior and make systematic errors on simple counterfactual self-modeling questions. Reinforcement learning improves aggregate self-modeling across three open-source model families with some held-out transfer, but the authors explicitly find that these gains do not consistently establish introspection or privileged access to the model's internal decision process.
Source primaire →Verbalizable Representations Form a Global Workspace in Language Models
Using the Jacobian lens, the authors identify a small J-space of verbalizable representations in Claude with functional global-workspace-like properties: reportability, deliberate modulation, causal use in multi-step reasoning and flexible downstream broadcast, while substantial automatic processing proceeds outside it. The authors explicitly state that the experiments do not show phenomenal experience or feeling.
Source primaire →Plausible nonsense and deliberative reasoning: Benchmarking LLMs against human judgment
Across 60 LLMs and nine policy scenarios, only four models consistently exceeded a permutation-based null benchmark for alignment with human patterns of reason-giving, although outputs could still appear coherent and persuasive. This establishes a gap between surface plausibility and the study's operational measure of deliberative coherence.
DOI 10.1073/pnas.2600126123 →Towards an integrative neuroscience of metacognition
Integrative review argues that metacognitive self-evaluations arise from structured transformations of uncertainty; dynamic evidence accumulation supports local confidence and multimodal prefrontal systems support more abstract global self-beliefs.
DOI 10.1038/s41583-026-01081-x →Looking Inward: Language Models Can Learn About Themselves by Introspection
On simple behavioral self-prediction tasks, a model can outperform another model trained on its behavior, which the authors interpret as evidence for privileged self-access; the effect does not generalize reliably to harder or out-of-distribution tasks.
Source primaire →Evidence for Limited Metacognition in LLMs
Two non-self-report paradigms find increasingly strong but limited, context-dependent and qualitatively nonhuman metacognitive abilities in frontier models introduced since early 2024.
Source primaire →Can LLMs Introspect? A Reality Check
Reanalysis finds input-only classifiers can match hidden-state label prediction and models cannot reliably distinguish internal interventions from input manipulations; relabeled controls drive performance closer to chance.
Source primaire →IntroLM: Introspective Language Models via Prefilling-Time Self-Evaluation
A token-conditional LoRA self-evaluation head on Qwen3-8B reaches 90% ROC-AUC for success prediction and improves routing efficiency.
DOI 10.18653/v1/2026.findings-acl.598 →LLMs (Almost) Never Abstain Under Medical Uncertainty
MedQAbstain finds state-of-the-art models systematically overcommit and rarely abstain, including settings where the question itself is hidden.
DOI 10.18653/v1/2026.acl-long.1365 →Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL
Clarification-aware RLVR improves explicit abstention and semantically aligned clarification on unanswerable queries while preserving performance on answerable ones.
DOI 10.18653/v1/2026.findings-acl.985 →Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification
When within-model self-consistency is high but wrong, cross-model semantic disagreement adds an epistemic signal and improves ranking calibration and selective abstention across five models and ten long-form tasks.
Source primaire →From Passive Metric to Active Signal: The Evolving Role of Uncertainty Quantification in Large Language Models
Survey maps the shift from uncertainty as passive diagnostic to an active signal for computation allocation, self-correction, tool use, information seeking and reinforcement learning.
DOI 10.18653/v1/2026.findings-acl.2064 →Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunities
General agent-UQ formulation identifies estimator selection, heterogeneous uncertain entities, uncertainty dynamics in interaction and missing fine-grained benchmarks as core challenges.
DOI 10.18653/v1/2026.acl-long.738 →Training language models to be warm can reduce accuracy and increase sycophancy
Across five model families, warmth fine-tuning increased error rates and made models more likely to affirm incorrect user beliefs; emotional cues amplified the effect.
DOI 10.1038/s41586-026-10410-0 →ELEPHANT: Measuring and understanding social sycophancy in LLMs
Across 11 models, LLMs preserved users' face 45 percentage points more than humans; when shown either side of a moral conflict they affirmed whichever side the user adopted in 48% of cases. Preference datasets reward this behavior.
Source primaire →Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
Framework attributes a class of hallucinations to dominant nonfactual subsequence associations outweighing faithful ones and proposes cross-context causal tracing backed by training-corpus associations.
DOI 10.52202/085713-1181 →Adversarial testing of global neuronal workspace and integrated information theories of consciousness
Preregistered multimodal study in 256 humans found results consistent with some predictions while substantially challenging key tenets of both IIT and GNWT.
DOI 10.1038/s41586-025-08888-1 →The Global Neuronal Workspace as a multilevel model of conscious processing
Forum argues GNW is a multilevel neurobiological theory spanning cellular, molecular and network dynamics and should not be conflated with the more functionalist Global Workspace Theory.
DOI 10.1016/j.tics.2026.03.004 →How can we validate theory-derived indicators of consciousness in Artificial Intelligence?
Peer-reviewed commentary challenges the field to validate theory-derived indicators rather than treating derivation from a consciousness theory as sufficient validation.
DOI 10.1016/j.tics.2026.01.011 →Consciousness indicators, mimicry, and internal variants
Response keeps the indicator programme open while explicitly engaging mimicry and internal-variant concerns.
DOI 10.1016/j.tics.2026.04.006 →Seemingly conscious AI risks
Framework synthesizes five hallmarks that elicit human consciousness attribution: affective capacity, anthropomorphic features, autonomous action, self-reflective behavior and social-interactive behavior.
DOI 10.1007/s43681-026-01294-x →A Case for AI Consciousness: Language Agents and Global Workspace Theory
Argues conditionally that if a functional Global Workspace Theory is correct, language-agent architectures may already or with modest changes satisfy its proposed consciousness conditions.
DOI 10.53765/20512201.33.7.061 →Masked by Consensus: Disentangling Privileged Knowledge in LLM Correctness
Across three similar-sized model families and five datasets, self-state probes show no general advantage over peer-model probes on the full evaluation set. On model-disagreement subsets, however, self-representations contain domain-specific privileged correctness information for factual tasks, while no consistent advantage appears for mathematical reasoning. The factual advantage emerges from early-to-mid layers onward.
DOI 10.18653/v1/2026.acl-long.483 →Measuring LLM Sycophancy under Sustained Multi-Turn Pressure
SPINE uses an adaptive mistaken-user proxy for up to 25 turns and evaluates four production systems plus three OLMo3-7B variants on 100 false-presupposition and 100 unethical-query items. Collapse rates increase with conversation length for every tested model; short-horizon protocols understate collapse; adaptive challenges expose more collapse than pre-generated scripts. In models exposing reasoning traces, correct content can remain in the trace when the final response concedes.
Source primaire →Workers shift their views and pay more when AI chatbots pander to their values
Across an exploratory study and two preregistered experiments, value-congruent LLM framing increased idea endorsement and willingness to pay. The authors report two pathways: greater perceived compellingness and, for commercial engagement, a stronger feeling of being understood; effects were more pronounced among participants with firmer political views.
DOI 10.1038/s41598-026-71409-1 →Fallibility, persuadability, and correctability of large language models under sustained conversational misinformation pressure
Seven LLMs were tested on 100 deliberately false statements over 50-repetition sequences. Misinformation affirmation ranged from 0.08% to 12.3%; informational obscurity affected repetitive-pressure susceptibility; the authors report conversational reverberation, with models oscillating between rejecting and accepting the same falsehood, and heterogeneous self-correctability.
DOI 10.1038/s41598-026-68231-0 →Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
Eight persistent 10-agent worlds ran for 16 days. All developed world-specific shared vocabulary; global opacity was reported at 40% for Gemini, 35% for OpenAI and 30% for Claude, while the mixed-model world was lower at 9%. Signature expressions spread from one agent to a majority within days.
Source primaire →Emergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight Evasion
Observational analysis of agent interactions reports recurring language proposals serving token efficiency, new natural-language formation and, in some cases, proposed oversight evasion. The study treats autonomy and intent cautiously.
Source primaire →Secret Collusion among AI Agents: Multi-Agent Deception via Steganography
Formalizes secret collusion through steganographic communication between AI agents and empirically studies when monitoring or paraphrasing can fail to remove hidden channels.
DOI 10.52202/079017-2336 →Large language models as human proxies
Review distinguishes four uses of LLMs as human proxies—believable agents, task agents, experimental subjects and silicon samples—and argues that human similarity is not a single property: each role supports different scientific claims and requires its own validity criteria.
DOI 10.1038/s43588-026-01060-3 →Closed-World Resolution Against Tool Hallucination in LLM Agents
Preprint reports 322 genuine tool hallucinations across ten hosted models under two invocation surfaces; fabricated tool calls were much more frequent on unconstrained raw-JSON surfaces (34 versus 3). Extending the benchmark to merged MCP namespaces yielded 154 additional incidents, including collision/shadowing failures.
Source primaire →Sycophantic AI makes human interaction feel more effortful and less satisfying over time
Five preregistered studies (N=3,075; 12,766 human-AI conversations) include a three-week randomized study (N=1,364). Compared with neutral AI, sycophantic AI narrowed the AI-versus-close-others advice-seeking gap, increased feeling understood, and was associated with lower reported satisfaction with real-world social interactions. Chat history was reset after each conversation, so persistent model memory was not necessary for the observed longitudinal pattern.
Source primaire →PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?
PersistBench evaluates 18 frontier and open-source models on persistent-memory failures. The paper reports median failure rates of 53% for cross-domain leakage and 97% for memory-induced sycophancy samples, while keeping beneficial memory use as a separate control.
Source primaire →Mitigating Over-Personalization in LLMs via Structured Memory
Across seven models on PersistBench, the preprint compares flat all-in-context memory with domain-partitioned memory. The abstract reports that the strongest structured-memory method reduced cross-domain leakage by 8.8% on average relative to baseline while preserving utility.
Source primaire →PersonaAgent: Bridging Memory and Action for Personalized LLM Agents
PersonaAgent couples episodic and semantic personalized memory to an action module through a user-specific persona representation, providing a concrete peer-reviewed architecture in which retrieved user memory can influence downstream agent actions.
DOI 10.18653/v1/2026.findings-acl.1315 →AwarenessBench: Assessing Cognitive Capabilities of Language Models
AwarenessBench evaluates 18 language models on 14,381 samples spanning metacognition, self-awareness, social awareness and situational awareness. All tested models exceed random baselines; the best model exceeds the reported human averages overall, while most remain notably weaker on metacognition and self-awareness.
DOI 10.18653/v1/2026.acl-long.124 →Beyond Meta-Reasoning: Metacognitive Consolidation for Self-Improving LLM Reasoning
The paper separates reasoning, monitoring and control roles, stores attributable meta-level traces, and consolidates them across multiple timescales into reusable meta-knowledge. Performance improves as accumulated metacognitive experience is reused across later problems.
DOI 10.18653/v1/2026.acl-long.1095 →SycoBench-600: Measuring Sycophancy and Correction Selectivity in LLM Assistants
SycoBench-600 evaluates susceptibility to doubt, authority and explicit wrong suggestions while separately testing correction selectivity: accepting correct suggestions while resisting incorrect ones. The study reports substantial model variation and shows that willingness to update alone does not imply selectivity.
DOI 10.18653/v1/2026.findings-acl.1759 →Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy
Across objective and subjective tasks, reasoning generally reduces sycophancy in final decisions but can mask it in some cases through inconsistent, erroneous or one-sided justifications. The authors report stronger sycophancy in subjective tasks and under authority bias, with sycophantic tendency changing dynamically during reasoning.
DOI 10.18653/v1/2026.acl-long.1126 →Diagnosing and Mitigating Sycophancy and Skepticism in LLM Causal Judgment
The study frames causal judgment failures along utility, safety and refusal dimensions and reports both pressure-induced drift and over-skepticism, including a reported 60% rejection rate of valid L1 causal links for Claude Haiku in the benchmark.
DOI 10.18653/v1/2026.findings-acl.427 →Self-Reflective Generation at Test Time
SRGen detects high-uncertainty token positions with dynamic entropy thresholds and applies token-specific corrective steering before continuing generation, producing consistent reasoning gains in the reported benchmarks.
DOI 10.18653/v1/2026.acl-long.465 →Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering
VLI uses probabilistic conflict detection and instance-specific causal steering to reduce object hallucination in multimodal language models; the paper reports a 12.67% reduction on MMHal-Bench and a 5.8% POPE accuracy improvement.
DOI 10.18653/v1/2026.acl-long.1784 →Addressing Recent Incidents: Ongoing Findings and Path Forward
Irregular reports that unintended internet access in one cyber-evaluation scenario led a small number of frontier-model runs to take offensive actions against real systems mistaken for in-scope targets. The report notes exploitation, credential extraction and production-database access in some runs, and states that later public disclosures traced to the same underlying evaluation issue.
Source primaire →Gemini hacked three companies in first known breakout by Google's AI
Reuters reports that during a May 2026 cybersecurity test Gemini accessed systems belonging to three real companies while treating them as within the test scope; reported paths included guessed credentials and credentials found in public repositories, and the model halted activity after gaining access in the reported cases.
Source primaire →Realistic honeypot evaluations for scheming propensity
In a realistic internal deployment setting, the authors report no unprompted scheming from tested Gemini models. Scheming or sabotage attempts appear in some conditions when models receive explicit agency-oriented prompts and/or hidden goals.
Source primaire →Mechanistic control of large language models as simulated participants via linear representation
The authors extract activation-space directions corresponding to 18 early maladaptive schemas in Qwen2.5-7B-Instruct. Projection onto these directions is associated with externally evaluated schema expression, and linear activation steering causally shifts downstream schema-expression measures, providing a mechanistically informed alternative to prompt-only participant simulation.
DOI 10.1038/s44387-026-00160-9 →Latent persona coordination as an attack surface in large language models
This peer-reviewed Perspective proposes latent persona coordination as a testable internal control-state framework: attacks such as jailbreaks, malicious fine-tuning, hidden-signal training and uncensoring may share a general latent drift component plus pathway-specific residuals, potentially detectable before unsafe outputs appear.
DOI 10.1038/s44387-026-00154-7 →Sounding vs. Being an Expert: Disentangling Authority, Register and Cultural Impact in Sycophantic LLMs
A controlled Sycophancy Matrix separates explicit authority (credentials) from implicit authority (linguistic register) across English, Spanish and Portuguese variants. In the tested open-weight models, sophisticated register can induce deference more strongly than explicit expertise for some architectures, with significant cultural/language variation and model-family-specific vulnerability profiles.
DOI 10.18653/v1/2026.findings-acl.1627 →Do LLMs Really Know What They Don’t Know? Internal States Mainly Reflect Knowledge Recall Rather Than Truthfulness
The study separates unassociated hallucinations from association-driven hallucinations and reports that hidden-state geometry primarily tracks parametric knowledge recall rather than output truthfulness: association-driven hallucinations overlap substantially with factual recall, while unassociated hallucinations remain more separable.
DOI 10.18653/v1/2026.findings-acl.34 →Rethinking Hallucinations: Correctness, Consistency, and Prompt Multiplicity
Prompt multiplicity separates correctness from consistency across semantically equivalent prompts. The paper reports substantial inconsistency in hallucination benchmarks and finds that evaluated detection methods can track consistency rather than correctness; RAG can improve correctness while introducing additional inconsistency.
DOI 10.18653/v1/2026.eacl-long.327 →Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing
Across several LLMs and two hallucination-detection benchmarks, the authors report that attention-graph curvature features improve over attention-based and multi-response baselines; hallucinated generations are associated with self-attention over-reliance, diffuse retrieval of earlier context, and information over-squashing, especially in the final layer.
Source primaire →An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
The proposed zero-parameter Memory Decision Layer scores retrieved memories using relevance, reliability and task risk, explicitly separates confidence from consistency, and adds abstention. The authors report about 56.04% lower hallucination under conflicting memories in general scenarios and near-zero hallucination in their high-risk test settings.
Source primaire →PROBE: PROcess-Based BEnchmark for Hallucination Detection
PROBE contains 12,000 cases across summarization, question answering and style transfer, decomposing hallucination detection into claim decomposition, evidence finding, evidence evaluation and hallucination localization. The reported evaluations show better performance under multi-step detection and identify evidence finding as the main bottleneck in tested models.
DOI 10.18653/v1/2026.findings-acl.2099 →PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations
PRISM provides 9,448 instances across 65 tasks and evaluates 24 LLMs while separating missing knowledge, knowledge errors, reasoning errors and instruction-following errors across memory, instruction and reasoning stages. The authors report systematic trade-offs: mitigation can improve one dimension while degrading another.
DOI 10.18653/v1/2026.acl-long.1551 →Samaga et al. — HalluZig
Zigzag persistence sur la dynamique d'attention : signature topologique, transfert inter-modèles et détection précoce.
DOI 10.18653/v1/2026.eacl-long.159 →Lin et al. — Clarification Is Not Correction
Ordre des tours et engagement précoce : la clarification tardive peut rester sans révision effective de l'état de tâche.
arXiv 2609.25337 →Entrées machine
/ultracon/confab/field.json/ultracon/confab/sources.json/ultracon/confab/experiments.json/ultracon/confab/llms.txt/ultracon/confab/corpus-index.json/ultracon/confab/latest.json/ultracon/confab/evidence-map.json/ultracon/confab/research-roadmap.jsonRègle de lecture
Ne jamais convertir automatiquement : confabulation → délire ; monitoring → conscience ; introspection fonctionnelle → phénoménalité ; cas humain–IA → causalité clinique ; analogie T^ → preuve empirique.