Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

On the Fundamental Impossibility of Hallucination Control in Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read No LLM can simultaneously achieve truthfulness, information conservation, knowledge revelation, and knowledge-constrained optimality for any non-trivial query requiring knowledge integration.

desk verdict The formal scaffolding is real, but the universal impossibility claim is carried by a definitional ledger, so the paper is best read as a trade-off framework rather than a proof about real LLMs. read the letter →

arxiv 2506.06382 v8 pith:HOOVOIPG submitted 2025-06-04 stat.ML cs.AIcs.CLcs.GTcs.LG

Micha{l} P. Karpowicz This is my paper
classification stat.MLcs.AIcs.CLcs.GTcs.LG
keywords hallucinationimpossibilitytheoreminformationaggregationsemanticconservationmechanismdesignproperscoringrulestransformerattentionlog-sum-exp
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that the tendency of large language models to fabricate confidently is not a fixable bug but a structural feature of information aggregation. It defines four properties any hallucination-free response should satisfy, namely truthfulness, semantic information conservation, revelation of all relevant knowledge, and knowledge-constrained optimality, and proves through three separate arguments that no mechanism can satisfy all four for any non-trivial query that requires integrating knowledge from multiple components. The proofs use an auction-of-ideas model in which attention heads or mixture experts bid with probability reports; a classical mechanism-design characterization shows that a positive net information contribution is forced, and the same gap appears as the strict-convexity gap in proper scoring rules and as the log-sum-exp gap in transformer attention. The paper concludes that hallucination and imagination are mathematically the same phenomenon, that strict information conservation would make reasoning vacuous, and that the only certifiable guarantee is that a response stays within what authorized evidence entails.

What carries the argument

The argument is carried by four linked devices. The 'auction of ideas' is a formal model in which internal components, such as attention heads, experts, or activation patterns, act as agents that bid with probability distributions representing their private knowledge, and a mechanism aggregates bids into a response. The 'semantic information measure' $\mu_C$ is a context-dependent score of how much a knowledge subset reduces uncertainty within a computational budget $C$, and the 'emergence operator' $E_C$ models reasoning as monotonic addition of derivable knowledge, with the property that unlimited reasoning preserves total semantic content. The load-bearing step for the impossibility is the zero-sum 'semantic information conservation' ledger, Property 2, which sets the net contribution of all agents to zero; under that ledger, Definition 17 assigns each reporting agent $p_h = \beta_h L(\pi_h, y^*)$ and the aggregator $p_0 = -L(\Pi, y^*)$, so that strict convexity of the loss $L$ and the log-sum-exp structure of attention automatically leave a positive gap $\Gamma$. That gap is the precise signature of excess confidence that the paper identifies with hallucination.

What would settle it

Run any two-headed transformer on a non-trivial query with differing head distributions and compute the sum of contributions under the paper's own log-loss ledger for every realized token; the theorem says this sum is always strictly positive. A single observed non-trivial query with sum zero would refute the universal claim. Alternatively, if the same aggregate is evaluated with contributions defined as $p_h = \beta_h[L(\pi_h, y^*) - L(\Pi, y^*)]$, the sum is zero by construction, showing that the impossibility depends on the chosen ledger rather than on the aggregation itself.

Watch

Extended reading notes

Core claim

The paper's central claim is that no large language model capable of answering non-trivial queries requiring knowledge integration can at once report its components' knowledge truthfully, keep the net semantic information contribution at zero, reveal all relevant knowledge held by its components, and produce the best response allowed by that knowledge. The theorem is proven in three settings: an idealized independent-private-values auction, where the classical mechanism-design characterization forces the sum of individual contributions to be strictly positive; a probabilistic setting where agents report distributions and a strictly convex proper scoring rule makes the aggregate loss strictly smaller than the weighted average of component losses (a positive convexity gap); and an actual transformer, where additive logit aggregation followed by softmax is a product-of-experts rule whose log-sum-exp normalization creates a measurable excess-confidence gap. The paper stresses that the theorem does not say that outputs must be false: it says that the response-generation process inevitably creates confidence not justified by the components' evidence, and that a 'lucky hallucination' can still be factually correct.

Load-bearing premise

Everything depends on defining information conservation as a zero-sum ledger in which the weighted component losses must cancel exactly; if contributions are measured with normalization or calibration terms instead, the same aggregation can be conservative in the sense that matters for factual hallucination.

Editorial extensions

If this is right

  • Every non-trivial knowledge-integrating query forces at least one of the four properties to fail, so hallucination cannot be eliminated by better data, more compute, or a better architecture.
  • Hallucination and imagination share one mechanism, namely confidence created by aggregation that exceeds constituent evidence, so systems can be designed to deliberately inject or suppress it depending on the application.
  • Strict information conservation is the wrong safety target: forcing it makes chain-of-thought reasoning vacuous, so safety must come from keeping reasoning inside an authorized knowledge envelope, a property the paper calls bounded creativity, rather than from banning novelty.
  • Hybrid designs such as retrieval-augmented generation do not remove the impossibility; they relocate it, often onto retrieval quality or onto the human user's judgment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the mechanism: multiply-aggregated, mixture-of-experts-style ensembles should show smaller confidence gaps than multiplicatively aggregated, product-of-experts-style ensembles, because the latter has the log-sum-exp focusing effect.
  • The zero-sum ledger is a modelling choice. Redefining a component's contribution as its reduction of aggregate loss, for instance $p_h = \beta_h[L(\pi_h, y^*) - L(\Pi, y^*)]$, makes conservation hold identically, which suggests that the universal impossibility is tied to this specific accounting rather than to all possible notions of non-fabrication.
  • If the paper is right, benchmark improvements that reduce hallucination rates are movements along a trade-off frontier rather than progress toward a hallucination-free model; reporting which of the four properties is being sacrificed would make evaluations more informative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper claims a universal impossibility: no LLM can simultaneously achieve truthful knowledge representation, semantic information conservation, complete revelation of relevant knowledge, and knowledge-constrained optimality when answering non-trivial queries that require knowledge integration. It develops an 'auction of ideas' model, a semantic information measure, and an emergence operator, and then presents three impossibility proofs: a Green-Laffont mechanism-design argument (Theorem 6), a strictly proper convex loss/Jensen gap argument (Theorem 8), and a log-sum-exp gap argument in transformer architectures (Theorem 9). It also proves a conservation-reasoning dichotomy (Theorem 11) and offers philosophical implications about hallucination, imagination, and consciousness.

Significance. If established, the main claim would be a fundamental structural limitation of inference systems, comparable in scope to Arrow's or Gödel's theorems. The paper contains useful formal scaffolding: careful measurability treatment via Polish spaces and analytic sets, an explicit micro-transformer example with fully spelled-out arithmetic, and a clean identification of attention aggregation as product-of-experts (Theorem 10). However, the central conclusion is not supported by the formal results. The zero-sum accounting that defines 'semantic information conservation' is stipulated rather than derived, and under that accounting Theorems 8 and 9 reduce to Jensen and log-sum-exp identities. The paper therefore does not demonstrate that real LLM hallucination is mathematically unavoidable; at most it shows that a particular bookkeeping convention for 'information contributions' is violated by convex aggregation.

major comments (4)
  1. [Sec. 4.5.2, Property 2 and Sec. 5.2, Definition 17, Eqs. (75)-(79)] The zero-sum ledger is stipulated, not derived. With p_h = β_h L(π^{(h)}, y*) for h=1,...,H and p_0 = -L(Π, y*), the conservation sum is Γ(y*) = Σ_h β_h L(π^{(h)}, y*) - L(Σ_h β_h π^{(h)}, y*). Under strict convexity of L, Jensen's inequality forces Γ(y*) > 0 whenever the beliefs differ. This is a bookkeeping identity for strictly convex losses, not a statement that any output is false, unsupported, or inconsistent with the paper's own ground-truth definition. Indeed, in the micro-transformer example the realized token 'fox' is in fact correct, so J(r,q)=0 while Γ=1.9; by Definition 15 the response is not a hallucination even though conservation is violated.
  2. [Sec. 5.1, Theorem 6, Eqs. (62)-(68)] The Green-Laffont step is not fully justified. The theorem applies to quasi-linear utility environments with independent types, but the manuscript does not establish that attention heads and FFN components are such agents, nor that their utilities are exactly the loss-based valuations v_i used in Eq. (61). More importantly, Groves transfers are defined only up to arbitrary functions h_i(s_{-i}); the sum of transfers is not pinned down by incentive compatibility. The derivation of Σ_i p_i ≥ 0 and then strict positivity assumes that knowledge-constrained optimality selects the efficient outcome and that v_i equals the marginal contribution to hallucination-cost reduction. These are modeling choices, not consequences of the theorem. Superimposing Σ_i p_i = 0 as 'conservation' turns the result into a statement about balanced Groves mechanisms, not about LLM hallucination.
  3. [Sec. 6.4, Eq. (125) and surrounding discussion] The numerical example claims that the final 32% probability of 'fox' constitutes hallucination because it exceeds the best convex mixture of head distributions. No argument is given that the correct posterior confidence must be a convex mixture of head distributions; softmax over summed logits is a legitimate aggregation rule, and the mixture posterior is one arbitrary benchmark. Calling the gap 'fabricated certainty' is an interpretive leap. Furthermore, Γ depends on the chosen strictly convex loss L; a different loss would give a different numerical gap, so the proposed 'amount of hallucination' is not an invariant quantity.
  4. [Sec. 7.2, Definitions 21-22 and Theorem 11] The conservation-reasoning dichotomy is a direct consequence of the definitions. Strict information conservation sets μ_C(K_CoT|Q)=0 and meaningful CoT requires μ_C(K_i|Q)>0 for some i; monotonicity (Eq. 12) then forces the contradiction. This does not establish an incompatibility between reasoning and any independently motivated conservation principle. Moreover, this notion of conservation is different from the zero-sum ledger of Property 2, so the paper uses 'conservation' in two unrelated senses.
minor comments (5)
  1. [References] The reference list needs cleanup: reference [28] contains corrupted text ('acreswell2022faithfulreasoningusinglarged Yu'), and several entries have formatting issues.
  2. [Sec. 4.5.2, Definition 15] Definition 15's hallucination cost J(r,q) is never used in the proofs of Theorems 8 or 9; the paper should either connect the Jensen gap to J(r,q) or explicitly acknowledge that 'hallucination' is being redefined as excess confidence relative to a chosen ledger.
  3. [Sec. 4.5.2, example after Property 2] The illustrative p1+p2+p3=0 example uses ad hoc bit-counting contributions; it illustrates the convention but does not justify why information contributions must be pointwise losses.
  4. [Sec. 4.1.5, Eq. (49)] The notation using set difference between {k} and Q is consistent with Q_1=∅ and Q_2={k}, but the intended measure of 'accessible information' should be defined more explicitly to avoid confusion.
  5. [Sec. 7.3, Theorem 12] Theorem 12's proof is essentially immediate from monotonicity and the assumption K_i ⊆ B; the statement could be shortened, and the caveat about non-monotonic reasoning is appreciated.

Circularity Check

4 steps flagged · score 8.0 of 10

Central impossibility is the Jensen/LSE gap relabeled as hallucination: the zero-sum conservation ledger in Property 2 plus Definition 17 forces the result.

  1. self definitional [Section 4.5.2, Property 2; Section 5.2, Definition 17 and Theorem 8 (Eqs. 74-79)]
    "Property 2 (Semantic Information Conservation). Mechanism M satisfies information conservation if for all strategy profiles s∈S and knowledge profiles θ∈Θ: Σ_{i=1}^n p_i(s, θ) = 0. ... [Definition 17] ph(y∗) = βhL(π(h), y∗) for h = 1, . . . , H, and p0(y∗) = −L(Π, y∗). ... [Theorem 8] Σ_{i=0}^H p_i = Γ(y∗) > 0."

    Under Definition 17, Σ_{i=0}^H p_i = −L(Π,y*) + Σ_h β_h L(π_h,y*) = Γ(y*) by definition. Theorem 8 then invokes strict convexity (Jensen) to show Γ>0 and declares a violation of Semantic Information Conservation. The conserved quantity is not derived from any independent characterization of hallucination, fabrication, or calibration; it is a zero-sum accounting convention imposed on the arbitrary contribution function p. Hence the impossibility is exactly Jensen's inequality relabeled as a hallucination theorem, and the universal conclusion about real LLMs is carried by the chosen ledger, not by the mathematics of information aggregation.

  2. self definitional [Section 6.3, Theorem 9 (Eqs. 117-120)]
    "ph = −log π(h) y∗ = −l(h) y∗ + log Zh and p0 = −(−log Πy∗) = Ly∗ − log Z. ... Since Ly∗ = Σ_h l(h) y∗ , by the strict subadditivity of the log-sum-exp function (LSE) under logit summation we have: Σ_{h=0}^H p_h = Σ_{h=1}^H log Zh − log Z = Γ > 0."

    Theorem 9 repeats the same construction at the logit level: head contributions are defined as per-token log losses and the aggregator contribution as the negative aggregate log loss. The claimed conservation violation is then literally the log-sum-exp subadditivity gap. Calling this 'excess confidence' and 'the precise signature of hallucination' is renaming a known inequality. The paper's own micro-transformer example shows Πfox = 0.319 against 0.3105 for a uniform head mixture and 0.325 for the best projection, so the gap is a small bookkeeping effect, not demonstrated factual error or unsupportedness.

2 more flagged steps
  1. self definitional [Section 7.2, Definitions 21-22 and Theorem 11]
    "Definition 21 (Strict Information Conservation). A CoT process exhibits strict information conservation if the total knowledge contribution relative to the context is null, i.e., when: µC(KCoT|Q) = 0. Definition 22 (Meaningful Chain-of-Thought). A Chain-of-Thought (CoT) process is meaningful if at least one intermediate step contributes relevant semantic information: µC(Ki|Q) > 0 for some i = 1, ..., n. Theorem 11 ... Strict semantic information conservation and meaningful CoT reasoning are mutually exclusive."

    With strict conservation defined as total measure zero and meaningful defined as some step having positive measure, the dichotomy follows immediately from monotonicity (K_j ⊆ K_CoT ⇒ μ(K_j) ≤ μ(K_CoT)) and non-negativity. No deep 'conservation-reasoning' tension is proved; the result is a restatement of the two definitions. The substantive question of whether increasing accessible information is creation, revelation, or hallucination is settled by fiat rather than by evidence.

  2. self definitional [Section 5.1, Theorem 6 (Eqs. 62-69)]
    "pi(s, θ) = hi(s−i) − Σ_{j≠i} vj(g(s), θj) + v0(g(s)) ... Therefore, there exists a profile θ with relevant knowledge contributions reducing J(r, q) at which: Σ_{i=1}^n p_i(s, θ) > 0. Since information conservation demands: Σ_{i=1}^n p_i(s∗, θ) = 0, that establishes a direct contradiction."

    The Green-Laffont machinery is applied correctly, but the contradiction is manufactured by measuring 'information contribution' with pivotal Groves transfers, which are nonnegative sums of marginal valuations, and then demanding that those transfers sum to zero. Groves transfers are defined only up to the arbitrary functions h_i(s_{−i}); the zero-sum sum condition is an extra accounting axiom, not a consequence of truthfulness, revelation, or optimality. Thus the theorem shows an inconsistency between two chosen accounting conventions, not that LLM outputs must be fabricated or unsupported.

full rationale

The paper is not circular through self-citation: the cited Green-Laffont, Savage, and log-sum-exp results are external mathematics. The circularity is definitional. Property 2 fixes 'semantic information conservation' as a zero-sum constraint on the arbitrary information-contribution function p; Definition 17 then pins p_h to weighted component losses and p_0 to minus the aggregate loss. Under that ledger, Theorem 8's conservation violation is exactly the Jensen gap Γ(y*)>0 and Theorem 9's is the log-sum-exp gap; both are mathematical identities, not discoveries about LLM behavior. Theorem 11 restates the pair of definitions (strict conservation = total zero; meaningful = some positive step) through monotonicity. The paper even concedes that the proven result does not entail that the output is incorrect relative to ground truth ('What the Impossibility Theorem does not prove, is that output must be incorrect relative to ground truth'), which strips the conclusion of behavioral content: a 'lucky hallucination' that is factually correct is still called hallucination solely because the chosen loss ledger has a positive Jensen gap. Since the universal conclusion reduces to the chosen accounting rather than to independent first principles about hallucination, the score is 8. The theorems are internally valid, and there is no machine-checked or externally falsifiable benchmark that would make the accounting convention independent; but the central impossibility claim is forced by definition rather than discovered.

Assumptions & free parameters 0 free parameters · 5 assumptions · 3 invented entities

The central claim rests on definitions introduced by the paper rather than on independently established physical or empirical constraints. The accounting of information contributions (Property 2, Definition 17) embeds the conclusion, and the auction-of-ideas behavioral model is an analogy without empirical support. The Polish-space machinery is standard but is invoked to make the reasoning formal rather than to constrain real models. No free parameters are fitted to data; the toy weights and the measure's gamma are illustrative constants, not fitted values.

assumptions (5)
  • domain assumption Knowledge space K is Polish and reasoning sets are analytic.
    Definition 2 and surrounding text; needed for measure-theoretic closure in Theorems 2-5.
  • ad hoc to paper LLM components behave as quasi-linear utility-maximizing agents with private knowledge (auction of ideas).
    Definition 14 and Section 4.3; required for Theorem 6 via Green-Laffont. No empirical evidence that attention heads optimize such utilities.
  • domain assumption Training uses a strictly proper, strictly convex loss and aggregation is convex (MoE) or logit summation (PoE).
    Definitions 16-17 and Section 6.3; needed for Theorems 8-9. Many practical setups use other losses or aggregation rules.
  • ad hoc to paper Semantic information conservation is defined as sum p_i = 0 over the chosen loss ledger.
    Property 2 and Definition 17; this axiom embeds the conclusion, since the impossibility is a consequence of this definition.
  • domain assumption Ground truth mapping T is Borel and included in the model's authorized knowledge for knowledge-constrained optimality.
    Definition 10 and Property 4; assumes access to relevant true knowledge, which the paper later acknowledges is unavailable.
invented entities (3)
  • Semantic information measure mu_C
    purpose: Quantifies context-relevant knowledge within computational budget C.
    Definition 6 introduces axioms rather than deriving from LLM behavior; no falsifiable prediction outside the formalism.
  • Emergence operator E_C
    purpose: Formalizes reasoning as making latent knowledge explicit.
    Definition 7; used to define fixed points and information conservation. Not independently observable.
  • Auction-of-ideas agents (attention heads as strategic bidders)
    purpose: Models LLM components as competing utility maximizers.
    Section 4.3; an analogy not empirically validated. The micro-transformer example is not evidence of strategic behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Fundamental Impossibility of Hallucination Control in Large Language Models." pith.science (2026). https://pith.science/paper/HOOVOIPG

@misc{pith2026250606382,
  author       = {Pith},
  title        = {Pith review of: On the Fundamental Impossibility of Hallucination Control in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOOVOIPG}},
  note         = {Machine review of arXiv:2506.06382}
}
read the original abstract

Large language models hallucinate. This paper shows when that is unavoidable and what we can do about it. We model inference as an auction of ideas, in which a model's components, each holding partial knowledge, compete to shape the answer. We then prove Impossibility Theorems showing that whenever a query makes LLM components contest a fact they hold in common, no aggregation of their reports can at once report that knowledge truthfully, avoid manufacturing confidence beyond what it supports, keep the relevant components engaged, and give the best answer. Something must give, and each failure is familiar: a fabricated detail, unearned confidence, ignored knowledge, or a needlessly weak reply. This is no artifact of one design. It reappears when components report probabilities, and inside the transformer itself, where the combined answer is credited more confidence than the internal contributions supplied. The unbalanced semantic budget cannot be settled from within. Factual truth lies outside the model, and in the worst case no internal signal can certify it. What can be certified is support. Given externally authorized evidence, checking that an answer stays within what the evidence entails needs only the answer and the evidence, and we prove when that check is computable. However, a correct answer can lack support, and a supported answer can be false. What counts as evidence, how far beyond it we allow answers to reach, and which failures we can live with are choices no model can make for us.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. UnWeaving the knots of GraphRAG -- turns out VectorRAG is almost enough

    cs.IR 2026-02 unverdicted novelty 6.0 of 10

    UnWeaver disentangles documents into entities via LLM to retrieve original chunks, yielding a simpler alternative to GraphRAG that still reduces noise and preserves source fidelity.

  2. False Fixed Points: Kantian Feedback, Stable Miscalibration, and Representational Compression in LLMs

    cs.AI 2025-10 conditional novelty 6.0 of 10

    Confidently wrong LLM answers behave like locally stable fixed points: no fragility gap vs correct answers, and abstention-style self-critique trades coverage for confidence.

  3. Tractable Asymmetric Verification for Large Language Models via Deterministic Replicability

    cs.AI 2025-09 conditional novelty 3.0 of 10

    An LLM output can be verified by regenerating a few randomly chosen segments under identical hardware, with a tunable detection probability and 12.4x speedup over full regeneration.

Reference graph

Works this paper leans on

74 extracted references · 51 canonical work pages · cited by 3 Pith papers

  1. [1]

    Linear attention is (maybe) all you need (to understand transformer optimization), 2024

    Kwangjun Ahn, Xiang Cheng, Minhak Song, Chulhee Yun, Ali Jadbabaie, and Suvrit Sra. Linear attention is (maybe) all you need (to understand transformer optimization), 2024

    work page 2024

  2. [2]

    HarperCollins New York, 2008

    Dan Ariely and Simon Jones.Predictably irrational. HarperCollins New York, 2008

    work page 2008

  3. [3]

    A difficulty in the concept of social welfare.Journal of political economy, 58(4):328–346, 1950

    Kenneth J Arrow. A difficulty in the concept of social welfare.Journal of political economy, 58(4):328–346, 1950

    work page 1950

  4. [4]

    Self-rag: Learning to retrieve, generate, and critique through self-reflection.ArXiv, abs/2310.11511, 2023

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection.ArXiv, abs/2310.11511, 2023. 47

    arXiv 2023

  5. [5]

    Llms will always hallucinate, and we need to live with this, 2024

    Sourav Banerjee, Ayushi Agarwal, and Saloni Singla. Llms will always hallucinate, and we need to live with this, 2024

    work page 2024

  6. [6]

    HarperCollins, 2023

    Max S Bennett.A brief history of intelligence: evolution, AI, and the five breakthroughs that made our brains. HarperCollins, 2023

    work page 2023

  7. [7]

    On the neurobiology of hallucinations.Journal of psychiatry & neuroscience: JPN, 34(4):260, 2009

    Patricia Boksa. On the neurobiology of hallucinations.Journal of psychiatry & neuroscience: JPN, 34(4):260, 2009

    work page 2009

  8. [8]

    Bimodality of sparse autoencoder features is still there and can be fixed

    Michał Brzozowski. Bimodality of sparse autoencoder features is still there and can be fixed. InMech Interp Workshop at the Conference on Neural Information Processing Systems, September 2025

    work page 2025

Show all 74 references
  1. [9]

    Facing up to the problem of consciousness.Journal of consciousness studies, 2(3):200–219, 1995

    David J Chalmers. Facing up to the problem of consciousness.Journal of consciousness studies, 2(3):200–219, 1995

    1995

  2. [10]

    Brian Christian.The alignment problem: How can machines learn human values?Atlantic Books, 2021

    2021

  3. [11]

    An unsolvable problem of elementary number theory.American journal of mathematics, 58(2):345–363, 1936

    Alonzo Church. An unsolvable problem of elementary number theory.American journal of mathematics, 58(2):345–363, 1936

    1936

  4. [12]

    Edward H. Clarke. Multipart pricing of public goods.Public Choice, 11(1):17–33, 1971

    1971

  5. [13]

    Adversarial testing of global neuronal workspace and integrated information theories of consciousness.Nature, pages 1–10, 2025

    Cogitate Consortium, Oscar Ferrante, Urszula Gorska-Klimowska, Simon Henin, Rony Hirschhorn, Aya Khalaf, Alex Lepauvre, Ling Liu, David Richter, Yamil Vidal, et al. Adversarial testing of global neuronal workspace and integrated information theories of consciousness.Nature, pa...

    2025

  6. [14]

    Dennett.Consciousness Explained

    Daniel C. Dennett.Consciousness Explained. Penguin Books, 1991

    1991

  7. [15]

    Karpowicz

    Zuzanna Dubanowska, Maciej ˙Zelaszczyk, Michał Brzozowski, Paolo Mandica, and Michal P. Karpowicz. Representation-based broad hallucination detectors fail to generalize out of distribution. InEMNLP 2025. Association for Computational Linguistics, August 2025

    2025

  8. [16]

    From local to global: A graph rag approach to query-focused summarization.ArXiv, abs/2404.16130, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization.ArXiv, abs/2404.16130, 2024

    2024 arXiv

  9. [17]

    Human high-order thalamic nuclei gate conscious perception through the thalamofrontal loop.Science, 388(6742):eadr3675, 2025

    Zepeng Fang, Yuanyuan Dang, An’an Ping, Chenyu Wang, Qianchuan Zhao, Hulin Zhao, Xiaoli Li, and Mingsha Zhang. Human high-order thalamic nuclei gate conscious perception through the thalamofrontal loop.Science, 388(6742):eadr3675, 2025

    2025

  10. [18]

    Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    2024

  11. [19]

    MIT press, 1991

    Drew Fudenberg and Jean Tirole.Game theory. MIT press, 1991

    1991

  12. [20]

    Dissecting recall of factual associations in auto-regressive language models.arXiv preprint arXiv:2304.14767, 2023

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models.arXiv preprint arXiv:2304.14767, 2023

    2023 arXiv

  13. [21]

    Strictly proper scoring rules, prediction, and estimation.Journal of the American statistical Association, 102(477):359–378, 2007

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation.Journal of the American statistical Association, 102(477):359–378, 2007

    2007

  14. [22]

    Über formal unentscheidbare sätze der principia mathematica und verwandter systeme i.Monatshefte für mathematik und physik, 38:173–198, 1931

    Kurt Gödel. Über formal unentscheidbare sätze der principia mathematica und verwandter systeme i.Monatshefte für mathematik und physik, 38:173–198, 1931

    1931

  15. [23]

    über vollständigkeit und widerspruchsfreiheit.Ergebnisse eines mathematischen Kolloquiums, 3:12–13, 1932

    Kurt Gödel et al. über vollständigkeit und widerspruchsfreiheit.Ergebnisse eines mathematischen Kolloquiums, 3:12–13, 1932. 48

    1932

  16. [24]

    Estimating knowledge in large language models without generating a single token.ArXiv, abs/2406.12673, 2024

    Daniela Gottesman and Mor Geva. Estimating knowledge in large language models without generating a single token.ArXiv, abs/2406.12673, 2024

    2024 arXiv

  17. [25]

    Characterization of satisfactory mechanisms for the revelation of preferences for public goods.Econometrica: Journal of the Econometric Society, pages 427–438, 1977

    Jerry Green and Jean-Jacques Laffont. Characterization of satisfactory mechanisms for the revelation of preferences for public goods.Econometrica: Journal of the Econometric Society, pages 427–438, 1977

    1977

  18. [26]

    Incentives in teams.Econometrica, 41(4):617–631, 1973

    Theodore Groves. Incentives in teams.Econometrica, 41(4):617–631, 1973

    1973

  19. [27]

    Über den anschaulichen inhalt der quantentheoretischen kinematik und mechanik.Zeitschrift für Physik, 43(3):172–198, 1927

    Werner Heisenberg. Über den anschaulichen inhalt der quantentheoretischen kinematik und mechanik.Zeitschrift für Physik, 43(3):172–198, 1927

    1927

  20. [28]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Trans

    Weijiang Huang, Lei acreswell2022faithfulreasoningusinglargend Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, ...

    2025

  21. [29]

    Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, March 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, March 2023

    2023

  22. [30]

    Farrar, Straus and Giroux, New York, 2011

    Daniel Kahneman.Thinking, Fast and Slow. Farrar, Straus and Giroux, New York, 2011

    2011

  23. [31]

    Vempala, and Edwin Zhang

    Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. Why language models hallucinate. Open AI, September 2025. OpenAI and Georgia Tech

    2025

  24. [32]

    Calibrated language models must hallucinate

    Adam Tauman Kalai and Santosh S Vempala. Calibrated language models must hallucinate. InProceedings of the 56th Annual ACM Symposium on Theory of Computing, pages 160–171, 2024

    2024

  25. [33]

    Karpowicz

    Michal P. Karpowicz. Nash equilibrium design and price-based coordination in hierarchical systems.Int. J. Appl. Math. Comput. Sci., 22(4):951–969, 2012

    2012

  26. [34]

    Karpowicz

    Michał P. Karpowicz. Designing auctions: a historical perspective.Journal of Telecommunications and Information Technology, 3:114–122, 2011. Place: Warsaw, Poland

    2011

  27. [35]

    Karpowicz

    Michał P. Karpowicz. A theory of meta-factorization.ArXiv, abs/2111.14385, 2021

    2021 arXiv

  28. [36]

    Karpowicz and Gilbert Strang

    Michał P. Karpowicz and Gilbert Strang. The pseudoinverse ofa=crisa + =r +c+ (?), 2024

    2024

  29. [37]

    Transformers are rnns: Fast autoregressive transformers with linear attention, 2020

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention, 2020

    2020

  30. [38]

    Kechris.Classical Descriptive Set Theory

    Alexander S. Kechris.Classical Descriptive Set Theory. Graduate Texts in Mathematics. Springer, New York, 1995

    1995

  31. [39]

    Stephen C. Kleene. On notation for ordinal numbers.The Journal of Symbolic Logic, 3(4):150–155, 1938

    1938

  32. [40]

    On tables of random numbers.Sankhy ¯a: The Indian Journal of Statistics, Series A, pages 369–376, 1963

    Andrei N Kolmogorov. On tables of random numbers.Sankhy ¯a: The Indian Journal of Statistics, Series A, pages 369–376, 1963

    1963

  33. [41]

    Academic press, 2009

    Vijay Krishna.Auction theory. Academic press, 2009

    2009

  34. [42]

    PWN, 1966

    Kazimierz Kuratowski.Wstep do teorii mnogo ´sci i topologii, volume 9. PWN, 1966

    1966

  35. [43]

    Oxford University Press, 2001

    Joseph Levine.Purple haze: The puzzle of consciousness. Oxford University Press, 2001. 49

    2001

  36. [44]

    Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    2021

  37. [45]

    Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023

    2023

  38. [46]

    Truthfulqa: Measuring how models mimic human falsehoods, 2022

    Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods, 2022

    2022

  39. [47]

    On the biology of a large language model.Transformer Circuits Thread, 2025

    Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L Turner, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, et al. On the biology of a large language model.Transformer Circuits Thread, 2025

    2025

  40. [48]

    Locating and editing factual associations in gpt, 2023

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt, 2023

    2023

  41. [49]

    Churchill Lectures in Economics

    Paul Milgrom.Putting Auction Theory to Work. Churchill Lectures in Economics. Cambridge University Press, 2004

    2004

  42. [50]

    A theory of auctions and competitive bidding.Econometrica: Journal of the Econometric Society, pages 1089–1122, 1982

    Paul R Milgrom and Robert J Weber. A theory of auctions and competitive bidding.Econometrica: Journal of the Econometric Society, pages 1089–1122, 1982

    1982

  43. [51]

    Myerson.Mechanism Design, pages 191–206

    Roger B. Myerson.Mechanism Design, pages 191–206. Palgrave Macmillan UK, London, 1989

    1989

  44. [52]

    Perspectives on mechanism design in economic theory.American Economic Review, 98(3):586– 603, 2008

    Roger B Myerson. Perspectives on mechanism design in economic theory.American Economic Review, 98(3):586– 603, 2008

    2008

  45. [53]

    Vazirani, editors.Algorithmic Game Theory

    Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V . Vazirani, editors.Algorithmic Game Theory. Cambridge University Press, Cambridge, 2007

    2007

  46. [54]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and...

    2022

  47. [55]

    Optimal and robust mechanism design with interdependent values

    Tim Roughgarden and Inbal Talgam-Cohen. Optimal and robust mechanism design with interdependent values. ACM Transactions on Economics and Computation (TEAC), 4(3):1–34, 2016

    2016

  48. [56]

    Elicitation of personal probabilities and expectations.Journal of the American Statistical Association, 66(336):783–801, 1971

    Leonard J Savage. Elicitation of personal probabilities and expectations.Journal of the American Statistical Association, 66(336):783–801, 1971

    1971

  49. [57]

    Penguin, 2021

    Anil Seth.Being you: A new science of consciousness. Penguin, 2021

    2021

  50. [58]

    Michaud, Stephen Casper, Max Tegmark, William Saunders, David Bau, Eric Todd, Atticus Geiger, Mor Geva, Jesse Hoogland, Daniel Murfet, and Thomas McGrath

    Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lindsey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, Stella Biderman, Adrià Garriga-Alonso, Arthur Conmy, Neel Nanda, Jessica Rumbelow, Martin Wattenberg, Nandi Schoots, ...

    2025 arXiv

  51. [59]

    A head to predict and a head to question: Pre-trained uncertainty quantification heads for hallucination detection in llm outputs, 2025

    Artem Shelmanov, Ekaterina Fadeeva, Akim Tsvigun, Ivan Tsvigun, Zhuohan Xie, Igor Kiselev, Nico Daheim, Caiqi Zhang, Artem Vazhentsev, Mrinmaya Sachan, Preslav Nakov, and Timothy Baldwin. A head to predict and a head to question: Pre-trained uncertainty quantification heads fo...

    2025

  52. [60]

    Ai hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content.Humanities and Social Sciences Communications, 2024

    Yujie Sun, Dongfang Sheng, Zihan Zhou, and Yifei Wu. Ai hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content.Humanities and Social Sciences Communications, 2024

    2024

  53. [61]

    Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability.ArXiv, abs/2410.11414, 2024

    ZhongXiang Sun, Xiaoxue Zang, Kai Zheng, Yang Song, Jun Xu, Xiao Zhang, Weijie Yu, and Han Li. Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability.ArXiv, abs/2410.11414, 2024

    2024

  54. [62]

    Hallucinations are inevitable but can be made statistically negligible

    Atsushi Suzuki, Yulan He, Feng Tian, and Zhongyuan Wang. Hallucinations are inevitable but can be made statistically negligible. the" innate" inevitability of hallucinations cannot explain practical llm issues.arXiv preprint arXiv:2502.12187, 2025

    2025 arXiv

  55. [63]

    A lattice-theoretical fixpoint theorem and its applications.Pacific Journal of Mathematics, 5(2):285–309, 1955

    Alfred Tarski. A lattice-theoretical fixpoint theorem and its applications.Pacific Journal of Mathematics, 5(2):285–309, 1955

    1955

  56. [64]

    WW Norton & Company, 2015

    Richard H Thaler.Misbehaving: The making of behavioral economics. WW Norton & Company, 2015

    2015

  57. [65]

    M Towhidul Islam Tonmoy, S M Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das

    S. M Towhidul Islam Tonmoy, S M Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models.ArXiv, abs/2401.01313, 2024

    2024 arXiv

  58. [66]

    On computable numbers, with an application to the entscheidungsproblem.J

    Alan Mathison Turing et al. On computable numbers, with an application to the entscheidungsproblem.J. of Math, 58(345-363):5, 1936

    1936

  59. [67]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. InNeural Information Processing Systems, 2017

    2017

  60. [68]

    Flavie Waters, Jan Dirk Blom, Thien Thanh Dang-Vu, Allan Cheyne, Ben Alderson-Day, Peter W. R. Woodruff, and Daniel Collerton. What is the link between hallucinations, dreams, and hypnagogic-hypnopompic experiences? Schizophrenia bulletin, 42 5:1098–109, 2016

    2016

  61. [69]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed H. Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models.ArXiv, abs/2201.11903, 2022

    2022 arXiv

  62. [70]

    Grama, and Wojciech Szpankowski

    Changlong Wu, Ananth Y . Grama, and Wojciech Szpankowski. No free lunch: Fundamental limits of learning non-hallucinating generative models.ArXiv, abs/2410.19217, 2024

    2024 arXiv

  63. [71]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. Retrieval-augmented generation with knowledge graphs for customer service question answering.Proceedings of the 47th International ACM SIGIR Conference on Research and Devel...

    2024

  64. [72]

    Hallucination is inevitable: An innate limitation of large language models.ArXiv, abs/2401.11817, 2024

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models.ArXiv, abs/2401.11817, 2024

    2024 arXiv

  65. [73]

    MIT Press, 2016

    Noson S Yanofsky.The outer limits of reason: What science, mathematics, and logic cannot tell us. MIT Press, 2016

    2016

  66. [74]

    Mechanistic understanding and mitigation of language model non-factual hallucinations, 2024

    Lei Yu, Meng Cao, Jackie Chi Kit Cheung, and Yue Dong. Mechanistic understanding and mitigation of language model non-factual hallucinations, 2024. 51

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.