Pith. sign in

REVIEW 4 major objections 5 minor 59 references

A Mathematical Theory of Discursive Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A small chance of external detection can push a network of imperfect LLMs to a truth-dominant equilibrium.

desk verdict The single-network Markov analysis is correct but elementary; the cross-network 'truth-dominant' result rests on a Poisson-mean error in Lemma 2.4, so the headline claim does not follow from the model as written. read the letter →

arxiv 2507.06565 v5 pith:MRO4MQBZ submitted 2025-07-09 cs.CL cs.LG

Juan B. Gutiérrez This is my paper
classification cs.CLcs.LG MSC 60J2068T5094A17
keywords discursivenetworksinvalidationlargelanguagemodelshallucinationmulti-agentverificationMarkovchainequilibriumFlaws-of-Otherstruth-dominantstate
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

Large language models cannot be made individually error-free: the paper proves a mathematical floor on the probability that any finite-loss model emits an invalid statement. It therefore proposes a different route to reliability: connect imperfect models into a discursive network in which statements circulate among agents, and give each false claim a chance of being caught by another agent. The core result is a threshold condition: if the per-statement detection probability $d$ is large enough relative to the fabrication hazard $\lambda$, the network's long-run share of false statements falls below the single-network baseline, moving the system into a truth-dominant state. The paper also derives how many mutually checking agents are needed to hold the false share under any target tolerance, and presents the FOO algorithm as the software counterpart of the detection mechanism.

What carries the argument

The central object is a two-state Markov chain on the proportions of agents endorsing a true statement $r$ versus a false statement $f$, with hazards $p$ (true-to-false drift), $q$ (false-to-true repair), $\lambda$ (fabrication), and $d$ (external detection). The argument runs on the comparison of two fixed points, $\pi_f^{\mathrm{single}}=(p+\lambda)/(p+\lambda+q)$ and $\pi_f^{\mathrm{cross}}=\lambda/d$; the inequality $\lambda/d < (p+\lambda)/(p+\lambda+q)$ is the threshold that separates an invalidation-dominant regime from a truth-dominant one. In the multi-agent version the effective correction hazard is $q_{\mathrm{eff}}(n)=q+(n-1)d$, which yields the required number of agents $n_{\min}=\lceil 1 + ((p+\lambda)(1/\varepsilon-1)-q)/d \rceil$ for a target tolerance $\varepsilon$. The Flaws-of-Others (FOO) algorithm, a configurable loop in which agents critique one another while a harmonizer merges verdicts, is the operational realization of the detection hazard $d$.

What would settle it

One can measure $d$ directly: present a set of false statements of varied subtlety to LLM critics and record the fraction detected per statement. If the detection rate differs substantially within the corpus, the model's single scalar $d$ is not well-defined, and the predicted equilibrium $\pi_f = \lambda/d$ will not reproduce the observed falsehood share in a persistent network.

Watch

Extended reading notes

Core claim

The paper treats every LLM, and every human, as a node in a discursive network whose utterances are either true or false, and models the proportions of each belief over time as a Markov chain. In the single-network model, false statements are created by drift $p$ and fabrication $\lambda$ and repaired at rate $q$, giving a steady-state false share $\pi_f = (p+\lambda)/(p+\lambda+q)$. When a second network checks the first and detects each false statement with probability $d$, the equilibrium false share becomes $\pi_f = \lambda/d$. Lemma 2.5 states that if $\lambda/d < (p+\lambda)/(p+\lambda+q)$, coupling the network to an external detector strictly lowers the false share below the single-network baseline; the paper calls this regime truth-dominant. The same framework yields a closed-form bound on the number of cross-checking agents needed to keep the long-run error below any tolerance $\varepsilon$.

Load-bearing premise

The load-bearing premise is that every false statement has one fixed probability $d$ of being detected by each external checker, independent of the statement's content, the checker's competence, and the network structure; if that uniformity fails, the equilibrium $\pi_f = \lambda/d$ and the threshold condition no longer follow.

Editorial extensions

If this is right

  • If Lemma 2.5 holds, any real discursive network that realizes a detection probability $d$ satisfying $\lambda/d < (p+\lambda)/(p+\lambda+q)$ will converge to a lower falsehood share than the same network in isolation.
  • The agent-count formula gives a planning rule: for fixed $p$, $q$, $\lambda$, and $d$, a target error $\varepsilon$ is reachable with a finite number of mutually checking agents, and the number grows linearly in $(p+\lambda)/d$ rather than exponentially.
  • Because Lemma 2.1 forces a positive invalidation floor for every finite-loss model, the paper's approach implies that post-generation cross-checking, rather than further fine-tuning alone, is the only route to arbitrarily low error rates.
  • The FOO algorithm turns the abstract detection hazard into a concrete protocol that can wrap any set of LLM agents, so the threshold condition is testable in practice by measuring the empirical detection rate.

Reading between the lines

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

  • The paper treats $d$ as a single scalar; a natural extension the author leaves implicit is a content-dependent detection rate $d(s)$, in which case the closed form $\pi_f = \lambda/d$ would need to be replaced by an average over the statement distribution, and the threshold condition would be more demanding.
  • The truth-dominant result is about aggregate equilibrium proportions, not about which particular statements are true; extending the model to track individual statements would require per-statement hazards and would likely show that some falsehoods survive even in the truth-dominant regime.
  • The same fixed-point comparison could be tested on human fact-checking or peer review: if human reviewers detect false claims at a rate $d$, the model predicts a measurable drop in the acceptance rate of false claims under double-blind review.
  • The energy-cost discussion in the paper suggests an optimization the author does not work out: because each additional agent yields diminishing returns, one can choose the smallest $n$ achieving $\varepsilon$ and then allocate remaining effort to raising $d$, which may be cheaper than adding another full agent.
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 manuscript claims to develop a mathematical theory of 'discursive networks' in which humans and LLMs are nodes exchanging true and false statements. It introduces the term invalidation, proves a lower bound on residual invalidation probability (Lemma 2.1), and builds three models: a two-state Markov chain (Lemma 2.2), a single-network model with fabrication and self-repair (Lemma 2.3), and a cross-network detection model (Lemma 2.4). The central theoretical claim is that coupling a network to an external detector lowers the long-run false-statement share from (p+λ)/(p+λ+q) to λ/d, and Lemma 2.5 states a threshold for this 'truth-dominant' regime. The paper also derives an agent-count formula (Lemma 2.7), presents the FOO multi-agent critique algorithm with blockchain logging, and reports Monte-Carlo simulations for the claimed equilibria.

Significance. If the cross-network model were sound, the λ/d design rule and the n_min formula would be practically useful, and the FOO implementation would operationalize the detection hazard in software. The manuscript is candid about the lack of empirical validation and about the limits of its entropy comparison, and it ships open-source code and reproducible simulations. However, the central derivation contains a factor error that breaks the claimed equilibrium, and the consistency condition used for the cross-network model makes the main 'prediction' a restatement of an imposed constraint rather than a derived consequence. The single-network stochastic primitives are also inconsistent with the transition matrix analyzed in Lemma 2.3, so the baseline is not a well-defined stochastic process as written. Because the qualitative claim (external scrutiny reduces error) is defensible and correctable by re-deriving the fixed points, the appropriate outcome is a major revision rather than outright rejection.

major comments (4)
  1. [Section 2.5.3, Lemma 2.4, Eq. (14)] The proof sets E[X_k] = λ_k, but X_k(t) ~ Poisson(λ_k T_k(t)) has mean λ_k T_k(t). After normalizing, the equilibrium condition is λ_k π_T,k = d_jk π_F,k + q_k π_F,k − p_k π_T,k; using Eq. (13) this reduces to λ_k π_T,k = d_jk π_F,k, which together with π_F,k/π_T,k = p_k/q_k requires λ_k/d_jk = p_k/q_k and yields π_F,k = p_k/(p_k+q_k), independent of λ_k and d_jk. The paper's consistency condition (12), λ_k = d_jk p_k/(p_k+q_k), is not the condition the stochastic model requires; for the values in Table 4, λ/d ≈ 0.289 while p/q = 0.4, so the stated stochastic process has no interior fixed point at the claimed equilibrium. This error is load-bearing for Lemma 2.5 and for the quantitative claims in Section 3.4.
  2. [Section 2.5.3, Eq. (12), and Section 2.5.4, Lemma 2.5] The consistency condition (12) is imposed specifically to make the derived π*_F,k = λ_k/d_jk coincide with p_k/(p_k+q_k), so the comparison in Lemma 2.5 is not a derived prediction. Under (12), π_cross_f = p/(p+q), and the inequality λ/d < (p+λ)/(p+λ+q) becomes p/(p+q) < (p+λ)/(p+λ+q), which holds automatically for any λ > 0. Thus the claimed truth-dominance result restates the imposed constraint rather than demonstrating an effect of cross-network detection on the false-statement share.
  3. [Section 2.5.2, Eqs. (8)-(9)] The stochastic primitives do not conserve the population: with ΔT(t) = −Z(t)+W(t) and ΔF(t) = X(t)+Z(t)−W(t), one has ΔT+ΔF = X(t) > 0, so T(t)+F(t) = n cannot remain true and no interior equilibrium exists for λ > 0. If X(t) is intended to represent true statements becoming false (as the interpretation paragraph of Lemma 2.3 states), the update should include −X(t) in ΔT. The proof of Lemma 2.3 analyzes the transition matrix T_λ directly and never derives it from the stated Poisson/Binomial primitives; this mismatch affects the single-network baseline used in Section 3.3 and the interpretation of q_eff in Lemma 2.6.
  4. [Section 2.5.5, Lemmas 2.6-2.7] Lemma 2.7 substitutes q_eff(n) = q+(n−1)d into the single-network formula to obtain π_f^(n) = (p+λ)/(p+λ+q+(n−1)d). This formula is inconsistent with the fixed point of Lemma 2.4: under the stated update equations (10)-(11), detection appears only in ΔF, so the corrected equilibrium condition is λπ_T = (n−1)d π_F, and an interior fixed point exists only if λ/d = (n−1)p/q. To obtain the claimed formula one must instead treat every detection as a correction from false to true (adding Y to ΔT), which is not what Eqs. (10)-(11) specify. The agent-count bound n_min in Section 3.5 therefore does not follow from the cross-network model as written.
minor comments (5)
  1. [Table 4, d row] The caption says d is 'Picked so that λ/d < p/(p+q)', but 0.055/0.19 ≈ 0.289 > p/(p+q) ≈ 0.286; the condition that actually matches Lemma 2.5 is λ/d < (p+λ)/(p+λ+q) = 0.6.
  2. [Section 2.5.1, interpretation after Lemma 2.2] The text says 'the two flip probabilities satisfy p+q=1', but Lemma 2.2 only assumes p,q ∈ (0,1); this sentence is unnecessary and potentially confusing, since the fixed-point and convergence results hold without p+q = 1.
  3. [Sections 3.5 and 4.2] The text refers to 'Proposition 2.7' in both places, but the cited statement is Lemma 2.7; the numbering should be corrected.
  4. [Throughout] Several typos remain, including 'ennounce' (Section 2.1), 'comapred' (Section 4.1), 'hamful' (Section 4.1), and 'F AILS' (Appendix A); the manuscript would benefit from copyediting.
  5. [Section 2.2, Theorem 2.1] Theorem 2.1 is a direct application of the chain rule for joint entropy, and Remark 2.1 concedes that no universal verification advantage is proved; labeling this statement a theorem may overstate its content, and the empirical concentration evidence should be clearly separated from the identity.

Circularity Check

3 steps flagged · score 8.0 of 10

The cross-network truth-dominance result is forced by the consistency condition (12) and a hand-picked detection rate d, so the central 'prediction' reduces to its inputs by construction.

  1. self definitional [Section 2.5.3, Lemma 2.4 (Eq. 12); Section 2.5.4, Lemma 2.5]
    "Assume the per-actor falsehood-generation rate satisfies the consistency condition λ_k = d_jk p_k/(p_k+q_k), which guarantees that the expected proportions sum to one. Then the Markov process has the mean fixed point π*_F,k = λ_k/d_jk ... If λ/d < (p+λ)/(p+λ+q), then π_cross_f < π_single_f."

    Equation (12) fixes λ/d = p/(p+q). Substituting this into Lemma 2.5's antecedent gives p/(p+q) < (p+λ)/(p+λ+q), which is an algebraic identity for every positive λ. The truth-dominant state is therefore not an emergent property of the coupled stochastic dynamics; it is guaranteed by the normalization constraint the authors impose to define the equilibrium. Moreover, the claimed fixed point π_F = λ/d becomes exactly p/(p+q), the single-network false share with fabrication set to zero. The comparison in Lemma 2.5 is thus a restatement of the imposed constraint, not a derived property of peer review.

  2. fitted input called prediction [Section 3.2, Table 4; Sections 3.4 and 3.5]
    "d 0.19 cross-network repair Picked so that λ/d < p/(p+q), just inside the truth-dominant region; see Lemma 2.5. ... n_min = 9 ... at least nine mutually detecting agents are necessary to guarantee that fewer than one statement in twenty remains false at equilibrium under this calibration."

    The value d=0.19 is not estimated from data or predicted by the theory; the table states it was picked precisely to place λ/d inside the truth-dominant region. The later 'predictions' — the cross-network false share π_cross = 0.29 and the agent requirement n_min = 9 — are computed from this same hand-picked d and then presented as quantitative findings of the framework. These numbers are arithmetic consequences of the chosen input, not independent predictions, so the user-facing claim that 'a small chance of peer review shifts the system to a truth-dominant state' is an artifact of the parameter selection.

1 more flagged steps
  1. other [Section 2.5.3, stochastic primitives and proof of Lemma 2.4]
    "X_k(t)∼Poisson(λ_k T_k(t)) ... Using the distributional means E[X_k] = λ_k (Poisson distribution) ... λ_k = π_F,k d_jk + π_F,k q_k − π_T,k p_k."

    This is a derivation error that manufactures the circular fixed point. The model defines X_k with mean λ_k T_k(t), so after normalization its contribution is λ_k π_T,k, not the λ_k used in the proof. With the correct mean, the equilibrium equations become p π_T = q π_F and λ π_T = d π_F, giving π_F = p/(p+q) and λ/d = p/q, not π_F = λ/d. The consistency condition (12) is therefore not the condition required by the stated stochastic model, and the claimed fixed point does not follow from the dynamics. The 'prediction' π_F = λ/d is an artifact of substituting the wrong mean and then imposing a normalization constraint, rather than a consequence of the model.

full rationale

The paper's single-network results (Lemmas 2.2 and 2.3) are self-contained Markov-chain computations and are not circular. The circularity enters with the cross-network model, which carries the paper's central claim. Lemma 2.4 imposes Eq. (12) as a 'consistency condition' to force the equilibrium to be a probability vector, and this condition fixes λ/d = p/(p+q). Lemma 2.5 then 'proves' that λ/d lies below the single-network threshold, but under Eq. (12) this inequality is an algebraic identity, so the truth-dominant state is built into the assumption rather than derived from the detection dynamics. The problem is compounded by a factor error in the proof of Lemma 2.4: the model defines X_k(t) ~ Poisson(λ_k T_k(t)), but the proof uses E[X_k] = λ_k, so the claimed fixed point π_F = λ/d is not a fixed point of the stated process under the correct mean. Table 4 also states that d was chosen so that λ/d sits inside the truth-dominant region, and the subsequent quantitative outputs (π_cross ≈ 0.29 and n_min = 9) are consequences of that hand-picked input. These are genuine reduction-by-construction steps, not merely self-citation. The self-citation to reference [61] in the acknowledgements is not load-bearing and does not affect the score. Overall, the central 'peer review shifts the system to truth-dominance' prediction reduces to an imposed normalization constraint plus a chosen parameter value, meriting a circularity score of 8.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The model's outputs are driven by four hand-chosen parameters, one ad hoc consistency condition, and several domain assumptions about independence and constant rates. The paper does not measure any parameter from data, so all quantitative predictions reduce to consequences of these choices. The main invented entities are terminological or algorithmic, none with independent falsifiable evidence.

free parameters (4)
  • p = 0.02
    True-to-false slip probability; chosen an order of magnitude below q so that internal repair remains visible (Table 4).
  • q = 0.05
    Internal repair probability; matched to self-evaluation AUROC ≈ 0.55 from Zhang et al. (Section 3.2).
  • lambda = 0.055
    Fabrication hazard; solved from π_f ≈ 0.60 with p and q fixed, placing the single network in the invalidation-dominant regime (Table 4).
  • d = 0.19
    Cross-network detection probability; picked so that λ/d is near the truth-dominant boundary to illustrate the regime flip (Table 4).
assumptions (5)
  • domain assumption Absolute continuity P_θ ≪ Q in Lemma 2.1
    Requires the model to assign zero probability to sequences absent from training data, which is false for softmax-based LLMs; the application of the lemma to real models is therefore questionable (Section 2.1).
  • domain assumption Constant per-statement hazards p, q, λ, d across all actors, statements, and time
    All actors and statements share identical flip, fabrication, and detection rates; no heterogeneity, context, or learning is modeled (Sections 2.5.1-2.5.5).
  • domain assumption Independence and memorylessness of detection channels in Lemma 2.6
    Internal and external corrections are independent exponential clocks, enabling the effective hazard q + (n-1)d; this is a continuous-time approximation not derived from the discrete model.
  • domain assumption Verification concentration advantage
    The FOO algorithm's premise that critique outputs are higher-probability than generative outputs is supported only by cited empirical concentration ratios; the paper admits no general proof (Remark 2.1).
  • ad hoc to paper Consistency condition λ_k = d_jk p_k/(p_k + q_k)
    Introduced in Lemma 2.4 to make the claimed fixed point a normalized probability vector; not derived from the dynamical rules.
invented entities (3)
  • invalidation
    purpose: Replaces 'hallucination' to encompass factual, logical, structural, and ethical breaches in LLM output and human discourse.
    A terminological reclassification of known error types; no new falsifiable content beyond existing error taxonomies.
  • scientific epithesis
    purpose: Names the ethical transgression of claiming authorship with only superficial or no contribution to an artifact produced in a discursive network.
    Conceptual/ethical category with no empirical or falsifiable handle proposed in the paper.
  • Flaws-of-Others (FOO) algorithm
    purpose: Operationalizes cross-network peer review: agents critique each other's outputs and a harmonizer merges verdicts.
    A software protocol whose effectiveness is not demonstrated; no benchmarks or reproducible evaluations are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Mathematical Theory of Discursive Networks." pith.science (2026). https://pith.science/paper/MRO4MQBZ

@misc{pith2026250706565,
  author       = {Pith},
  title        = {Pith review of: A Mathematical Theory of Discursive Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRO4MQBZ}},
  note         = {Machine review of arXiv:2507.06565}
}
read the original abstract

Large language models (LLMs) turn writing into a live exchange between humans and software. We characterize this new medium as a discursive network that treats people and LLMs as equal nodes and tracks how their statements circulate. We define the generation of erroneous information as invalidation (any factual, logical, or structural breach) and show it follows four hazards: drift from truth, self-repair, fresh fabrication, and external detection. We develop a general mathematical model of discursive networks that shows that a network governed only by drift and self-repair stabilizes at a modest error rate. Giving each false claim even a small chance of peer review shifts the system to a truth-dominant state. We operationalize peer review with the open-source Flaws-of-Others (FOO) algorithm: a configurable loop in which any set of agents critique one another while a harmonizer merges their verdicts. We identify an ethical transgression, epithesis, that occurs when humans fail to engage in the discursive network. The takeaway is practical and cultural: reliability in this new medium comes not from perfecting single models but from connecting imperfect ones into networks that enforce mutual accountability.

Figures

Figures reproduced from arXiv: 2507.06565 by the authors.

Figure 1
Figure 1. Venn-style illustration of members of a broader set of invalid outputs produced by [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FOO consensus loop. An arbitrary set of agents [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. (a) Single-network emergent-invalidation dynamics corresponding to Lemma 2.3. Twenty independent Monte-Carlo runs of 100 steps are averaged. The blue curve shows the mean proportion πr(t) of actors endorsing the true statement r; the red curve shows the mean proportion πf (t) = 1 − πr(t) endorsing the false statement f. Shaded bands mark point-wise 95 % confidence intervals. Dashed horizontal lines denote the fixed … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Long-run false share vs. number of agents. The orange curve shows the analytic steady-state falsehood share πf (n) = p+λ p+λ+q+(n−1)d −1 for the calibrated hazards (p, q, λ, d) = (0.02, 0.05, 0.055, 0.19). Each dot marks an integer n; the dashed horizontal line is the…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 22 canonical work pages

  1. [1]

    Attention Is All You Need

    Ashish Vaswani et al. “Attention Is All You Need”. In:arXiv preprint arXiv:1706.03762 (2017).doi:10.48550/arXiv.1706.03762

    arXiv doi:10.48550/arxiv.1706.03762 2017

  2. [2]

    Language Models are Few-Shot Learners

    Tom B Brown et al. “Language Models are Few-Shot Learners”. In:arXiv preprint arXiv:2005.14165(2020).doi:10.48550/arXiv.2005.14165

    arXiv doi:10.48550/arxiv.2005.14165 2020

  3. [3]

    Survey of hallucination in natural language generation

    Ziwei Ji et al. “Survey of hallucination in natural language generation”. In:ACM Computing Surveys55.12 (2022), pp. 1–38.doi:10.1145/3571730

    doi:10.1145/3571730 2022

  4. [4]

    On faithfulness and factuality in abstractive summarization

    Joshua Maynez et al. “On faithfulness and factuality in abstractive summarization”. In:Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics. 2020, pp. 1906–1919.doi:10.18653/v1/2020.acl-main.173

    doi:10.18653/v1/2020.acl-main.173 2020

  5. [6]

    Ethical and Social Risks of Harm from Language Models

    Laura Weidinger et al. “Ethical and Social Risks of Harm from Language Models”. In: arXiv preprint arXiv:2112.04359(2021).doi:10.48550/arXiv.2112.04359

    arXiv doi:10.48550/arxiv.2112.04359 2021

  6. [7]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis et al. “Retrieval-augmented generation for knowledge-intensive nlp tasks”. In:Advances in Neural Information Processing Systems. Vol. 33. 2020, pp. 9459–9474. url:https://dl.acm.org/doi/abs/10.5555/3495724.3496517

    arXiv 2020

  7. [8]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang et al. “Self-consistency improves chain of thought reasoning in language models”. In:arXiv preprint arXiv:2203.11171(2023).doi:10.48550/arXiv.2203. 11171

    arXiv doi:10.48550/arxiv.2203.11171 2023

  8. [9]

    Large language models in medicine

    Arun James Thirunavukarasu et al. “Large language models in medicine”. In:Nature Medicine29.8 (2023), pp. 1930–1940.doi:10.1038/s41591-023-02448-8

    doi:10.1038/s41591-023-02448-8 2023

Show all 59 references
  1. [10]

    ChatGPT utility in healthcare education, research, and practice: sys- tematic review on the promising perspectives and valid concerns

    Malik Sallam. “ChatGPT utility in healthcare education, research, and practice: sys- tematic review on the promising perspectives and valid concerns”. In:Healthcare11.6 (2023), p. 887.doi:10.3390/healthcare11060887

    2023 doi

  2. [11]

    Universal and transferable adversarial attacks on aligned language models

    Andy Zou et al. “Universal and transferable adversarial attacks on aligned language models”. In:arXiv preprint arXiv:2307.15043(2023).doi:10.48550/arXiv.2307. 15043

    2023 doi arXiv

  3. [12]

    Jailbreaking black box large language models in twenty queries

    Patrick Chao et al. “Jailbreaking black box large language models in twenty queries”. In:arXiv preprint arXiv:2310.08419(2023).doi:10.48550/arXiv.2310.08419

    2023 doi arXiv

  4. [13]

    TruthfulQA: Measuring how mod- els mimic human falsehoods

    Stephanie Lin, Jacob Hilton, and Owain Evans. “TruthfulQA: Measuring how mod- els mimic human falsehoods”. In:arXiv preprint arXiv:2109.07958(2022).doi:10. 48550/arXiv.2109.07958. 35

    2022 doi arXiv

  5. [14]

    Truthful AI: Developing and governing AI that does not lie

    Owain Evans et al. “Truthful AI: Developing and governing AI that does not lie”. In: arXiv preprint arXiv:2110.06674(2021).doi:10.48550/arXiv.2110.06674

    2021 doi arXiv

  6. [15]

    Excavating AI: The Politics of Images in Machine-Learning Training Sets

    Kate Crawford. “Excavating AI: The Politics of Images in Machine-Learning Training Sets”. In:KDD ’21 Workshop on Machine Learning in Society. 2021.doi:10.1007/ s00146-021-01301-1

    2021

  7. [16]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

    Lei Huang et al. “A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions”. In:ACM Computing Surveys(2025). doi:10.1145/3703155

    2025 doi

  8. [17]

    Self-Contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation

    Niels M¨ undler et al. “Self-Contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation”. In:arXiv preprint arXiv:2305.15852(2023). doi:10.48550/arXiv.2305.15852

    2023 doi arXiv

  9. [18]

    Investigating the Robustness of Deductive Reasoning with Large Language Models

    Fabian Hoppe, Filip Ilievski, and Jan-Christoph Kalo. “Investigating the Robustness of Deductive Reasoning with Large Language Models”. In:arXiv preprint arXiv:2502.04352 (2025).doi:10.48550/arXiv.2502.04352

    2025 doi arXiv

  10. [19]

    RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models

    Samuel Gehman et al. “RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models”. In:Findings of the Association for Computational Linguistics: EMNLP 2020(2020).doi:10.48550/arXiv.2009.11462

    2020 doi arXiv

  11. [20]

    JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models

    Saibo Geng et al. “JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models”. In:arXiv preprint arXiv:2501.10868(2025).doi:10.48550/ arXiv.2501.10868

    2025 doi arXiv

  12. [21]

    Stanford, CA: Stanford University Press, 1957

    Leon Festinger.A Theory of Cognitive Dissonance. Stanford, CA: Stanford University Press, 1957

    1957

  13. [22]

    Cambridge, UK: Polity Press, 2001.doi:10.1086/343211

    Stanley Cohen.States of Denial: Knowing about Atrocities and Suffering. Cambridge, UK: Polity Press, 2001.doi:10.1086/343211

    2001 doi

  14. [23]

    Woddstock, NY: Overlook Books, 1974.url:https://lccn.loc.gov/73407323

    Erving Goffman.The Presentation of Self in Everyday Life. Woddstock, NY: Overlook Books, 1974.url:https://lccn.loc.gov/73407323

    1974

  15. [24]

    Living with Television: The Violence Profile

    George Gerbner and Larry Gross. “Living with Television: The Violence Profile”. In: Journal of Communication26.2 (1976), pp. 172–199.doi:10.1111/j.1460- 2466. 1976.tb01397.x

    1976 doi

  16. [25]

    New York: Mc- Graw–Hill, 1964

    Marshall McLuhan.Understanding Media: The Extensions of Man. New York: Mc- Graw–Hill, 1964

    1964

  17. [26]

    Herman and Noam Chomsky.Manufacturing Consent: The Political Econ- omy of the Mass Media

    Edward S. Herman and Noam Chomsky.Manufacturing Consent: The Political Econ- omy of the Mass Media. New York: Pantheon Books, 1988

    1988

  18. [27]

    New York: Harcourt Brace, 1951

    Hannah Arendt.The Origins of Totalitarianism. New York: Harcourt Brace, 1951

    1951

  19. [28]

    New York: Knopf, 1965

    Jacques Ellul.Propaganda: The Formation of Men ’s Attitudes. New York: Knopf, 1965

    1965

  20. [29]

    Judgment under Uncertainty: Heuristics and Biases

    Amos Tversky and Daniel Kahneman. “Judgment under Uncertainty: Heuristics and Biases”. In:Science185.4157 (1974), pp. 1124–1131.doi:10.1126/science.185. 4157.1124

    1974

  21. [30]

    The Case for Motivated Reasoning

    Ziva Kunda. “The Case for Motivated Reasoning”. In:Psychological Bulletin108.3 (1990), pp. 480–498.doi:10.1037/0033-2909.108.3.480. 36

    1990 doi

  22. [31]

    Discourse Network Analysis: Policy Debates as Dynamic Networks

    Philip Leifeld. “Discourse Network Analysis: Policy Debates as Dynamic Networks”. In: Policy Studies Journal42.3 (2014), pp. 465–487.doi:10.1093/oxfordhb/9780190228217. 013.25

    2014

  23. [32]

    Political Discourse Networks and the Conflict over Software Patents in Europe

    Philip Leifeld and Sebastian Haunss. “Political Discourse Networks and the Conflict over Software Patents in Europe”. In:European Journal of Political Research51.3 (2012), pp. 382–409.doi:10.1111/j.1475-6765.2011.02003.x

    2012

  24. [33]

    Kittler.Discourse Networks 1800/1900

    Friedrich A. Kittler.Discourse Networks 1800/1900. Stanford University Press, 1990

    1900

  25. [34]

    Automatic Detection of Generated Text Is Easiest When Humans Are Fooled

    Dani Ippolito et al. “Automatic Detection of Generated Text Is Easiest When Humans Are Fooled”. In:Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics. 2020, pp. 1808–1822.doi:10.18653/v1/2020.acl-main.164

    2020 doi

  26. [35]

    GLTR: Statistical Detection and Visualization of Generated Text

    Sebastian Gehrmann, Hendrik Strobelt, and Alexander M. Rush. “GLTR: Statistical Detection and Visualization of Generated Text”. In:Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28 - August 2, 2019, Volume 3...

    2019 doi

  27. [36]

    Political Discourse on Social Media: Echo Chambers, Gate- keepers, and the Price of Bipartisanship

    Kiran Garimella et al. “Political Discourse on Social Media: Echo Chambers, Gate- keepers, and the Price of Bipartisanship”. In:Proceedings of the Web Conference 2018. 2018, pp. 913–922.doi:10.1145/3178876.3186139

    2018

  28. [37]

    Are AI-Generated Text Detectors Robust to Adversarial Per- turbations?

    Guanhua Huang et al. “Are AI-Generated Text Detectors Robust to Adversarial Per- turbations?” In:arXiv preprintarXiv:2406.01179 (2024).doi:10.48550/arXiv.2406. 01179

    2024 doi arXiv

  29. [38]

    A Practical Examination of AI-Generated Text Detectors for Large Language Models

    Brian Tufts, Xuandong Zhao, and Lei Li. “A Practical Examination of AI-Generated Text Detectors for Large Language Models”. In:Proceedings of the 2024 Annual Meet- ing of the Association for Computational Linguistics. 2024.doi:10.48550/arXiv. 2412.05139

    2024 doi arXiv

  30. [39]

    Can AI–Generated Text Be Reliably Detected?

    Vinu Sankar Sadasivan et al. “Can AI–Generated Text Be Reliably Detected?” In: Transactions on Machine Learning Research(2025).doi:10 . 48550 / arXiv . 2303 . 11156

    2025

  31. [41]

    The Curious Case of Neural Text Degeneration

    Ari Holtzman et al. “The Curious Case of Neural Text Degeneration”. In:International Conference on Learning Representations. 2020.url:https : / / openreview . net / forum?id=rygGQyrFvH. 37

    2020

  32. [42]

    Exploiting Cloze Questions for Few Shot Text Classification and Natural Language Inference

    Timo Schick and Hinrich Sch¨ utze. “Exploiting Cloze Questions for Few Shot Text Classification and Natural Language Inference”. In:Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics. Association for Computational Lingui...

    2021 doi

  33. [43]

    Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?

    Sewon Min et al. “Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?” In:Proceedings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing. Association for Computational Linguistics, 2022, pp. 11048– 11064.doi:10.18653/v1/2022.em...

    2022 doi

  34. [44]

    Language Models (Mostly) Know What They Know

    Saurav Kadavath et al. “Language Models (Mostly) Know What They Know”. In: arXiv preprint arXiv:2207.05221(2022).doi:10.48550/arXiv.2207.05221.url: https://arxiv.org/abs/2207.05221

    2022 doi arXiv

  35. [45]

    On the Opportunities and Risks of Foundation Models

    Rishi Bommasani and et al. “On the Opportunities and Risks of Foundation Models”. In:arXiv:2108.07258(2022).doi:10.48550/arXiv.2108.07258

    2022 doi arXiv

  36. [46]

    Augmented language models: a survey

    Gr´ egoire Mialon et al. “Augmented language models: a survey”. In:arXiv preprint arXiv:2302.07842(2023).doi:10.48550/arXiv.2302.07842

    2023 doi arXiv

  37. [47]

    Chain of Thought Prompting Elicits Reasoning in Large Lan- guage Models

    Jason Wei and et al. “Chain of Thought Prompting Elicits Reasoning in Large Lan- guage Models”. In:NeurIPS(2022)

    2022

  38. [48]

    Training language models to follow instructions with human feedback

    Long Ouyang et al. “Training language models to follow instructions with human feedback”. In:Advances in neural information processing systems35 (2022), pp. 27730– 27744.url:https://proceedings.neurips.cc/paper_files/paper/2022/file/ b1efde53be364a73914f58805a001731-Paper-Conf...

    2022

  39. [49]

    Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm

    L. Reynolds and T. McDonell. “Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm”. In:arXiv preprint arXiv:2102.07350(2021).doi: 10.48550/arXiv.2102.07350

    2021 doi arXiv

  40. [50]

    ProxyLLM: An LLM-Driven Framework for Cus- tomer Support Through Text-Style Transfer

    Sehyeong Jo and Jungwon Seo. “ProxyLLM: An LLM-Driven Framework for Cus- tomer Support Through Text-Style Transfer”. In:Proceedings of The Web Conference (WWW) 2025. arXiv:2412.09916. 2025.doi:10.48550/arXiv.2412.09916

    2025 doi arXiv

  41. [51]

    Generative Agents: Interactive Simulacra of Human Behavior

    Joon Sung Park et al. “Generative Agents: Interactive Simulacra of Human Behavior”. In:Proceedings of CHI ’23. ACM, 2023, pp. 1–18.doi:10.1145/3586183.3606763

    2023

  42. [52]

    The Typing Cure: Experiences with Large Language Model Chat- bots for Mental Health Support

    Inhwa Song et al. “The Typing Cure: Experiences with Large Language Model Chat- bots for Mental Health Support”. In:arXiv preprintarXiv:2401.14362 (2024).doi: 10.48550/arXiv.2401.14362

    2024 doi arXiv

  43. [53]

    Stigmergic epistemology, stigmergic cognition

    Leslie Marsh and Christian Onof. “Stigmergic epistemology, stigmergic cognition”. In: Cognitive Systems Research9.1 (2008). Perspectives on Social Cognition, pp. 136–149. issn: 1389-0417.doi:https://doi.org/10.1016/j.cogsys.2007.06.009.url: https://www.sciencedirect.com/scienc...

    2008 doi

  44. [54]

    Towards Mitigating Hallucination in Large Language Models via Self- Reflection

    Ziwei Ji et al. “Towards Mitigating Hallucination in Large Language Models via Self- Reflection”. In:Findings of the Association for Computational Linguistics: EMNLP 2023. 2023, pp. 1827–1843.doi:10.18653/v1/2023.findings-emnlp.123. 38

    2023 doi

  45. [55]

    Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation

    Xiaoying Zhang et al. “Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation”. In:arXiv preprint arXiv:2402.09267(2024).doi:10 . 48550/arXiv.2402.09267

    2024 doi arXiv

  46. [56]

    A systematic review of research on the meaning, ethics and practices of authorship across scholarly disciplines

    Ana Maruˇ si´ c, Lana Boˇ snjak, and Ana Jeronˇ ci´ c. “A systematic review of research on the meaning, ethics and practices of authorship across scholarly disciplines”. In:Plos one6.9 (2011), e23477.doi:journal.pone.0023477.url:https://doi.org/10. 1371/journal.pone.0023477

    2011

  47. [57]

    Energy and Policy Consid- erations for Deep Learning in NLP

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. “Energy and Policy Consid- erations for Deep Learning in NLP”. In:Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019, pp. 3645–3650.doi:10.18653/v1/ P19-1355.url:https://doi.org/10.18...

    2019 doi

  48. [58]

    On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

    Emily M. Bender et al. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In:Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. 2021, pp. 610–623.doi:10.1145/3442188.3445922.url:https: //doi.org/10.1145/3442188.3445922

    2021

  49. [59]

    Membership Inference Attacks Against Machine Learning Models

    Reza Shokri et al. “Membership Inference Attacks Against Machine Learning Models”. In:Proceedings of the 2017 IEEE Symposium on Security and Privacy. 2017, pp. 3–18. doi:10.1109/SP.2017.41.url:https://doi.org/10.1109/SP.2017.41

    2017 doi

  50. [60]

    Universal Adversarial Triggers for Attacking and Analyzing NLP

    Eric Wallace et al. “Universal Adversarial Triggers for Attacking and Analyzing NLP”. In:Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. 2019, pp. 2153–2162.doi:10.18653/v1/D19-1221.url:https://doi. org/10.18653/v1/D19-1221

    2019 doi

  51. [61]

    A Mathematical Theory of Discursive Networks

    Juan B. Guti´ errez. “A Mathematical Theory of Discursive Networks”. In:arXiv preprint arXiv:2507.06565(2025).doi:10.48550/arXiv.2507.06565.url:https://doi. org/10.48550/arXiv.2507.06565. 39 A Blockchain Implementation Details . Implementation note.We implement the loop inPyth...

    2025 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.