Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

HD-NDEs: Neural Differential Equations for Hallucination Detection in LLMs

T0 review · 5 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 Hallucination signals in LLMs are spread across the full token trajectory, and modeling that trajectory with neural differential equations raises detection AUC-ROC by over 14 percent on subtle true/false statements.

desk verdict A first application of Neural DEs to hallucination detection with a headline 14% AUC gain, but the gain is likely attributable to using all token states rather than DE dynamics; needs a simple pooling baseline and fuller specification. read the letter →

arxiv 2506.00088 v1 pith:Y2T4LT2E submitted 2025-05-30 cs.CL cs.AIcs.LG

Qing Li , Jiahui Geng , Zongxiong Chen , Derui Zhu , Yuxia Wang , Congbo Ma , Chenyang Lyu , Fakhri Karray This is my paper
classification cs.CLcs.AIcs.LG
keywords hallucinationdetectionneuraldifferentialequationsODEsCDEsSDEsLLMinternalstatesPCAprojectionAUC-ROC
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

A statement's truthfulness in an LLM is not readable from the last token alone; the signal lives in how hidden states evolve across all tokens. The paper proposes HD-NDEs, which treat per-token hidden states (after PCA compression) as a continuous trajectory and model its evolution with neural ODEs, CDEs, or SDEs, then classify the resulting latent path as factual or hallucinated. On a True-False benchmark built from near-identical factual and non-factual statements, the method reports an average AUC-ROC gain over 14 percent relative to final-token classifiers SAPLMA, MIND, and Probe@Exact across six LLMs. On four question-answering datasets, it is comparable to those baselines. This matters for deployment because it offers a route to catch false statements even when the model's final words look confident and correct.

What carries the argument

The central object is the latent trajectory $z(t)$ defined by a neural differential equation over token-index time. Each token's last-layer hidden state is projected by PCA to dimension $K=1024$ to form the sequence $y=(y_0,\dots,y_n)$; an encoder $h$ maps the start into $z(0)$, and the DE solver advances $z$ through a neural-network vector field $f$ (ODEs), a controlled path $Y(t)$ built from cubic splines (CDEs), or drift plus diffusion with Brownian motion (SDEs). The trajectory is aggregated by a function $k$ into $z^*$, and a linear layer with a sigmoid outputs the hallucination probability. Gradients come from adjoint methods, and the solvers are Euler or RK4 for ODEs and CDEs and Euler-Maruyama for SDEs. The machinery's job is to force the model to use the whole sequence's dynamics rather than a single endpoint.

What would settle it

Take a set of statements that are identical except for a single swapped word placed in the first third of the sentence, train HD-NDEs and a final-token classifier of similar capacity, and vary the error position; if the HD-NDE advantage over the final-token classifier disappears when the factual error sits early, the claim that full-trajectory dynamics cause the gain is contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, HD-NDEs establishes that the internal state sequence of an LLM can be treated as a continuous dynamical system, and that the trajectory itself carries truthfulness information that the final hidden state does not. The authors show that for pairs of statements differing by only a few words, the last-token representations of true and false statements overlap, while their trajectories diverge in the middle; Neural DEs trained on this trajectory recover the difference. The strongest demonstration is on the True-False dataset, where the best HD-NDE variants beat the three classification baselines by more than 14 percent AUC-ROC on average, with the largest single gap being CDEs outperforming SAPLMA by 24.3 percent on the Invention subset with Gemma-2-9B. On four QA datasets, the gains shrink and the method performs within 6 percent of the baselines, which the paper attributes to those datasets' answers being easy to separate from the final token alone.

Load-bearing premise

The method assumes that per-token hidden states, after PCA projection, form a smooth continuous trajectory in which token order behaves like time, so a learnable differential equation can represent them; if token embeddings do not evolve smoothly, the DE step adds nothing over a classifier on the final token.

Editorial extensions

If this is right

  • If correct, hallucination detection can be improved for open-source models without external retrieval or repeated sampling, simply by training on internal states.
  • Sequence-level classifiers that ignore temporal structure are leaving signal on the table; hidden-state dynamics are a usable feature family for truthfulness.
  • On tasks where true and false outputs have similar surface form, final-token probes will be near chance while trajectory models can separate them.
  • Middle layers can carry more hallucination signal than the final layer, so layer choice interacts with detection method.
  • The method generalizes out-of-domain with smaller AUC drops than the baselines in the paper's cross-domain test.

Reading between the lines

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

  • One consequence the paper leaves implicit is that the trajectory shape itself, including where and by how much true and false answers diverge, could be used to localize the hallucination within the output rather than only classifying the whole statement.
  • The PCA-plus-DE pipeline can be read as a learned smoothing of token embeddings; a direct comparison against a non-differential sequence model, such as an LSTM or transformer over the same PCA states, would test whether the DE formulation or simply sequence modeling drives the gain.
  • A testable extension is to inject factual errors at controlled token positions in otherwise identical statements and measure detection sensitivity as a function of error position; the paper's motivation predicts that sensitivity should stay high even for errors in the first third of the sequence.
  • The authors' comparative result on QA datasets suggests the method targets subtle-factuality cases more than general QA, so benchmark suites that mix both regimes would give a fuller picture.
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

5 major / 5 minor

Summary. The paper proposes HD-NDEs, a supervised method that models the sequence of per-token hidden states of an LLM as a continuous trajectory in a PCA-reduced latent space and applies neural ordinary, controlled, or stochastic differential equations (Neural ODEs, CDEs, SDEs) to obtain a representation z* that is fed to a linear classifier for hallucination detection. The method is evaluated on four True-False subsets and four question-answering datasets, across six open-source LLMs, using AUC-ROC as the metric. The paper reports that the proposed methods outperform SAPLMA, MIND, and Probe@Exact by over 14% on the True-False datasets, with smaller gains on question-answering datasets, and includes ablations on the number of hidden layers, PCA dimensions, middle layers, and an out-of-domain transfer setting.

Significance. If the reported results are substantiated, HD-NDEs would be a novel and practically relevant application of neural differential equations to hallucination detection, especially for the challenging case where non-factual content appears in the early or middle portion of a generated sequence. The paper deserves credit for evaluating on five datasets, six LLMs, and three Neural DE variants, and for identifying a plausible failure mode of final-token classifiers. However, the current evidence does not yet isolate the contribution of Neural DE dynamics: the method is not fully specified, a crucial non-ODE full-token baseline is missing, and the PCA/train-test procedure is not described. These issues are load-bearing for the central claim and should be addressed before the results can be relied upon.

major comments (5)
  1. [Section 3.2 / Figure 2] The function k(θk) that maps the latent states z=(z0,...,zn) to z* is never defined. The manuscript only says 'We derive z* from the latent states z using the function k(θk)' and then states that the classifier is a linear layer followed by a sigmoid. This is a central component of the method: if k is a learned pooling or attention over all z_t, then the model is effectively a sequence encoder and the Neural DE solver is only one module; if k selects the terminal state, the ODE and SDE variants may not use all token observations in the way the paper claims. Please define k explicitly, including its input, output, parameterization, and how its parameters θk are trained.
  2. [Section 3.1, Eqs. (1)-(3)] The initial condition z(0)=h(y;θ_h) has a type mismatch. The text defines y=y_{0:n}=(y0,...,yn) as a sequence, but h is described as a feedforward network from R^{dy} to R^{dz}. If h acts only on y0, then the Neural ODE trajectory in Eq. (1) is independent of y1,...,yn, which contradicts the claim that the method captures the full dynamics of the token sequence. If h acts on the entire sequence, its input dimension is variable and this needs to be specified. Please clarify what h actually consumes and how variable-length sequences are handled.
  3. [Section 4.5 / Appendix D] The train/test split and the PCA fitting procedure for HD-NDEs are not reported. Appendix D describes a random three-quarters/one-quarter split for SAPLMA, but no equivalent statement appears for the proposed method. It is also not stated whether PCA is fit on the training split only, on the full dataset before splitting, or separately for each dataset and LLM. If PCA is fit on all samples before classification, the AUC values in Tables 1 and 2 could be inflated by information leakage. Please specify the split, the PCA fitting procedure, and whether the reported numbers are averages over multiple random seeds.
  4. [Section 5.1 / Table 1] The central claim of an over-14% improvement is made against SAPLMA, MIND, and Probe@Exact, which use only final-token or key-token states. The paper does not include a non-ODE baseline that uses the same full trajectory of PCA-reduced token embeddings, such as mean or max pooling over all tokens followed by the same linear classifier, or a BiLSTM/Transformer over the token-indexed sequence. Without such an ablation, the reported gains could be due entirely to access to all token states rather than to Neural DE dynamics. This concern is reinforced by Section 5.2, which shows that the True-False difficulty is concentrated in overlapping, intermixed points across the sequence. Please add this baseline and discuss the results.
  5. [Section 3.3.2, Eqs. (8)-(9)] The adjoint sensitivity equations are incorrect as written. Eq. (8) should use a(t), not α(t), and should read da/dt = -a(t)^T ∂f/∂z. Eq. (9) should be dL/dθ = -∫ a(t)^T ∂f/∂θ dt, but the printed equation has ∂f/∂z and omits dt. Since the paper states that the parameters are updated using these adjoint methods, the current derivation is not reproducible. Please correct the equations or replace them with a precise reference to the standard adjoint method for Neural ODEs and its CDE/SDE analogues.
minor comments (5)
  1. [Eq. (4)] The Euler update in Eq. (4) is missing Δt: it should be z_{t+1}=z_t + Δt f(t,z_t). The text later says Δt can be 1, but the equation as written is dimensionally inconsistent.
  2. [Section 4.5 / Figure 2] The time points t0,...,tn and the time step Δt used in the experiments are never stated. Figure 2 mentions 'corresponding time points' and Section 3.3.1 says Δt 'could be simply 1', but the actual values used for the data should be specified for reproducibility.
  3. [Section 4.1 / Appendix B] The description of the QA datasets says 'we generate 1,000 samples in each of the four aforementioned datasets' but does not state whether these are balanced between truthful and hallucinated answers. Please clarify the class balance and how the 1,000 samples were drawn from the original datasets.
  4. [References] The HaluEval reference is duplicated as Li et al. 2023a and Li et al. 2023b, which are the same paper. In addition, Appendix C.1 refers to 'LVLMs' in the Unigram consistency method, but the experiments are on LLMs; this appears to be a typo.
  5. [Table 1 / Section 5.1] The statement 'over 14% improvement' should be accompanied by the exact baseline average used for the comparison, and the tables would benefit from error bars or significance testing; currently the reported numbers are point estimates without variance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported AUC gains come from a supervised train/test evaluation, not from a fitted parameter or self-citation chain.

full rationale

The paper's central claim is empirical: HD-NDEs outperforms SAPLMA, MIND, and Probe@Exact on held-out AUC-ROC test splits. The pipeline (PCA projection, Neural DE solver, latent mapping k, and classifier c) is trained with binary cross-entropy on labeled statements, and the reported numbers are measured on data not used for fitting. There is no fitted constant that is later renamed as a prediction, no equation that defines the target in terms of the model output, and no uniqueness theorem or ansatz smuggled in via a self-citation. The self-citations present in the manuscript (e.g., Geng et al. 2023; Li et al. 2024; Zhu et al. 2024) appear in related-work or background contexts and are not load-bearing for the method's predictive claim. The absence of a non-ODE sequence-aggregation baseline is a genuine experimental-completeness concern, but it is not circularity under the definition used here: the ODE modeling is not equivalent by construction to the evaluation metric, and the claim could in principle fail. Likewise, the use of GPT-4o to construct the True-False variants is a data-construction validity concern, not a case of the paper's derivation reducing to its own inputs. No specific circular step can be quoted, so the score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a large set of trained neural network weights (the DE models) plus hand-chosen hyperparameters such as PCA dimension, network depth, learning rate, and epoch count. The most non-standard assumption is that LLM token embeddings form a continuous dynamic system. No new physical or conceptual entities are introduced.

free parameters (5)
  • PCA dimension K = 1024
    Chosen to reduce computational cost; ablation (Figure 5) shows performance improves with more components, so the choice is a hand-selected trade-off.
  • Number of layers in f(·;θ_f) = 8 (CDEs), 10 (ODEs/SDEs)
    Selected based on ablations (Figure 4); depth materially changes AUC, making it a tuned hyperparameter.
  • Learning rate = 0.001
    Adam optimizer with fixed learning rate; no sensitivity analysis reported.
  • Training epochs = 50
    Fixed number of epochs without early stopping or validation-based selection; may overfit.
  • Time step Δt = 1 (per token)
    Assumes token positions are equally spaced in continuous time; alternative spacings are not explored.
assumptions (4)
  • standard math Existence and uniqueness of solutions to the ODE/SDE used in Equations 1-3.
    The method treats the latent dynamics as a well-posed initial value problem, requiring standard theorems for ODEs/SDEs.
  • domain assumption Token-level hidden states from the last layer of an LLM follow a continuous trajectory that can be approximated by a differential equation.
    This is the key modeling premise of the paper, stated implicitly in Section 3.1 and Figure 2.
  • domain assumption PCA projection to 1024 dimensions preserves the information needed to distinguish factual from non-factual statements.
    The paper relies on PCA without formally verifying that the discarded dimensions are irrelevant for the classification task.
  • domain assumption The GPT-4o-generated statements marked as non-factual in the True-False dataset are indeed non-factual.
    Section 4.1 describes using GPT-4o to produce factually opposite statements; incorrect labeling would directly affect the labels used in training and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HD-NDEs: Neural Differential Equations for Hallucination Detection in LLMs." pith.science (2026). https://pith.science/paper/Y2T4LT2E

@misc{pith2026250600088,
  author       = {Pith},
  title        = {Pith review of: HD-NDEs: Neural Differential Equations for Hallucination Detection in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y2T4LT2E}},
  note         = {Machine review of arXiv:2506.00088}
}
read the original abstract

In recent years, large language models (LLMs) have made remarkable advancements, yet hallucination, where models produce inaccurate or non-factual statements, remains a significant challenge for real-world deployment. Although current classification-based methods, such as SAPLMA, are highly efficient in mitigating hallucinations, they struggle when non-factual information arises in the early or mid-sequence of outputs, reducing their reliability. To address these issues, we propose Hallucination Detection-Neural Differential Equations (HD-NDEs), a novel method that systematically assesses the truthfulness of statements by capturing the full dynamics of LLMs within their latent space. Our approaches apply neural differential equations (Neural DEs) to model the dynamic system in the latent space of LLMs. Then, the sequence in the latent space is mapped to the classification space for truth assessment. The extensive experiments across five datasets and six widely used LLMs demonstrate the effectiveness of HD-NDEs, especially, achieving over 14% improvement in AUC-ROC on the True-False dataset compared to state-of-the-art techniques.

Figures

Figures reproduced from arXiv: 2506.00088 by the authors.

Figure 1
Figure 1. 1D PCA projection of hidden layer embed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Computation graph of HD-NDEs detecting hallucination via Neural CDEs. The statement is processed by [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. 2D PCA projection of the last hidden layer’s [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The impact of the number of hidden layers on [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: 2D PCA projection of the last hidden layer’s embedding for the final token on [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs

    cs.LG 2026-01 conditional novelty 6.0 of 10

    Spurious RLVR makes Qwen2.5-Math retrieve memorized answers via a layer 18-20 anchor and layer 21+ adapters, a shortcut that can be steered by scaling specific MLP keys.

  2. Decoding Memories: An Efficient Pipeline for Self-Consistency Hallucination Detection

    cs.CL 2025-08 conditional novelty 6.0 of 10

    A decoding pipeline reuses cached tokens and anneals sampling temperature to accelerate self-consistency hallucination detection by up to 3x without meaningful AUROC loss.

Reference graph

Works this paper leans on

57 extracted references · 20 canonical work pages · cited by 2 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Herv \'e Abdi and Lynne J Williams. 2010. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433--459

    2010

  4. [4]

    Fatemeh Azadi, Heshaam Faili, and Mohammad Javad Dousti. 2023. https://doi.org/10.18653/v1/2023.findings-acl.782 PMI -align: Word alignment with point-wise mutual information without requiring parallel training data . In Findings of the Association for Computational Linguistics: ACL 2023, pages 12366--12377, Toronto, Canada. Association for Computational ...

    work page doi:10.18653/v1/2023.findings-acl.782 2023

  5. [5]

    Amos Azaria and Tom Mitchell. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.68 The internal state of an LLM knows when it ' s lying . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967--976, Singapore. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-emnlp.68 2023

  6. [6]

    Aaron Baier-Reinio and Hans De Sterck. 2020. https://arxiv.org/abs/2010.11358 N-ode transformer: A depth-adaptive variant of the transformer using neural ordinary differential equations . arXiv preprint arXiv:2010.11358

    arXiv 2020

  7. [7]

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2023. https://openreview.net/pdf?id=ETKGuby0hcs Discovering latent knowledge in language models without supervision . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net

    work page 2023

  8. [8]

    Bo Chang, Lili Meng, Eldad Haber, Frederick Tung, and David Begert. 2018. https://openreview.net/forum?id=SyJS-OgR- Multi-level residual networks from dynamical systems view . In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . OpenReview.net

    work page 2018

Show all 57 references
  1. [9]

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024. https://openreview.net/forum?id=Zj12nzlQbz INSIDE : LLM s' internal states retain the power of hallucination detection . In The Twelfth International Conference on Learning Representations

    2024

  2. [10]

    Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. 2018. https://proceedings.neurips.cc/paper/2018/hash/69386f6bb1dfed68692a24c8686939b9-Abstract.html Neural ordinary differential equations . In Advances in Neural Information Processing Systems 31: Annual Con...

    2018

  3. [11]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. https://lmsys.org/blog/2023-03-30-vicuna/ Vicuna: An open-source chatbot impressing gpt-4 with 90\

    2023

  4. [12]

    Jeongwhan Choi, Hwangyong Choi, Jeehyun Hwang, and Noseong Park. 2022. https://ojs.aaai.org/index.php/AAAI/article/view/20587 Graph neural controlled differential equations for traffic forecasting . In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-...

    2022

  5. [13]

    Jinhao Duan, Hao Cheng, Shiqi Wang, Chenan Wang, Alex Zavalny, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. 2023. https://arxiv.org/abs/2307.01379 Shifting attention to relevance: Towards the uncertainty estimation of large language models . ArXiv preprint, abs/2307.01379

    2023 arXiv

  6. [14]

    Subhabrata Dutta, Tanya Gautam, Soumen Chakrabarti, and Tanmoy Chakraborty. 2021. https://proceedings.neurips.cc/paper/2021/hash/2bd388f731f26312bfc0fe30da009595-Abstract.html Redesigning the transformer architecture with insights from multi-particle dynamical systems . In Adv...

    2021

  7. [15]

    Leonhard Euler. 1845. Institutionum calculi integralis, volume 4. impensis Academiae imperialis scientiarum

  8. [16]

    Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2023. https://arxiv.org/abs/2311.08298 A survey of language model confidence estimation and calibration . ArXiv preprint, abs/2311.08298

    2023 arXiv

  9. [17]

    Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. https://openreview.net/pdf?id=sE7-XhLxHA Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing . In The Eleventh International Conference on Learning Representations, ICLR 2...

    2023

  10. [18]

    Yuheng Huang, Jiayang Song, Zhijie Wang, Huaming Chen, and Lei Ma. 2023. https://arxiv.org/abs/2307.10236 Look before you leap: An exploratory study of uncertainty measurement for large language models . ArXiv preprint, abs/2307.10236

    2023 arXiv

  11. [19]

    Morrill James, Kidger Patrick, Yang Lingyi, and Lyons Terry. 2021. https://arxiv.org/abs/2106.11028 Neural controlled differential equations for online prediction tasks . ArXiv preprint, abs/2106.11028

    2021 arXiv

  12. [20]

    Sheo Yon Jhin, Heejoo Shin, Sujie Kim, Seoyoung Hong, Minju Jo, Solhee Park, Noseong Park, Seungbeom Lee, Hwiyoung Maeng, and Seungmin Jeon. 2024. https://arxiv.org/pdf/2109.01876 Attentive neural controlled differential equations for time-series classification and forecasting...

    2024 arXiv

  13. [21]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. https://arxiv.org/pdf/2310.06825 Mistral 7b . arXiv preprint arXiv:2310.06825

    2023 arXiv

  14. [22]

    Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. https://doi.org/10.18653/v1/P17-1147 T rivia QA : A large scale distantly supervised challenge dataset for reading comprehension . In Proceedings of the 55th Annual Meeting of the Association for Computational...

    2017 doi

  15. [23]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. https://arxiv.org/abs/2207.05221 Language models (mostly) know what they know . arXiv preprint arXiv:2207.05221

    2022 arXiv

  16. [24]

    Patrick Kidger. 2022. https://arxiv.org/abs/2202.02435 On neural differential equations . ArXiv preprint, abs/2202.02435

    2022 arXiv

  17. [25]

    Patrick Kidger, James Foster, Xuechen Li, and Terry J. Lyons. 2021 a . https://proceedings.neurips.cc/paper/2021/hash/9ba196c7a6e89eafd0954de80fc1b224-Abstract.html Efficient and accurate gradients for neural sdes . In Advances in Neural Information Processing Systems 34: Annu...

    2021

  18. [26]

    Patrick Kidger, James Foster, Xuechen Li, and Terry J. Lyons. 2021 b . http://proceedings.mlr.press/v139/kidger21b.html Neural sdes as infinite-dimensional gans . In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event...

    2021

  19. [27]

    Patrick Kidger, James Morrill, James Foster, and Terry J. Lyons. 2020 a . https://proceedings.neurips.cc/paper/2020/hash/4a5876b450b45371f6cfe5047ac8cd45-Abstract.html Neural controlled differential equations for irregular time series . In Advances in Neural Information Proces...

    2020

  20. [28]

    Patrick Kidger, James Morrill, James Foster, and Terry J. Lyons. 2020 b . https://proceedings.neurips.cc/paper/2020/hash/4a5876b450b45371f6cfe5047ac8cd45-Abstract.html Neural controlled differential equations for irregular time series . In Advances in Neural Information Proces...

    2020

  21. [29]

    Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, and Yarin Gal. 2024. https://arxiv.org/abs/2406.15927 Semantic entropy probes: Robust and cheap hallucination detection in llms . ArXiv preprint, abs/2406.15927

    2024 arXiv

  22. [30]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019 doi

  23. [31]

    V Levenshtein. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Proceedings of the Soviet physics doklady

    1966

  24. [32]

    Benjamin A Levinstein and Daniel A Herrmann. 2024. https://arxiv.org/pdf/2307.00175 Still no lie detector for language models: Probing empirical and conceptual roadblocks . Philosophical Studies, pages 1--27

    2024 arXiv

  25. [34]

    Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, JingBo Zhu, Xuebo Liu, and Min Zhang. 2022 b . https://doi.org/10.18653/v1/2022.acl-long.571 ODE transformer: An ordinary differential equation-inspired model for sequence generation . In Proceedings of the ...

    2022 doi

  26. [36]

    Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.397 H alu E val: A large-scale hallucination evaluation benchmark for large language models . In Proceedings of the 2023 Conference on Empirical Methods in Na...

    2023 doi

  27. [37]

    Qing Li, Chenyang Lyu, Jiahui Geng, Derui Zhu, Maxim Panov, and Fakhri Karray. 2024. https://arxiv.org/abs/2408.05767 Reference-free hallucination detection for large vision-language models . ArXiv preprint, abs/2408.05767

    2024 arXiv

  28. [38]

    Xuechen Li, Ting - Kam Leonard Wong, Ricky T. Q. Chen, and David Duvenaud. 2020. http://proceedings.mlr.press/v108/li20i.html Scalable gradients for stochastic differential equations . In The 23rd International Conference on Artificial Intelligence and Statistics, AISTATS 2020...

    2020

  29. [39]

    Yuxuan Liang, Kun Ouyang, Hanshu Yan, Yiwei Wang, Zekun Tong, and Roger Zimmermann. 2021. https://doi.org/10.24963/IJCAI.2021/207 Modeling trajectories with neural ordinary differential equations . In Proceedings of the Thirtieth International Joint Conference on Artificial In...

    2021 doi

  30. [40]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. https://doi.org/10.18653/v1/2022.acl-long.229 T ruthful QA : Measuring how models mimic human falsehoods . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa...

    2022 doi

  31. [41]

    Yiping Lu, Zhuohan Li, Di He, Zhiqing Sun, Bin Dong, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2019. https://arxiv.org/abs/1906.02762 Understanding and improving transformer from a multi-particle dynamic system point of view . ArXiv preprint, abs/1906.02762

    2019 arXiv

  32. [42]

    Yiping Lu, Aoxiao Zhong, Quanzheng Li, and Bin Dong. 2018. http://proceedings.mlr.press/v80/lu18d.html Beyond finite layer neural networks: Bridging deep architectures and numerical differential equations . In Proceedings of the 35th International Conference on Machine Learnin...

    2018

  33. [43]

    Potsawee Manakul, Adian Liusie, and Mark Gales. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.557 S elf C heck GPT : Zero-resource black-box hallucination detection for generative large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural...

    2023 doi

  34. [44]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.741 FA ct S core: Fine-grained atomic evaluation of factual precision in long form text generatio...

    2023 doi

  35. [45]

    Niels M \"u ndler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2023. https://arxiv.org/abs/2305.15852 Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation . ArXiv preprint, abs/2305.15852

    2023 arXiv

  36. [46]

    YongKyung Oh, Dongyoung Lim, and Sungil Kim. 2024. https://arxiv.org/abs/2402.14989 Stable neural stochastic differential equations in analyzing irregular time series data . ArXiv preprint, abs/2402.14989

    2024 arXiv

  37. [47]

    Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. 2025. https://openreview.net/forum?id=KRnsX5Em3W LLM s know more than they show: On the intrinsic representation of LLM hallucinations . In The Thirteenth International C...

    2025

  38. [48]

    Lev Semenovich Pontryagin, EF Mishchenko, VG Boltyanskii, and RV Gamkrelidze. 1962. The mathematical theory of optimal processes. Routledge

    1962

  39. [49]

    U ber die numerische aufl \

    Carl Runge. 1895. \"U ber die numerische aufl \"o sung von differentialgleichungen. Mathematische Annalen, 46(2):167--178

  40. [50]

    Da Song, Xuan Xie, Jiayang Song, Derui Zhu, Yuheng Huang, Felix Juefei-Xu, and Lei Ma. 2024. https://arxiv.org/pdf/2310.14211 Luna: A model-based universal analysis framework for large language models . IEEE Transactions on Software Engineering

    2024 arXiv

  41. [51]

    Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. 2024. https://doi.org/10.18653/v1/2024.findings-acl.854 Unsupervised real-time hallucination detection based on the internal states of large language models . In Findings of the Association...

    2024 doi

  42. [52]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca

    2023

  43. [53]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi \`e re, Mihir Sanjay Kale, Juliette Love, et al. 2024. https://arxiv.org/pdf/2403.08295 Gemma: Open models based on gemini research and technology . arXiv ...

    2024 arXiv

  44. [54]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. https://arxiv.org/abs/2307.09288 Llama 2: Open foundation and fine-tuned chat models . ArXiv preprint, abs/2...

    2023 arXiv

  45. [55]

    Yuxia Wang, Revanth Gangi Reddy, Zain Muhammad Mujahid, Arnav Arora, Aleksandr Rubashevskii, Jiahui Geng, Osama Mohammed Afzal, Liangming Pan, Nadav Borenstein, Aditya Pillai, et al. 2023. https://arxiv.org/abs/2311.09000 Factcheck-gpt: End-to-end fine-grained document-level f...

    2023 arXiv

  46. [56]

    Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. 2024. https://arxiv.org/abs/2403.18802 Long-form factuality in large language models . ArXiv preprint, abs/2403.18802

    2024 arXiv

  47. [57]

    Tianhang Zhang, Lin Qiu, Qipeng Guo, Cheng Deng, Yue Zhang, Zheng Zhang, Chenghu Zhou, Xinbing Wang, and Luoyi Fu. 2023 a . https://doi.org/10.18653/v1/2023.emnlp-main.58 Enhancing uncertainty-based hallucination detection with stronger focus . In Proceedings of the 2023 Confe...

    2023 doi

  48. [58]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023 b . https://arxiv.org/abs/2309.01219 Siren's song in the ai ocean: a survey on hallucination in large language models . ArXiv preprint, abs/2309.01219

    2023 arXiv

  49. [59]

    Derui Zhu, Dingfan Chen, Qing Li, Zongxiong Chen, Lei Ma, Jens Grossklags, and Mario Fritz. 2024. https://aclanthology.org/2024.findings-naacl.294 P o LLM graph: Unraveling hallucinations in large language models via state transition dynamics . In Findings of the Association f...

    2024

Pith tools

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