pith. sign in

arxiv: 2607.00447 · v1 · pith:6BMKK4RBnew · submitted 2026-07-01 · 💻 cs.CL

Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors

Yangfan Hu , Xuhan Tong , Haoyue Bai , Xi Ding , Shashank Muralidhar Bharadwaj , Siyang Cao , Robert Nowak , Jiawei Zhang This is my paper

Pith reviewed 2026-07-02 13:32 UTC · model grok-4.3

classification 💻 cs.CL
keywords hallucinationinference misalignmentlatent key-task modelTrapQAScientistQAconstrained QAshortcut bias
0
0 comments X

The pith

Hallucinations in language models often arise from biased latent inference rather than absent knowledge.

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

The paper asks whether hallucinations occur because models lack facts or because they follow the wrong inference path despite having the facts. It formalizes inference misalignment as a mismatch where pretraining-frequency imbalances let statistically salient shortcuts override the answer supported by the prompt. A latent key-task model predicts this produces positive inference loss through two modes: task-retrieval bias during entity disambiguation and key-selection bias during action choice. The authors test the idea with TrapQA, a controlled testbed containing ScientistQA for disambiguation among similar scientists and Real-Life Constrained QA for everyday constraint following. Results indicate that the bias can occur even when relevant knowledge is present.

Core claim

The authors claim that hallucination can arise from biased latent inference rather than absent knowledge alone. They formalize this with a latent key-task model in which pretraining-frequency imbalance causes a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. TrapQA, consisting of ScientistQA with factual probes and Real-Life Constrained QA, isolates these effects and shows that the misalignment persists when knowledge is confirmed present.

What carries the argument

The latent key-task model, which separates a prompt-supported reasoning path from a statistically salient latent association path whose dominance due to pretraining imbalance produces inference loss.

If this is right

  • Models can produce constraint-violating answers on tasks where the relevant facts are present but overshadowed by frequent pretraining patterns.
  • Task-retrieval bias will appear most clearly in entity disambiguation settings with similar candidates.
  • Key-selection bias will appear in action-choice settings where a salient but incorrect option competes with the prompt constraint.
  • Reducing hallucinations requires addressing inference path selection in addition to knowledge acquisition.

Where Pith is reading between the lines

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

  • Data curation that balances association frequencies during pretraining could reduce these specific failure modes.
  • Existing factual-recall benchmarks may underestimate hallucination risk if they do not control for latent shortcut strength.
  • The same misalignment mechanism may apply to other constrained generation tasks such as code synthesis or logical reasoning.

Load-bearing premise

The TrapQA testbed successfully isolates inference misalignment from confounding factors such as incomplete knowledge or model-specific architectural quirks.

What would settle it

If models with verified factual knowledge via the supplementary probes in ScientistQA still produce identical hallucination rates on high-frequency versus low-frequency association items, the claim that biased latent inference drives the errors would be falsified.

Figures

Figures reproduced from arXiv: 2607.00447 by Haoyue Bai, Jiawei Zhang, Robert Nowak, Shashank Muralidhar Bharadwaj, Siyang Cao, Xi Ding, Xuhan Tong, Yangfan Hu.

Figure 1
Figure 1. Figure 1: Hallucination as misaligned inference. In the motivating [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative examples from Real-Life Con [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Scientist QA construction pipeline. Starting from Wikipedia-linked scientist profiles, we [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example structured scientist profile used in Scientist QA. [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example prepend_profiles prompt variant. The ellipses in the profile example indicate omitted attributes for readability. label. H Extended Empirical Results This appendix provides the full empirical break￾downs supporting Section 5. Unless otherwise stated, all tables refer to the retrieval-sensitive prepend_names condition over 2,925 Scientist QA questions. H.1 Full Probe-State Breakdowns For each pairwi… view at source ↗
read the original abstract

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper claims that hallucinations in large language models can arise from inference misalignment—where pretraining-frequency imbalances cause latent shortcut paths to dominate constraint-sensitive reasoning—rather than from absent knowledge alone. It formalizes this via a latent key-task model predicting task-retrieval bias in entity disambiguation and key-selection bias in action choice, and introduces the TrapQA testbed (ScientistQA with supplementary factual probes for scientist disambiguation, plus Real-Life Constrained QA) to empirically separate these factors, reporting that results support biased latent inference as a distinct hallucination mechanism.

Significance. If the separation between knowledge possession and inference bias is rigorously demonstrated, the work provides a useful diagnostic framework for hallucination causes that could inform more targeted mitigation beyond knowledge injection. The controlled testbed approach and explicit formalization of failure modes represent a constructive contribution to mechanistic understanding in the field.

major comments (1)
  1. [ScientistQA (abstract and TrapQA description)] ScientistQA component (abstract): the supplementary factual probes must be shown to independently confirm knowledge possession without inheriting the same pretraining-frequency biases (e.g., name-frequency imbalances or similar constraint structures) tested in the main disambiguation tasks. Without explicit controls demonstrating that probe success is decoupled from the entity-disambiguation bias, the claim that models possess the relevant facts yet follow the shortcut path due to latent associations remains unproven, directly undermining the inference-misalignment interpretation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential contribution. We address the major comment below.

read point-by-point responses

  1. Referee: [ScientistQA (abstract and TrapQA description)] ScientistQA component (abstract): the supplementary factual probes must be shown to independently confirm knowledge possession without inheriting the same pretraining-frequency biases (e.g., name-frequency imbalances or similar constraint structures) tested in the main disambiguation tasks. Without explicit controls demonstrating that probe success is decoupled from the entity-disambiguation bias, the claim that models possess the relevant facts yet follow the shortcut path due to latent associations remains unproven, directly undermining the inference-misalignment interpretation.

    Authors: We agree that demonstrating the factual probes' independence from the pretraining-frequency biases is essential to rigorously support the inference-misalignment interpretation. The probes are structured as direct factual queries, distinct from the constraint-sensitive disambiguation tasks, but we acknowledge that explicit decoupling controls are not sufficiently detailed in the current version. In the revised manuscript, we will add analyses (e.g., stratifying probe accuracy by entity frequency and testing frequency-matched controls) to confirm that probe success is not driven by the same biases, thereby strengthening the separation between knowledge possession and latent inference paths. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical diagnostic framework with independent testbed design

full rationale

The paper presents an empirical diagnostic framework via the TrapQA testbed (ScientistQA with factual probes and Real-Life Constrained QA) to distinguish inference misalignment from absent knowledge. The latent key-task model is introduced conceptually to predict failure modes from pretraining-frequency imbalance, but no equations, fitted parameters, or derivations are shown that reduce the central claim to inputs by construction. No self-citations, ansatzes, or uniqueness theorems are invoked as load-bearing. The separation between knowledge possession and biased inference is attempted through supplementary probes and controlled tasks, making the derivation self-contained against external benchmarks rather than self-referential.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review; ledger populated from explicit statements in the abstract.

axioms (1)
  • domain assumption Pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path
    Invoked to explain positive inference loss in the latent key-task model
invented entities (1)
  • latent key-task model no independent evidence
    purpose: Formalize inference misalignment between prompt-supported answer and statistically salient associations
    Introduced as the central modeling device

pith-pipeline@v0.9.1-grok · 5715 in / 1230 out tokens · 21338 ms · 2026-07-02T13:32:13.159258+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

72 extracted references · 11 canonical work pages

  1. [1]

    On Faithfulness and Factuality in Abstractive Summarization

    Maynez, Joshua and Narayan, Shashi and Bohnet, Bernd and McDonald, Ryan. On Faithfulness and Factuality in Abstractive Summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.173

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

  2. [2]

    2019 , url=

    Hallucinations in Neural Machine Translation , author=. 2019 , url=

    2019

  3. [3]

    Challenges in Data-to-Document Generation

    Wiseman, Sam and Shieber, Stuart and Rush, Alexander. Challenges in Data-to-Document Generation. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. 2017. doi:10.18653/v1/D17-1239

    work page doi:10.18653/v1/d17-1239 2017

  4. [4]

    2025 , eprint=

    Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations , author=. 2025 , eprint=

    2025

  5. [5]

    On the Origin of Hallucinations in Conversational Models: Is it the Datasets or the Models?

    Dziri, Nouha and Milton, Sivan and Yu, Mo and Zaiane, Osmar and Reddy, Siva. On the Origin of Hallucinations in Conversational Models: Is it the Datasets or the Models?. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.387

    work page doi:10.18653/v1/2022.naacl-main.387 2022

  6. [6]

    2023 , eprint=

    How Language Model Hallucinations Can Snowball , author=. 2023 , eprint=

    2023

  7. [7]

    2025 , eprint=

    H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs , author=. 2025 , eprint=

    2025

  8. [8]

    T ruthful QA : Measuring How Models Mimic Human Falsehoods

    Lin, Stephanie and Hilton, Jacob and Evans, Owain. T ruthful QA : Measuring How Models Mimic Human Falsehoods. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.229

    work page doi:10.18653/v1/2022.acl-long.229 2022

  9. [9]

    An Explanation of In-Context Learning as Implicit Bayesian Inference , url =

    Xie, Sang Michael and Raghunathan, Aditi and Liang, Percy and Ma, Tengyu , booktitle =. An Explanation of In-Context Learning as Implicit Bayesian Inference , url =

  10. [10]

    2026 , eprint=

    Demonstrations, CoT, and Prompting: A Theoretical Analysis of ICL , author=. 2026 , eprint=

    2026

  11. [11]

    Language Models are Few-Shot Learners , url =

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel and Wu, Jeffrey and Winte...

  12. [12]

    H alu E val: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models

    Li, Junyi and Cheng, Xiaoxue and Zhao, Xin and Nie, Jian-Yun and Wen, Ji-Rong. H alu E val: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.397

    work page doi:10.18653/v1/2023.emnlp-main.397 2023

  13. [13]

    Small World of Words

    The “Small World of Words” English word association norms for over 12,000 cue words , author=. Behavior research methods , volume=. 2019 , publisher=

    2019

  14. [14]

    2025 , eprint=

    Why Language Models Hallucinate , author=. 2025 , eprint=

    2025

  15. [15]

    2025 , eprint=

    Hallucination is Inevitable: An Innate Limitation of Large Language Models , author=. 2025 , eprint=

    2025

  16. [16]

    2025 , eprint=

    Measuring the Impact of Lexical Training Data Coverage on Hallucination Detection in Large Language Models , author=. 2025 , eprint=

    2025

  17. [17]

    2024 , eprint=

    Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? , author=. 2024 , eprint=

    2024

  18. [18]

    2024 , eprint=

    Unfamiliar Finetuning Examples Control How Language Models Hallucinate , author=. 2024 , eprint=

    2024

  19. [19]

    2025 , eprint=

    Harnessing RLHF for Robust Unanswerability Recognition and Trustworthy Response Generation in LLMs , author=. 2025 , eprint=

    2025

  20. [20]

    2026 , eprint=

    Hallucination Detection and Mitigation in Large Language Models , author=. 2026 , eprint=

    2026

  21. [21]

    2026 , eprint=

    Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations , author=. 2026 , eprint=

    2026

  22. [22]

    2026 , eprint=

    Large Language Models Hallucination: A Comprehensive Survey , author=. 2026 , eprint=

    2026

  23. [23]

    2021 , eprint=

    Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metrics , author=. 2021 , eprint=

    2021

  24. [24]

    2024 , eprint=

    In Defense of Lazy Visual Grounding for Open-Vocabulary Semantic Segmentation , author=. 2024 , eprint=

    2024

  25. [25]

    2021 , eprint=

    The Curious Case of Hallucinations in Neural Machine Translation , author=. 2021 , eprint=

    2021

  26. [26]

    2022 , eprint=

    Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? , author=. 2022 , eprint=

    2022

  27. [27]

    2023 , eprint=

    LIMA: Less Is More for Alignment , author=. 2023 , eprint=

    2023

  28. [28]

    2022 , eprint=

    Training language models to follow instructions with human feedback , author=. 2022 , eprint=

    2022

  29. [29]

    2023 , eprint=

    Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback , author=. 2023 , eprint=

    2023

  30. [30]

    2020 , eprint=

    The Curious Case of Neural Text Degeneration , author=. 2020 , eprint=

    2020

  31. [31]

    2024 , eprint=

    Calibrated Language Models Must Hallucinate , author=. 2024 , eprint=

    2024

  32. [32]

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

    Ji, Ziwei and Lee, Nayeon and Frieske, Rita and Yu, Tiezheng and Su, Dan and Xu, Yan and Ishii, Etsuko and Bang, Ye Jin and Madotto, Andrea and Fung, Pascale , year=. Survey of Hallucination in Natural Language Generation , volume=. ACM Computing Surveys , publisher=. doi:10.1145/3571730 , number=

    work page doi:10.1145/3571730

  33. [33]

    2023 , eprint=

    Do Large Language Models Know What They Don't Know? , author=. 2023 , eprint=

    2023

  34. [34]

    2026 , eprint=

    PretrainRL: Alleviating Factuality Hallucination of Large Language Models at the Beginning , author=. 2026 , eprint=

    2026

  35. [35]

    Forty-first International Conference on Machine Learning , year=

    Understanding Finetuning for Factual Knowledge Extraction , author=. Forty-first International Conference on Machine Learning , year=

  36. [36]

    2024 , eprint=

    FLAME: Factuality-Aware Alignment for Large Language Models , author=. 2024 , eprint=

    2024

  37. [37]

    2024 , eprint=

    Learning or Self-aligning? Rethinking Instruction Fine-tuning , author=. 2024 , eprint=

    2024

  38. [38]

    2025 , eprint=

    TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  39. [39]

    2024 , eprint=

    R-Tuning: Instructing Large Language Models to Say `I Don't Know' , author=. 2024 , eprint=

    2024

  40. [40]

    2023 , eprint=

    Sources of Hallucination by Large Language Models on Inference Tasks , author=. 2023 , eprint=

    2023

  41. [41]

    A is B" fail to learn

    The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A" , author=. 2024 , eprint=

    2024

  42. [42]

    2023 , eprint=

    Why Does ChatGPT Fall Short in Providing Truthful Answers? , author=. 2023 , eprint=

    2023

  43. [43]

    2022 , eprint=

    Towards Improving Faithfulness in Abstractive Summarization , author=. 2022 , eprint=

    2022

  44. [44]

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

    Huang, Lei and Yu, Weijiang and Ma, Weitao and Zhong, Weihong and Feng, Zhangyin and Wang, Haotian and Chen, Qianglong and Peng, Weihua and Feng, Xiaocheng and Qin, Bing and Liu, Ting , year=. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions , volume=. ACM Transactions on Information Systems , publis...

    work page doi:10.1145/3703155

  45. [45]

    2023 , eprint=

    Deep reinforcement learning from human preferences , author=. 2023 , eprint=

    2023

  46. [46]

    2020 , eprint=

    Fine-Tuning Language Models from Human Preferences , author=. 2020 , eprint=

    2020

  47. [47]

    2019 , eprint=

    Evaluating the Factual Consistency of Abstractive Text Summarization , author=. 2019 , eprint=

    2019

  48. [48]

    2020 , eprint=

    Asking and Answering Questions to Evaluate the Factual Consistency of Summaries , author=. 2020 , eprint=

    2020

  49. [49]

    2021 , eprint=

    SummEval: Re-evaluating Summarization Evaluation , author=. 2021 , eprint=

    2021

  50. [50]

    ToTTo : A Controlled Table-To-Text Generation Dataset

    Parikh, Ankur and Wang, Xuezhi and Gehrmann, Sebastian and Faruqui, Manaal and Dhingra, Bhuwan and Yang, Diyi and Das, Dipanjan. ToTTo : A Controlled Table-To-Text Generation Dataset. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.89

    work page doi:10.18653/v1/2020.emnlp-main.89 2020

  51. [51]

    2022 , eprint=

    TRUE: Re-evaluating Factual Consistency Evaluation , author=. 2022 , eprint=

    2022

  52. [52]

    2023 , eprint=

    Evaluating Hallucinations in Chinese Large Language Models , author=. 2023 , eprint=

    2023

  53. [53]

    2023 , eprint=

    HalOmi: A Manually Annotated Benchmark for Multilingual Hallucination and Omission Detection in Machine Translation , author=. 2023 , eprint=

    2023

  54. [54]

    2024 , eprint=

    HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models , author=. 2024 , eprint=

    2024

  55. [55]

    Unified Hallucination Detection for Multimodal Large Language Models

    Chen, Xiang and Wang, Chenxi and Xue, Yida and Zhang, Ningyu and Yang, Xiaoyan and Li, Qiang and Shen, Yue and Liang, Lei and Gu, Jinjie and Chen, Huajun. Unified Hallucination Detection for Multimodal Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/...

    work page doi:10.18653/v1/2024.acl-long.178 2024

  56. [56]

    2025 , eprint=

    Fact, Fetch, and Reason: A Unified Evaluation of Retrieval-Augmented Generation , author=. 2025 , eprint=

    2025

  57. [57]

    RAGT ruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models

    Niu, Cheng and Wu, Yuanhao and Zhu, Juno and Xu, Siliang and Shum, KaShun and Zhong, Randy and Song, Juntong and Zhang, Tong. RAGT ruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v...

    work page doi:10.18653/v1/2024.acl-long.585 2024

  58. [58]

    2024 , eprint=

    WildHallucinations: Evaluating Long-form Factuality in LLMs with Real-World Entity Queries , author=. 2024 , eprint=

    2024

  59. [59]

    2024 , eprint=

    Measuring short-form factuality in large language models , author=. 2024 , eprint=

    2024

  60. [60]

    2024 , eprint=

    Long-form factuality in large language models , author=. 2024 , eprint=

    2024

  61. [61]

    FA ct S core: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation

    Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh. FA ct S core: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.1...

    work page doi:10.18653/v1/2023.emnlp-main.741 2023

  62. [62]

    New Embedding Models and API Updates , year =

  63. [63]

    Models & Pricing , year =

  64. [64]

    2026 , month = jan, howpublished =

    Peter Steinberger , title =. 2026 , month = jan, howpublished =

    2026

  65. [65]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  66. [66]

    2025 , eprint=

    Gemini: A Family of Highly Capable Multimodal Models , author=. 2025 , eprint=

    2025

  67. [67]

    2025 , eprint=

    DeepSeek-V3 Technical Report , author=. 2025 , eprint=

    2025

  68. [68]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  69. [69]

    2022 , eprint=

    WebGPT: Browser-assisted question-answering with human feedback , author=. 2022 , eprint=

    2022

  70. [70]

    2023 , eprint=

    WebGLM: Towards An Efficient Web-Enhanced Question Answering System with Human Preferences , author=. 2023 , eprint=

    2023

  71. [71]

    2025 , eprint=

    Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models , author=. 2025 , eprint=

    2025

  72. [72]

    2026 , howpublished =

    2026