Pith. sign in

REVIEW 4 major objections 6 minor 58 references

An Annotated Reading of 'The Singer of Tales' in the LLM Era

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read An LLM in a single pass composes like an oral bard, not a writer.

desk verdict A thoughtful, genuinely original interpretive essay mapping Parry-Lord concepts onto LLM components, but the copyright conclusion is an overreach and the 'mechanism' claim is more analogy than demonstrated mechanism. read the letter →

arxiv 2502.05148 v1 pith:G2WY4NMC submitted 2025-02-07 cs.CY cs.CL

Kush R. Varshney This is my paper · ORCID
classification cs.CYcs.CL
keywords oral-formulaictheorylargelanguagemodelsgenerativeAIauthorshipcopyrightpost-literatecompositiondecodingmechanism
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the Parry-Lord oral-formulaic theory of how illiterate bards compose epic poetry is the right high-level mechanism for understanding how large language models generate text. It claims that in a single inference pass, an LLM composes under time pressure from learned formulas and themes, much as a singer of tales composes in performance, and that this is 'not writing but akin to oral composition.' If the claim holds, LLM outputs belong to a third, post-literate paradigm of composition, with consequences for how we judge their aesthetics and whether their output is copyrightable. The paper reads Lord's 'The Singer of Tales' chapter by chapter, annotating each passage with LLM equivalents: formulas become constrained heuristic decoding, themes become concepts, and learning stages become pretraining and alignment.

What carries the argument

The carrying object is the oral-formulaic mechanism, defined by Parry and Lord: composition in performance using formulas — 'a group of words which is regularly employed under the same metrical conditions to express a given essential idea' — and themes, the recurring groups of ideas that structure a song. The paper maps each component onto LLMs: tokenization corresponds to the bard's sound-group 'word'; the attention mechanism is what makes analogy-based phrase creation possible; constrained heuristic decoding (greedy or beam search) corresponds to metrical fit; pretraining, fine-tuning, and alignment correspond to the singer's three learning stages; and the first pass of inference, before any editing or inference-time compute, corresponds to the moment of performance. Illiteracy is the load-bearing shared property: neither the bard nor the LLM works from a fixed written original.

What would settle it

Measure the formula-like repetition structure of LLM text generated under single-pass greedy decoding and compare it with oral epic corpora; if the structure is indistinguishable from randomly assembled fluent text, or if it does not change when attention or decoding constraints are removed, the claimed mechanism is not there. Alternatively, test the copyright corollary directly: if single-pass LLM output with no inference-time compute is consistently ruled copyrightable in cases where a human provides only a prompt, the originality-based boundary the paper draws is not what the law tracks.

Watch

Extended reading notes

Core claim

The central claim is that LLM generation, at its first pass, is mechanistically the same kind of activity as oral narrative composition: it is real-time, single-pass, formula-based, and constrained, and therefore should not be classified as writing. The author states that 'we can rightly say that the LLM is as illiterate as the singer of tales,' because both work with sound-group-like tokens and neither has access to a fixed text or an original. From this it follows that LLM responses, like a bard's performance, are not original expressions in the copyright sense, and that human-involvement is the wrong test for copyrightability; the right question is whether originality is part of the generation mechanism. The paper also proposes a new aesthetics for AI-created text, parallel to the way oral poetry was long misjudged by literary standards.

Load-bearing premise

The argument assumes that the resemblance between bards and LLMs is mechanistic rather than merely metaphorical: learned formulas and themes must correspond in a causal way to tokenization, attention, and constrained decoding, and the first pass of inference, without editing or inference-time compute, must be the right object of comparison.

Editorial extensions

If this is right

  • LLM-generated text should be evaluated by its own aesthetic standards, not by the aesthetics developed for written literature.
  • Copyright for LLM output should not turn on whether a human was involved, but on whether the generation mechanism has originality as part of it; by this test, single-pass LLM responses are not subject to copyright.
  • Hallucination is an inherent feature of the oral-formulaic mechanism, and using common factoids keeps hallucination rates low, just as bards stick to common names and places.
  • Adding inference-time compute, memory, or editing moves an LLM from oral composition toward writing, which changes its legal and aesthetic status.
  • Author attribution for AI co-creation should be fine-grained rather than binary, following the 'author function' analysis.

Reading between the lines

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

  • Extension: if the mechanism analogy is causal, then under time-constrained single-pass decoding LLM text should show a measurable formula-like reuse structure similar to oral epic corpora, and removing attention or decoding constraints should weaken that structure.
  • Extension: the paper's copyright argument implies that as LLMs gain inference-time compute and editing, the same model's outputs could straddle the oral/written boundary, so copyright analysis would need to be per-interaction rather than per-model.
  • Extension: the post-literate frame suggests comparing LLM 'themes' with cross-cultural epic themes, such as the trickster slot, as a quantitative handle on model bias and hallucination when structural slots get filled by historical artifacts.
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 / 6 minor

Summary. The paper offers an annotated reading of Alfred Lord's 'The Singer of Tales' (1960) as a lens for understanding large language models. It alternates verbatim passages from Lord with commentary, mapping the bard's apprenticeship (listening, formula learning, ornamentation) onto LLM data curation, pretraining, and alignment; formulas onto heuristic and constrained decoding; themes onto concepts and large concept models; and single-pass generation onto oral composition in performance. The paper concludes that LLMs are 'as illiterate as the singer of tales,' that first-pass LLM generation belongs to a post-literate rather than a literate paradigm, and that LLM responses should not be subject to copyright. It also makes ancillary claims about new aesthetics, attribution, hallucination, and authorship.

Significance. The paper is a readable and genuinely interdisciplinary experiment: it juxtaposes Lord's text with LLM research and draws a number of non-obvious parallels (apprenticeship stages vs. pretraining and fine-tuning, formulas vs. constrained decoding, themes vs. concepts, orality vs. single-pass inference). It is honest about some limitations, notably the exclusion of inference-time compute and the risk of anthropomorphism. If the analogy is meant only as an interpretive lens, the paper is a useful prompt for discussions of AI aesthetics and attribution. Its significance in the stronger sense—as a claim that LLMs share the oral-formulaic mechanism and that this entails noncopyrightability—is not established. The copyright and 'illiteracy' conclusions would need substantially more argument, and the mechanism claim would need a formal or at least falsifiable account of the constraint that plays the role of meter. The paper does not provide that, so its contribution is currently suggestive rather than demonstrative.

major comments (4)
  1. [Section 5 (Chapter Three, The Formula)] The central equivalence between Parry-Lord formulas and LLM decoding is asserted rather than demonstrated. Parry's formula is defined as 'a group of words which is regularly employed under the same metrical conditions to express a given essential idea'; the metrical condition is a hard, enumerable constraint that must be satisfied in real time. The paper's proposed LLM analogues—'style or harmlessness constraints'—are not constraints of this kind: they are learned, prompt-dependent, and not representable as a fixed condition that a generated span either satisfies or fails. Greedy or beam decoding is a heuristic over a probability distribution, not a constraint-satisfaction procedure over a metrical scheme, and repetition in LLM output is a statistical regularity rather than evidence of formulaic composition. Since this section is the load-bearing joint for the claim that 'LLMs have a mechanism quite similar to the oral-formulaic mechanism,' the paper needs either to identify a specifiable hard constraint under which decoding selects from a structured repertoire of formulas or to explicitly downgrade the claim from mechanism to analogy.
  2. [Section 5 (Chapter Three, The Formula) and Section 10 (In Conclusion)] The copyright conclusion does not follow from the analogy. The sentence 'phrases from patterns are not original expressions of ideas' conflates descriptive originality (non-novelty of pattern) with copyright originality, which is a legal standard involving human authorship, minimal creativity, and the idea/expression distinction. Even if LLM responses were entirely formulaic in the oral-formulaic sense, it would not follow that 'neither the singer's poem nor LLM responses should be considered subject to copyright'; oral traditional works can be fixed and protected, and statutory and doctrinal copyright analysis is not engaged anywhere in the paper. The policy claim in Section 10 that LLMs 'are not authors' repeats this leap. At minimum, the conclusion should be framed as a policy argument requiring separate legal support, not as an implication of the literary analogy.
  3. [Section 1 and Section 8 (Chapter Six)] The paper's object of comparison is not stable. Section 1 restricts the analysis to single-pass generation without inference-time compute or editing, stating that the first pass 'is not writing but akin to oral composition.' Section 8, however, says that memory-enhanced architectures and inference-time compute are 'the LLM equivalent of giving the bard reading and writing' and that these capabilities improve fact retrieval, mathematics, and reasoning. This concedes that LLMs can operate in a literate mode, yet the paper does not explain why the first pass—rather than the full interaction including prompting, sampling, and revision—should ground general conclusions about 'LLM responses,' their aesthetics, and their copyrightability. The restriction needs justification, or the conclusions must be explicitly limited to the single-pass setting.
  4. [Section 2 (Foreword)] By the paper's own account, a mechanism must be explanatory, counterfactual, and constructive (citing Craver and Darden). The annotated reading gives correspondences between stages of bardic apprenticeship and LLM training, and between formulas and themes and decoding and concepts, but it does not specify what counterfactual prediction would fail if the analogy were false, nor does it construct an account of how next-token prediction implements constrained formula selection. The plate-tectonics and germ-theory examples cited in Section 2 are causal mechanisms with such predictive content; the oral-formulaic analogy as presented has none. Thus, even if the literary parallels are suggestive, the paper does not deliver the 'mechanism' it promises in the introduction.
minor comments (6)
  1. [Abstract and Section 1] The abstract and Section 1 contain the duplicated phrase 'the the' ('point out the the similarities'), and the abstract has a stray space in 'composed ,'; please proofread the text.
  2. [Section 1] The sentence 'This is a question never asked in this way before, but we will attempt to answer it here in.' ends with 'here in' rather than 'herein' or 'here in this paper'; please fix.
  3. [Section 3 (Chapter One)] The claim that 'LLMs’ output is characteristically different from human writing [47]' is presented as a settled fact, but the cited work is a preprint; consider adding nuance about the strength of the evidence.
  4. [Section 8 (Chapter Six)] The discussion of Pihel's 'post-literate' is very brief, but 'post-literate paradigm' is a key concept for the paper; please integrate the quote and state how LLM generation resembles and differs from hip-hop freestyling.
  5. [Section 4 (Chapter Two)] The footnote about Google naming an LLM 'Bard' is tangential; either develop it or remove it.
  6. [Section 6 (Chapter Four)] The conjecture about oral-formulaic theory inspiring LCMs is speculative and should be explicitly labeled as such rather than presented as a plausible historical influence.

Circularity Check

1 steps flagged · score 6.0 of 10

The formula equivalence is arrived at by broadening Parry's definition until LLM decoding fits, making the central analogy partly definitional.

  1. self definitional [Section 5, 'Chapter Three, The Formula', commentary paragraph beginning 'Parry-Lord formulas are heuristic solutions...']
    "Parry-Lord formulas are heuristic solutions to constrained optimization problems that must be solved in real-time. ... LLMs are rarely prompted to meet metrical constraints, but often have to meet other style or harmlessness constraints. Decoding, the part of LLM inference that produces actual response tokens from probabilities, is often heuristic and constrained as well. Greedy decoding produces many repetitions of the kind seen in Homer; ... Thus, formulas are equally a part of LLMs as they are of the oral tradition."

    Parry's definition, quoted immediately before, makes 'the same metrical conditions' a defining property of a formula; the commentary replaces that property with generic 'style or harmlessness constraints' and treats 'heuristic and constrained' decoding as sufficient. The observed repetition in greedy/beam decoding is then relabeled 'formula,' so the claim that formulas are equally part of LLMs follows from the broadened definition rather than from evidence of a shared hard constraint. The same section concedes 'the constraints are different — metrical ones for bards and style or harmlessness constraints for LLMs,' confirming that the defining condition was dropped rather than matched.

full rationale

The paper is an interpretive essay rather than a fitted model, so most fitted-input and numerical-reduction failure modes do not apply. The training-stage and audience analogies (Chapters 2 and 4) are independent observations, and the Section 1 decision to restrict attention to single-pass inference is an explicit scope assumption rather than a hidden derivation. The main circularity is localized to Chapter 3: Parry's definition of 'formula' is tied to 'the same metrical conditions,' but the commentary replaces that defining condition with generic 'style or harmlessness constraints,' declares decoding 'heuristic and constrained,' and concludes that 'formulas are equally a part of LLMs.' The known fact that greedy and beam decoding produce repetition is thus renamed 'formula' under a broadened definition; the central analogy is partly established by definitional construction rather than by demonstrating a shared hard constraint analogous to meter. The copyright conclusion inherits this step. The cited prior works by the author [4, 52, 53, 54] are contextual and non-load-bearing; the argument does not reduce to them.

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

The paper has no free parameters and introduces no physical or mathematical entities. Its load-bearing assumptions are domain assumptions about the validity of the oral-LLM analogy and about authorship. The central claim rests on these assumptions because it is an interpretive argument, not a derivation.

assumptions (3)
  • domain assumption Single-pass inference without revision is the correct unit of comparison to oral composition.
    Section 1 says inference-time compute follows the first pass as post-processing, so the first pass 'can be said to be not writing but akin to oral composition.' This frames all later claims.
  • domain assumption Parry-Lord formulas and themes have functional analogs in LLM decoding and concepts, and constraints are learned implicitly, as for bards.
    Chapter 3 equates formulas with 'heuristic solutions to constrained optimization problems' and decoding heuristics; Chapter 4 maps themes to concepts. This is the central analogy, assumed rather than proven.
  • domain assumption LLMs, like bards, do not seek originality and are not authors, so authorship and copyright analysis should proceed from the generation mechanism rather than human involvement.
    Chapter 5 argues, via Foucault and Barthes, that authorship is a social construct and LLM responses should not be copyrighted because originality is not part of their mechanism. This normative premise is load-bearing for the policy conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Annotated Reading of 'The Singer of Tales' in the LLM Era." pith.science (2026). https://pith.science/paper/G2WY4NMC

@misc{pith2026250205148,
  author       = {Pith},
  title        = {Pith review of: An Annotated Reading of 'The Singer of Tales' in the LLM Era},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2WY4NMC}},
  note         = {Machine review of arXiv:2502.05148}
}
read the original abstract

The Parry-Lord oral-formulaic theory was a breakthrough in understanding how oral narrative poetry is learned, composed, and transmitted by illiterate bards. In this paper, we provide an annotated reading of the mechanism underlying this theory from the lens of large language models (LLMs) and generative artificial intelligence (AI). We point out the the similarities and differences between oral composition and LLM generation, and comment on the implications to society and AI policy.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 49 canonical work pages

  1. [1]

    Daniel Banks. 2010. From Homer to Hip Hop: Orature and Gri ots, Ancient and Present. Classical World 103, 2 (Winter 2010), 238–245

    work page 2010

  2. [2]

    Roland Barthes. 1972. Myth Today. In Mythologies, Annette Lavers (Ed.). Hill and Wang, 109–159

    work page 1972

  3. [3]

    Roland Barthes. 2016. The Death of the Author. In Readings in the Theory of Religion. Routledge, 141–145

    work page 2016

  4. [4]

    Varshney, Lav R

    Sourya Basu, Prasanna Sattigeri, Karthikeyan Natesan R amamurthy, Vijil Chen- thamarakshan, Kush R. Varshney, Lav R. Varshney, and Payel D as. 2023. Equi- Tuning: Group Equivariant Fine-Tuning of Pretrained Models. In Proceedings of the AAAI Conference on Artificial Intelligence . 6788–6796

    work page 2023

  5. [5]

    Bender, Timnit Gebru, Angelina McMillan-Major , and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major , and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?. In Proceedings of the ACM Conference on Fairness, Accountabili ty, and Transparency. 610–623

    work page 2021

  6. [6]

    William Benzon. 2023. ChatGPT Intimates a Tantalizing F uture, Its Core LLM is Organized on Multiple Levels, and It Has Broken the Idea of Thinking. SSRN:4336442

    work page 2023

  7. [7]

    Leonard Bereska and Stratis Gavves. 2024. Mechanistic I nterpretability for AI Safety - A Review. Transactions on Machine Learning Research (Sept. 2024). https://openreview.net/forum?id=ePUVetPKu6

    work page 2024

  8. [8]

    Erik Brynjolfsson. 1993. The Productivity Paradox of In formation Technology. Commun. ACM 36, 12 (1993), 66–77

    work page 1993

Show all 58 references
  1. [9]

    Kim, Alexander Karpekov, Alec Helbli ng, Zijie J

    Aeree Cho, Grace C. Kim, Alexander Karpekov, Alec Helbli ng, Zijie J. Wang, Seongmin Lee, Benjamin Hoover, and Duen Horng Chau. 2024. Tr ansformer Explainer: Interactive Learning of Text-Generative Model s. In IEEE VIS

    2024

  2. [10]

    Monojit Choudhury. 2023. Generative AI Has a Language P roblem. Nature Human Behaviour 7, 11 (Nov. 2023), 1802–1803

    2023

  3. [11]

    Craver and Lindley Darden

    Carl F. Craver and Lindley Darden. 2013. In Search of Mechanisms: Discoveries Across the Life Sciences . University of Chicago Press

    2013

  4. [12]

    Payel Das, Subhajit Chaudhury, Elliot Nelson, Igor Mel nyk, Sarath Swami- nathan, Sihui Dai, Aurélie Lozano, Georgios Kollias, Vijil Chenthamarakshan, Soham Dan, and Pin-Yu Chen. 2024. Larimar: Large Language Mo dels with Episodic Memory Control. In Proceedings of the Internat...

    2024

  5. [13]

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Ya rin Gal. 2024. Detecting Hallucinations in Large Language Models Using Semantic Ent ropy. Nature 630, 8017 (2024), 625–630

    2024

  6. [14]

    Myles Foley, Ambrish Rawat, Taesung Lee, Yufang Hou, Ga briele Picco, and Giulio Zizzo. 2023. Matching Pairs: Attributing Fine-Tune d Models to Their Pre-Trained Large Language Models. arXiv:2306.09308

    2023 arXiv

  7. [15]

    Michel Foucault. 1969. What is an Author? Technical Report. Société Français de Philosophie

    1969

  8. [16]

    Tairan Fu, Raquel Ferrando, Javier Conde, Carlos Arria ga, and Pedro Reviriego

  9. [17]

    Edd Gent. 2024. Coming to Theaters: An AI-Generated Bol lywood Movie. IEEE Spectrum (Dec. 2024)

    2024

  10. [18]

    Katy Ilonka Gero. 2023. AI and the Writer: How Language Models Support Cre- ative Writers. Ph. D. Dissertation. Columbia University

    2023

  11. [19]

    Kevin Gu, Eva Tuecke, Dmitriy Katz, Raya Horesh, David A lvarez-Melis, and Mikhail Yurochkin. 2024. CharED: Character-Wise Ensemble Decoding for Large Language Models. arXiv:2407.11009

    2024 arXiv

  12. [20]

    Jessica He, Stephanie Houde, and Justin D. Weisz. 2025. Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co- Creation. In Pro- ceedings of the CHI Conference on Human Factors in Computing Systems

    2025

  13. [21]

    Will Douglas Heaven. 2024. Why Does AI Hallucinate? MIT Technology Review 127, 4 (July–Aug. 2024), 20–21

    2024

  14. [22]

    Xiaomeng Hu, Pin-Yu Chen, and Tsung-Yi Ho. 2023. RADAR: Robust AI-Text Detection via Adversarial Learning. In Advances in Neural Information Process- ing Systems. 15077–15095

    2023

  15. [23]

    Shengzhi Huang, Yong Huang, Yinpeng Liu, Zhuoran Luo, a nd Wei Lu. 2025. Are Large Language Models Qualified Reviewers in Originalit y Evaluation? In- formation Processing & Management 62, 3 (2025), 103973

    2025

  16. [24]

    Hubert, Kim N

    Kent F. Hubert, Kim N. Awa, and Darya L. Zabelina. 2024. T he Current State of Artificial Intelligence Generative Language Models is More Creative than Hu- mans on Divergent Thinking Tasks. Scientific Reports 14 (Feb. 2024), 3440

    2024

  17. [25]

    David Jensen. 2024. Explaining the Mechanisms of AI Sys tems to Facilitate Governance. In Trustworthiness and Responsibility in AI – Causality, Lear ning, and Verification, Vaishak Belle, Hana Chockler, Shannon Vallor, Kush R. Vars h- ney, Joost Vennekens, and Sander Beckers (E...

    2024

  18. [26]

    Minna Skafte Jensen. 2017. The Challenge of Oral Epic to Homeric Scholarship. Humanities 6, 4 (Dec. 2017), 97

    2017

  19. [27]

    Lingjie Jiang, Shaohan Huang, Xun Wu, and Furu Wei. 2024 . Textual Aesthetics in Large Language Models. arXiv:2411.02930

    2024 arXiv

  20. [28]

    Samuel G. B. Johnson, Amir-Hossein Karimi, Yoshua Beng io, Nick Chater, To- bias Gerstenberg, Kate Larson, Sydney Levine, Melanie Mitc hell, Iyad Rahwan, Bernhard Schölkopf, and Igor Grossmann. 2024. Imagining an d Building Wise Machines: The Centrality of AI Metacognition. ar...

    2024

  21. [29]

    Adam Tauman Kalai and Santosh S. Vempala. 2024. Calibra ted language models must hallucinate. In Proceedings of the ACM Symposium on Theory of Computing. 160–171

    2024

  22. [30]

    Ching-Yun Ko, Pin-Yu Chen, Payel Das, Youssef Mroueh, S oham Dan, Georgios Kollias, Subhajit Chaudhury, Tejaswini Pedapati, and Luca Daniel. 2024. Large Language Models can be Strong Self-Detoxifiers

    2024

  23. [31]

    Tharindu Kumarage and Huan Liu. 2023. Neural Authorshi p Attribution: Sty- lometric Analysis on Large Language Models. In International Conference on Cyber-Enabled Distributed Computing and Knowledge Discov ery. 51–54

    2023

  24. [32]

    LCM Team, Loïc Barrault, Paul-Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alastruey, Pierre Andrews, Mariano Cor ia, Guillaume Couairon, Marta R. Costa-jussà, David Dale, Hady Elsahar, K evin Heffernan, João Maria Janeiro, Tuan Tran, Christophe Ropers, Eduardo Sá...

    2024 arXiv

  25. [33]

    Claude Lévi-Strauss. 1955. The Structural Study of Myt h. Journal of American Folklore 68, 270 (1955), 428–444

    1955

  26. [34]

    Alfred B. Lord. 1960. The Singer of Tales. Harvard University Press

    1960

  27. [35]

    Ximing Lu, Melanie Sclar, Skyler Hallinan, Niloofar Mi reshghallah, Jiacheng Liu, Seungju Han, Allyson Ettinger, Liwei Jiang, Khyathi Chandu , Nouha Dziri, and Yejin Choi. 2024. AI as Humanity’s Salieri: Quantifying Lin guistic Creativity of Language Models via Systematic Attr...

    2024 arXiv

  28. [36]

    Rohin Manvi, Anikait Singh, and Stefano Ermon. 2024. Ad aptive Inference- Time Compute: LLMs Can Predict If They Can Do Better, Even Mid -Generation. arXiv:2410.02725

    2024 arXiv

  29. [37]

    Lucas Monteiro Paes, Dennis Wei, Hyo Jin Do, Hendrik Str obelt, Ronny Luss, Amit Dhurandhar, Manish Nagireddy, Karthikeyan Natesan Ramamurthy, Prasanna Sattigeri, Werner Geyer, and Soumya Ghosh. 2024. M ulti-Level Expla- nations for Generative Language Models

    2024

  30. [38]

    Muhammad Umair Nasir, Sam Earle, Julian Togelius, Stev en James, and Christo- pher Cleghorn. 2024. LLMatic: Neural Architecture Search v ia Large Language Models and Quality Diversity Optimization. In Proceedings of the Genetic and Evolutionary Computation Conference. 1110–1118

    2024

  31. [39]

    Daly, Michael Hind, Werner Geyer, Ambrish Rawat, Kush R

    Inkit Padhi, Manish Nagireddy, Giandomenico Cornacch ia, Subhajit Chaud- hury, Tejaswini Pedapati, Pierre Dognin, Keerthiram Murugesan, Erik Miehling, Martín Santillán Cooper, Kieran Fraser, Giulio Zizzo, Muhammad Zaid Hameed, Mark Purcell, Michael Desmond, Qian Pan, Zahra Ash...

    2024 arXiv

  32. [40]

    Artidoro Pagnoni, Ram Pasunuru, Pedro Rodriguez, John Nguyen, Benjamin Muller, Margaret Li, Chunting Zhou, Lili Yu, Jason Weston, L uke Zettlemoyer, Gargi Ghosh, Mike Lewis, Ari Holtzman, and Srinivasan Iyer. 2024. Byte Latent Kush R. Varshney Transformer: Patches Scale Better...

    2024

  33. [41]

    Core Francisco Park, Andrew Lee, Ekdeep Singh Lubana, Y ongyi Yang, Maya Okawa, Kento Nishi, Martin Wattenberg, and Hidenori Tanaka . 2024. ICLR: In- Context Learning of Representations

    2024

  34. [42]

    Erik Pihel. 1996. A Furified Freestyle: Homer and Hip Hop . Oral Tradition 11, 2 (Oct. 1996), 249–269

    1996

  35. [43]

    Morgan S. Porter. 2024. Moral Agency in Silico: Explori ng Free Will in Large Language Models. arXiv:2410.23310

    2024 arXiv

  36. [44]

    Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia , Prateek Mittal, and Peter Henderson. 2023. Fine-Tuning Aligned Language Model s Compromises Safety, Even When Users Do Not Intend To! arXiv:2310.03693

    2023 arXiv

  37. [45]

    Orlando Reade. 2025. Why Is the Right Obsessed With Epic Poetry? The Nation (Jan. 2025)

    2025

  38. [46]

    Redman, Juan M

    William T. Redman, Juan M. Bello-Rivas, Maria Fonobero va, Ryan Mohr, Yannis Kevrekidis, and Igor Mezic. 2024. Identifying Equivalent T raining Dynamics. In Advances in Neural Information Processing Systems

    2024

  39. [47]

    Alex Reinhart, David West Brown, Ben Markey, Michael La udenbach, Kachatad Pantusen, Ronald Yurko, and Gordon Weinberg. 2024. Do LLMs W rite like Hu- mans? Variation in Grammatical and Rhetorical Styles. arXi v:2410.16107

    2024 arXiv

  40. [48]

    Jean-Francois Santucci, Albert Doja, and Laurent Capo cchi. 2020. A Discrete- Event Simulation of Claude Lévi-Strauss’ Structural Analysis of Myths Based on Symmetry and Double Twist Transformations. Symmetry 12, 10 (2020), 1706

    2020

  41. [49]

    Shaily. 2025. AI Generated Gods and Impacts on the Infor mation Ecosystem. https://shaily99.substack.com/p/ai-generated-gods-and-impacts-on

    2025

  42. [50]

    Ben Shneiderman and Michael Muller. 2023. On AI Anthrop omor- phism. https://medium.com/human-centered-ai/on-ai-an thropomorphism- abff4cecc5ae

    2023

  43. [51]

    Shannon Vallor. 2024. The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. Oxford University Press

    2024

  44. [52]

    Varshney

    Kush R. Varshney. 2024. Decolonial AI Alignment: Openn ess, Viśes.a-Dharma, and Including Excluded Knowledges. In Proceedings of the AAAI/ACM Confer- ence on AI, Ethics, and Society . 1467–1481

    2024

  45. [53]

    Varshney, Zahra Ashktorab, Djallel Bouneffouf, Matthew Riemer, and Justin D

    Kush R. Varshney, Zahra Ashktorab, Djallel Bouneffouf, Matthew Riemer, and Justin D. Weisz. 2025. Scopes of Alignment. InProceedings of the AAAI Workshop on AI Governnace

    2025

  46. [54]

    Varshney and Lav R

    Kush R. Varshney and Lav R. Varshney. 2023. A Banal Accou nt of a Safety- Creativity Tradeoff in Generative AI. In Joint Proceedings of the ACM Interna- tional Conference on Intelligent User Interfaces Workshop s

    2023

  47. [55]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko reit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. A ttention Is All You Need. In Advances in Neural Information Processing Systems

    2017

  48. [56]

    Xinyi Wang, Antonis Antoniades, Yanai Elazar, Alfonso Amayuelas, Alon Al- balak, Kexun Zhang, and William Yang Wang. 2024. Generaliza tion v.s. Mem- orization: Tracing Language Models’ Capabilities Back to P retraining Data. arXiv:2407.14985

    2024 arXiv

  49. [57]

    Wenjun Zeng, Yuchi Liu, Ryan Mullins, Ludovic Peran, Jo e Fernandez, Hamza Harkous, Karthik Narasimhan, Drew Proud, Piyush Kumar, Bha ktipriya Rad- harapu, Olivia Sturman, and Oscar Wahltinez. 2024. ShieldG emma: Generative AI Content Moderation Based on Gemma. arXiv:2407.21772

    2024 arXiv

  50. [2024]

    Why Do Large Language Models (LLMs) Struggle to Count L etters?

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.