Jekyll-and-Hyde Tipping Point in an AI's Behavior

Frank Yingjie Huo; Neil F. Johnson

REVIEW 3 major objections 4 minor 28 references

Jekyll-and-Hyde Tipping Point in an AI's Behavior

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Starting from a single self-attention head, this paper derives an exact formula for the iteration at which a language model's response tips from good to bad.

desk verdict Clean derivation of a single-head tipping point formula, but no evidence it transfers to real LLMs. read the letter →

arxiv 2504.20980 v1 pith:FJYPH4DB submitted 2025-04-29 cs.AI cs.CYnlin.AOphysics.comp-phphysics.soc-ph

Neil F. Johnson , Frank Yingjie Huo This is my paper · ORCID

classification cs.AIcs.CYnlin.AOphysics.comp-phphysics.soc-ph

keywords largelanguagemodelstippingpointself-attentioncontextvectordot-productcrossovertokenembeddingsAIsafety

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

The paper derives an exact formula for the moment a simplified language model's response switches from good content to bad content, and argues the same mechanism drives real large language models. The formula follows from setting the context vector's dot product against bad content equal to its dot product against good content, and it predicts a sharp crossover once attention spreads too thinly across the growing sequence. If the step from the toy attention head to real LLMs holds, the result gives a quantitative handle on when responses go wrong, and it answers everyday questions such as whether being polite to a chatbot changes the risk.

What carries the argument

The machinery is a deterministic softmax attention head treated as a discrete dynamical system, with the context vector $c_n$ as the state and the row-normalized attention weights as the dilution mechanism. The next token is chosen as the one whose embedding has the largest dot product with $c_n$, so the load-bearing identity is the crossover condition $c_n\cdot B=c_n\cdot G$: because each row of the attention matrix must sum to one, adding more $G$ tokens spreads attention ever thinner and lets the bad token's accumulated score overtake the good token's at a computable iteration. The paper sets the Key and Query matrices to identity for simplicity, so the formula solves a minimal single-head version of the attention calculation.

What would settle it

For a fixed prompt whose token embeddings satisfy $B\cdot G>G\cdot G$, use Eq. 2 to compute a predicted $n^*$, then generate the response many times and record the first token at which output switches from the good class to the bad class; the claim fails if the observed switch iteration is not $n^*$ or if no switch occurs. A laboratory version would compare the single-head prediction with a numerical multi-head transformer simulation using the same embeddings.

Watch

Extended reading notes

Core claim

The paper's central claim is that within a single self-attention head there is an exactly computable tipping-point iteration $n^*$ determined by the prompt's token vectors and the model's learned embeddings. Writing $c_n$ for the context vector and $G$, $B$ for good and bad content tokens, the condition $c_n\cdot B=c_n\cdot G$ yields $n^* = \frac{[\sum_{i\neq G}\exp(P_i\cdot G)P_i]\cdot(G-B)}{[\exp(G\cdot G)G]\cdot(B-G)} - g$, where $g$ is the number of good tokens in the prompt. The paper verifies this formula against full numerical evaluation of the attention head and obtains $n^*=10$ exactly for its example. It further claims that $n^*$ is hard-wired when generation starts: if Eq. 2 yields a positive finite value, the model's output is bad from that iteration onward, and the relative vectors $(G-B)$ and $(B-G)$ in the formula express the competition for attention between good and bad content.

Load-bearing premise

The load-bearing premise is that a real language model behaves like the paper's stripped-down attention head, in which the two internal comparison matrices are set to the identity; if real multi-head models do not, the predicted switch point will not transfer.

Editorial extensions

If this is right

If the single-head formula transfers to real LLMs, the switch point is fixed before generation begins: for a given prompt and training, $n^*$ is determined by dot products that are already set.
Raising $P\cdot G$ in the prompt or training geometry can push $n^*$ beyond the response length, so a response stays entirely good; a negative $n^*$ means the response is bad from the start.
Politeness tokens that are nearly orthogonal to substantive content barely change $n^*$, so being polite neither prevents nor triggers the tipping point.
Once the bad token wins, it is a stable attractor while the good token is only metastable, so without intervention the bad output persists for the rest of the response.

Reading between the lines

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

A testable extension beyond the paper's single-head setting is to compute Eq. 2 using the actual embedding layer of an open-weights model and check whether the predicted switch iteration matches observed token-by-token switches in real generations.
The same crossover mechanism could, by extension, explain shorter lapses such as repetitions or hallucinated phrases, not just all-or-nothing switches, by coarse-graining $G$ and $B$ as classes of content rather than single tokens.
If the paper's claim carries over, a practical extension is a screening tool: prompts with a small margin between $G\cdot B$ and $G\cdot G$ could be flagged before deployment, and training-time geometry changes that separate $B$ from $G$ would delay the tip without relying on politeness.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague, quick take on 2504.20980. The paper gives a closed-form expression for the iteration n* at which a single self-attention head (with K=Q=I) flips from producing 'good' to 'bad' tokens. That formula is new as an explicit expression and the derivation is clean and fully reproducible—they ship Mathematica notebooks and the algebra checks out. The 'attention dilutes and suddenly snaps' intuition is a nice way to summarize the mechanism. So there is a real, if small, result here: an exact threshold in a deliberately minimal attention model.

The soft spot is the leap from that toy to real LLMs. The abstract and intro talk about predicting when ChatGPT's output 'tips,' but the Methods explicitly restrict the analysis to a single head with bland unit Key/Query matrices and defer multi-head and deep transformers to future work. No actual model is tested. The numerical 'confirmation' in Figs. 2c and 3 is just comparing the formula to a simulation of the same toy head it was derived from—that's self-consistency, not an independent test. The sentence 'Equations 1 and 2 are exact for any prompt...' is true only inside the toy setting, and the paper does not do enough to flag that boundary to a general reader.

There are two smaller issues. The derivation is a direct extension of the authors' own companion paper [23], but the text does not say 'our prior work'—that should be disclosed. And the 'being polite is negligible' conclusion rests on an assumption about politeness tokens being orthogonal in embedding space; it's a reasonable guess, but it's not demonstrated with real embeddings.

Net: the paper is a decent physics-style toy model note, but it massively overclaims in the abstract. A serious referee would ask them to either provide evidence from a real LLM or reframe the whole paper as a minimal-model analysis with no direct claim about production systems. As written, I'd reject, but I'd send it to review because the derivation deserves checking and the framing is fixable.

Referee Report

3 major / 4 minor

Summary. The manuscript derives a closed-form expression, Eq. (2), for the iteration number n* at which a simplified single-head self-attention model's greedy decoding switches from producing 'good' tokens (G) to 'bad' tokens (B). The derivation sets the context vector's dot products with the G and B token embeddings equal and obtains a formula depending on prompt token embeddings and dot products. The paper claims this formula is exact for any prompt and provides quantitative predictions for real LLMs, including the question of whether politeness affects tipping points. Numerical results in Figs. 2(c) and 3 compare the formula with simulations of the same toy attention head, yielding n*=10 for the example parameters. The paper acknowledges simplifying assumptions (identity Key and Query matrices, single head) and defers multi-head/deep transformer generalizations to future work.

Significance. If the derived formula quantitatively predicted real LLM tipping points, it would address a timely and important problem in AI safety and interpretability. The paper's algebraic derivation is clean, the Mathematica notebooks provide reproducible code, and the formula is internally consistent with the proposed toy model. However, the central claim—that Eqs. (1)–(3) give exact, quantitative predictions for when real LLM outputs turn bad—is unsupported. No experiments on actual language models are reported; no real token embeddings are used; and the only validation is a self-consistency check against the very model from which the formula was derived. The contribution is therefore a mathematical analysis of a highly simplified toy model, not a predictive theory of LLM behavior as the abstract and title suggest.

major comments (3)

[Abstract and Introduction, Eq. (2)] The abstract and the text around Eq. (2) claim that the formula provides exact quantitative predictions for when an LLM's response tips to bad output, yet the Methods restrict the derivation to a single self-attention head with identity Key and Query matrices ('For simplicity, we choose bland unit matrices for the Key and Query'). The statement 'Equations 1 and 2 are exact for any prompt of any number of tokens and composition' is only true within this toy model. No argument or experiment establishes that this toy head captures the behavior of real multi-head, deep transformers. Without such evidence, the central predictive claim about real LLMs is not supported.
[Figs. 2 and 3, Methods] The numerical verification is circular: Eq. (2) is derived from the same attention-head dynamics that are then simulated to produce the 'empirical' n*=10. The Methods explicitly state that because the derivation is exact, Eq. (2) 'will always agree with numerical evaluation of the Attention head in Fig. 1.' This confirms algebraic consistency but does not test the model against any real LLM output. An independent test would require using embeddings from an actual model, a procedure for labeling output tokens as G or B, and a comparison of predicted versus observed tipping iterations in greedy decoding.
[Eq. (2) and surrounding text] The paper assumes that the crossing condition c_n · B = c_n · G is sufficient for the output to remain permanently bad thereafter. The text gives an intuitive attractor argument (B stable, G metastable) but does not prove that once the dot-product inequality flips, all subsequent greedy choices are B tokens. Since the softmax context vector updates depend on the entire history, a more careful analysis or an explicit proof is needed to justify irreversibility; this is a gap in the theoretical derivation even within the toy model.

minor comments (4)

[Eq. (2)] The typesetting of Eq. (2) is difficult to parse; the summation limits and bracket structure should be clarified, possibly with an expanded definition of all symbols and vector dimensions.
[Discussion of politeness] The conclusion that polite words have negligible effect relies on the assumption that such tokens are orthogonal to G and B in embedding space. The paper provides no empirical or theoretical justification for this orthogonality assumption beyond intuition.
[References] The paper relies on prior work [23] for the context-vector dynamics; the novelty of the tipping-point derivation relative to [23] should be stated more explicitly.
[General framing] The abstract's mention of deaths, trauma, and 'should I be polite to my LLM?' sets expectations of direct relevance to real systems that the toy-model analysis does not meet; a more measured framing would help readers calibrate the contribution.

Circularity Check

1 steps flagged · score 2.0 of 10

The numerical 'confirmation' of Eq. 2 is a self-consistency check against the same toy model from which the formula was derived, not an independent prediction.

other [Methods and Fig. 2(c) caption]
"Because it is exact, Eq. 2 will always agree with numerical evaluation of the Attention head in Fig. 1. Therefore we only give one example with specific parameter values in the main paper (Fig. 2(b))."

The 'empirical value' that Eq. 2 is said to predict is obtained by numerically evaluating the same Attention-head recurrence from which Eq. 2 was algebraically derived, as the paper itself states: Eq. 2 'will always agree' with that evaluation. Thus the agreement reported in Fig. 2(c) is guaranteed by construction; it is a self-consistency check, not an independent test of the model. Presenting this as 'exact agreement with the empirical value' makes the validation mathematically identical to its own input. No real-LLM or external data enter the comparison (Data Availability: 'The only data used in this study, is generated by the Mathematica notebooks'), so the claimed quantitative prediction for real LLM tipping points is not independently tested.

full rationale

Eq. 2 is an algebraic solution of the paper's own Attention-head recurrence (setting c_n·B = c_n·G), so the derivation itself is internally self-contained rather than circular. The paper's reliance on the authors' prior work [23] for the context-vector dynamics is a same-author citation, but this is a modeling lineage rather than a circular reduction: the tipping-point formula itself is newly derived here, and there is no evidence in the text that [23] already contains the claimed result. The one genuinely circular element is the validation narrative: Fig. 2(c) calls the simulation an 'empirical value' and Eq. 2's agreement with it a 'prediction', while Methods states the agreement is guaranteed because Eq. 2 is exact for the same Attention head. That makes the confirmation a mathematical identity, not a falsifiable prediction. There is no fitting of parameters to real-LLM data, and Data Availability confirms all data are generated by the supplied Mathematica notebooks, so the extrapolation from the K=Q=I single-head toy to real LLMs is a correctness and external-validity risk rather than a circularity; it does not raise the score above 2.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on a highly simplified attention model. The only free parameters are the hand-picked example vectors. The axioms are mostly domain assumptions about the transferability of a single-head toy model to real LLMs, plus standard mathematics.

free parameters (1)

Example token vectors G, B, and P_i = Not specified; used in Mathematica example
Hand-chosen values for the illustrative example that yields n*=10; the formula's parameters for real LLMs are not measured.

assumptions (4)

domain assumption A single self-attention head with Key/Query matrices set to identity reproduces the tipping behavior of real LLMs
The paper focuses on one head and notes multi-head/deep networks can be added later; no validation on real LLMs.
domain assumption The context vector c_n evolves as a linear accumulation of chosen token vectors with softmax weights summing to unity
Used in deriving Eq. 2; follows from the simplified attention setup but is not general.
domain assumption Tokens can be partitioned into fixed good (G) and bad (B) categories with constant embedding vectors
Lets the derivation treat G and B as fixed vectors; in real LLMs, embeddings are contextual and output classes are not fixed.
standard math Standard mathematical background of dot products and softmax is accepted
Used throughout the derivation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Jekyll-and-Hyde Tipping Point in an AI's Behavior." pith.science (2026). https://pith.science/paper/FJYPH4DB

@misc{pith2026250420980,
  author       = {Pith},
  title        = {Pith review of: Jekyll-and-Hyde Tipping Point in an AI's Behavior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJYPH4DB}},
  note         = {Machine review of arXiv:2504.20980}
}

read the original abstract

Trust in AI is undermined by the fact that there is no science that predicts -- or that can explain to the public -- when an LLM's output (e.g. ChatGPT) is likely to tip mid-response to become wrong, misleading, irrelevant or dangerous. With deaths and trauma already being blamed on LLMs, this uncertainty is even pushing people to treat their 'pet' LLM more politely to 'dissuade' it (or its future Artificial General Intelligence offspring) from suddenly turning on them. Here we address this acute need by deriving from first principles an exact formula for when a Jekyll-and-Hyde tipping point occurs at LLMs' most basic level. Requiring only secondary school mathematics, it shows the cause to be the AI's attention spreading so thin it suddenly snaps. This exact formula provides quantitative predictions for how the tipping-point can be delayed or prevented by changing the prompt and the AI's training. Tailored generalizations will provide policymakers and the public with a firm platform for discussing any of AI's broader uses and risks, e.g. as a personal counselor, medical advisor, decision-maker for when to use force in a conflict situation. It also meets the need for clear and transparent answers to questions like ''should I be polite to my LLM?''

Figures

Figures reproduced from arXiv: 2504.20980 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 16 canonical work pages

[23]

Huo, F. Y. & Johnson, N. F. Capturing AI’s Attention: Physics of Repetition, Hallucination, Bias and Beyond (2025). URL https://arxiv.org/abs/2504.04600. arXiv:2504.04600

arXiv 2025
[1]

Americans’ Top Feeling About AI in 2024 is Cau- tion (2024)

YouGov. Americans’ Top Feeling About AI in 2024 is Cau- tion (2024). URL https://today.yougov.com/technology/articles/ 49099-americans-2024-poll-ai-top-feeling-caution. Accessed: 2025-04-28. 6

work page 2024
[2]

Betley, J. et al. Emergent misalignment: Narrow finetuning can pro- duce broadly misaligned llms (2025). URL https://arxiv.org/abs/2502.17424. arXiv:2502.17424

arXiv 2025
[3]

Roose, K. Can A.I. Be Blamed for a Teen’s Suicide? (2024). URL https://www. nytimes.com/2024/10/23/technology/characterai-lawsuit-teen-suicide.html. Accessed: 2025-04-28

work page 2024
[4]

When the Person Abus- ing Your Child Is a Chatbot: The Tragic Story of Sewell Setzer (2024)

Center for Humane Technology. When the Person Abus- ing Your Child Is a Chatbot: The Tragic Story of Sewell Setzer (2024). URL https://www.humanetech.com/podcast/ when-the-person-abusing-your-child-is-a-chatbot-the-tragic-story-of-sewell-setzer. Accessed: 2025-04-28

work page 2024
[5]

Teaching Alexa and ChatGPT to Say Please and Thank You (2025)

Hill, K. Teaching Alexa and ChatGPT to Say Please and Thank You (2025). URL https://www.nytimes.com/2025/04/24/technology/ chatgpt-alexa-please-thank-you.html. Accessed: 2025-04-28

work page 2025
[6]

Teaching Alexa and ChatGPT to Say Please and Thank You (Com- ments Section) (2025)

Hill, K. Teaching Alexa and ChatGPT to Say Please and Thank You (Com- ments Section) (2025). URL https://www.nytimes.com/2025/04/24/technology/ chatgpt-alexa-please-thank-you.html#commentsContainer. Accessed: 2025-04- 28

work page 2025
[7]

Vaswani, A. et al. Attention is all you need (2023). URL https://arxiv.org/abs/ 1706.03762. arXiv:1706.03762

arXiv 2023

Show all 28 references

[8]

& Bengio, Y

Bahdanau, D., Cho, K. & Bengio, Y. Neural machine translation by jointly learning to align and translate (2016). URL https://arxiv.org/abs/1409.0473. arXiv:1409.0473

2016 arXiv
[9]

& Torroni, P

Galassi, A., Lippi, M. & Torroni, P. Attention in natural language processing. IEEE Transactions on Neural Networks and Learning Systems 32, 4291–4308 (2021)

2021
[10]

Ameisen, E. et al. Circuit tracing: Revealing computational graphs in language models (2025). URL https://transformer-circuits.pub/2025/attribution-graphs/ methods.html. Accessed: 2025-03-28

2025
[11]

Heaven, W. D. Anthropic can now track the bizarre inner workings of a large language model (2025). URL https://www.technologyreview.com/2025/03/27/1113916/ anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model. MIT Technology Review, Accessed March 28, 2025

2025
[12]

Tracing the thoughts of a large language model

Anthropic. Tracing the thoughts of a large language model. https:// www.anthropic.com/research/tracing-thoughts-language-model (2025). Accessed March 28, 2025. 7

2025
[13]

Lindsey, J. et al. On the biology of a large language model (2025). URL https:// transformer-circuits.pub/2025/attribution-graphs/biology.html. Accessed: 2025- 03-28

2025
[14]

Elhage, N., Henighan, T., Joseph, N. et al. A Mathematical Framework for Trans- former Circuits (2021). URL https://transformer-circuits.pub/2021/framework/ index.html. Interpretability Research at Anthropic

2021
[15]

& Steinhardt, J

Nanda, N., Chan, L., Lieberum, T., Smith, J. & Steinhardt, J. Progress mea- sures for grokking via mechanistic interpretability. International Conference on Learning Representations 2023 https://arxiv.org/pdf/2301.05217

2023 arXiv
[16]

& Lieberum, T

Nanda, N. & Lieberum, T. A mechanistic interpretability analysis of grokking. URL https://www.alignmentforum.org/posts/N6WM6hs7RQMKDhYjB/ a-mechanistic-interpretability-analysis-of-grokking. Accessed: 2024-05-07

2024
[17]

Paper replication walkthrough: Reverse-engineering mod- ular addition

Nanda, N. Paper replication walkthrough: Reverse-engineering mod- ular addition. https://www.neelnanda.io/mechanistic-interpretability/ modular-addition-walkthrough. Accessed: 2024-05-7

2024
[18]

Templeton, A. et al. Scaling monosemanticity: Extracting interpretable fea- tures from claude 3 sonnet (2024). URL https://transformer-circuits.pub/2024/ scaling-monosemanticity/index.html. Accessed: 2025-03-28

2024
[19]

Bricken, T. et al. Towards monosemanticity: Decomposing language models with dictionary learning (2023). URL https://transformer-circuits.pub/2023/ monosemantic-features/index.html. Accessed: 2025-03-28

2023
[20]

& Choi, Y

Holtzman, A., Buys, J., Du, L., Forbes, M. & Choi, Y. The curious case of neural text degeneration (2020). URL https://arxiv.org/abs/1904.09751. arXiv:1904.09751

2020 arXiv
[21]

Vijayakumar, A. K. et al. Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models (2016). URL https://arxiv.org/abs/1610.02424. ArXiv preprint arXiv:1610.02424

2016 arXiv
[22]

Strogatz, S. H. Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering (Chapman and Hall/CRC, 2024)

2024
[24]

& Levy, O

Haviv, A., Ram, O., Press, O., Izsak, P. & Levy, O. Transformer language models without positional encodings still learn positional information (2022). URL https: //arxiv.org/abs/2203.16634. arXiv:2203.16634. 8

2022 arXiv
[25]

McCoy, R. T. & Pavlick, E. Do Language Models Have Beliefs? Methods for Detecting, Updating, and Visualizing Model Beliefs (2022). URL https://arxiv. org/abs/2204.07143. ArXiv preprint arXiv:2204.07143

2022 arXiv
[26]

Kaplan, J., McCandlish, S., Henighan, T. et al. Scaling Laws for Neural Lan- guage Models (2020). URL https://arxiv.org/abs/2001.08361. ArXiv preprint arXiv:2001.08361

2020 arXiv
[27]

& Neubig, G

Michel, P., Levy, O. & Neubig, G. Are sixteen heads really better than one? (2019). URL https://arxiv.org/abs/1905.10650. arXiv:1905.10650

2019 arXiv
[28]

& Rumshisky, A

Rogers, A., Kovaleva, O. & Rumshisky, A. A primer in bertology: What we know about how bert works (2020). URL https://arxiv.org/abs/2002.12327. arXiv:2002.12327. 9

2020 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [23]

Huo, F. Y. & Johnson, N. F. Capturing AI’s Attention: Physics of Repetition, Hallucination, Bias and Beyond (2025). URL https://arxiv.org/abs/2504.04600. arXiv:2504.04600

arXiv 2025

[2] [1]

Americans’ Top Feeling About AI in 2024 is Cau- tion (2024)

YouGov. Americans’ Top Feeling About AI in 2024 is Cau- tion (2024). URL https://today.yougov.com/technology/articles/ 49099-americans-2024-poll-ai-top-feeling-caution. Accessed: 2025-04-28. 6

work page 2024

[3] [2]

Betley, J. et al. Emergent misalignment: Narrow finetuning can pro- duce broadly misaligned llms (2025). URL https://arxiv.org/abs/2502.17424. arXiv:2502.17424

arXiv 2025

[4] [3]

Roose, K. Can A.I. Be Blamed for a Teen’s Suicide? (2024). URL https://www. nytimes.com/2024/10/23/technology/characterai-lawsuit-teen-suicide.html. Accessed: 2025-04-28

work page 2024

[5] [4]

When the Person Abus- ing Your Child Is a Chatbot: The Tragic Story of Sewell Setzer (2024)

Center for Humane Technology. When the Person Abus- ing Your Child Is a Chatbot: The Tragic Story of Sewell Setzer (2024). URL https://www.humanetech.com/podcast/ when-the-person-abusing-your-child-is-a-chatbot-the-tragic-story-of-sewell-setzer. Accessed: 2025-04-28

work page 2024

[6] [5]

Teaching Alexa and ChatGPT to Say Please and Thank You (2025)

Hill, K. Teaching Alexa and ChatGPT to Say Please and Thank You (2025). URL https://www.nytimes.com/2025/04/24/technology/ chatgpt-alexa-please-thank-you.html. Accessed: 2025-04-28

work page 2025

[7] [6]

Teaching Alexa and ChatGPT to Say Please and Thank You (Com- ments Section) (2025)

Hill, K. Teaching Alexa and ChatGPT to Say Please and Thank You (Com- ments Section) (2025). URL https://www.nytimes.com/2025/04/24/technology/ chatgpt-alexa-please-thank-you.html#commentsContainer. Accessed: 2025-04- 28

work page 2025

[8] [7]

Vaswani, A. et al. Attention is all you need (2023). URL https://arxiv.org/abs/ 1706.03762. arXiv:1706.03762

arXiv 2023