Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

A single fitted law predicts when induction heads form during language-model pretraining, and it depends only on batch size and context size.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 21:01 UTC pith:KNUGQMHZ

load-bearing objection A useful empirical map of when induction heads form, but the headline law is an in-sample fit—the frequency/reliability Pareto frontier is the more solid contribution. the 4 major comments →

arxiv 2511.16893 v3 pith:KNUGQMHZ submitted 2025-11-21 cs.CL

Predicting the Emergence of Induction Heads in Language Model Pretraining

Tatsuya Aoyama , Ethan Gotlieb Wilcox , Nathan Schneider This is my paper
classification cs.CL
keywords induction headsphase transitionin-context learningpretrainingbatch sizecontext sizebigram repetitiontoken-weighted updates
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the moment a transformer develops induction heads—attention heads that copy an in-context pattern like A,B,...,A,B—is not a chaotic accident of training but obeys a simple rule: the number of updates to the phase transition equals a constant divided by the square root of the product of batch size and context size. The same fitted constant works across models trained with very different batch and context sizes, which the authors take as evidence that the formation point is agnostic to model size. A second claim is that two measurable properties of the training text—the frequency of repeated bigrams and how reliably the repetition completes—determine whether induction heads form at all, with a clear Pareto frontier separating success from failure. Finally, local (non-i.i.d.) dependency between consecutive tokens is necessary and, with enough bigram repetition, sufficient; category structure and marginal distribution shape only matter near the boundary.

Core claim

The central discovery is the fitted law UPT = T/√(BC): the emergence step is predicted from batch size B and context size C alone via one constant T. Because T = U√(BC) stays roughly constant across configurations, IH formation is governed by token-weighted updates; the equivalent token form NPT = T√(BC) predicts observed emergence points across two orders of magnitude. Two data knobs—frequency P(A,B,...,A) and reliability P(B|A,B,...,A)—form a Pareto frontier below which IHs fail to appear. In synthetic Markov chains, local dependency plus high frequency and reliability is sufficient; categoriality and marginal shape matter only near the frontier.

What carries the argument

The load-bearing object is the fitted power law UPT = T/√(BC), which collapses batch- and context-size effects into a single 'token-weighted updates' quantity T; this is what makes pre-training prediction of the phase transition possible. The supporting mechanism is the two-knob statistical characterization of training data—frequency, P(A,B,...,A), the rate at which a repeated bigram's first token reappears in context, and reliability, P(B|A,B,...,A), the conditional probability that the repetition completes—which trace the Pareto frontier below which induction heads fail to form.

Load-bearing premise

The law's key premise is that batch size and context size affect IH emergence time independently as inverse square roots—a claim inferred from two one-dimensional sweeps rather than tested by varying both together; if the two factors interact, the law fails.

What would settle it

Train the same 2-layer, 8-head transformer on a grid of batch and context sizes that includes combinations absent from the separate sweeps—e.g., B=64 with C=128, or B=4 with C=2048—and measure the step at which the prefix-matching score first crosses the random-attention threshold; if the emergence times do not lie on U = T/√(BC) for a single T, the separable law collapses. A second check: train the same model with 4 layers or 16 heads and see whether T stays constant; if it moves, the claim of model-size agnosticism is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the law holds, the IH emergence point for a given architecture can be computed before training starts from batch and context size alone, with no need to train and probe intermediate checkpoints.
  • Because the law is claimed to be agnostic to model size, predictions about phase-transition timing could transfer across model scales within the same architecture family.
  • The Pareto frontier gives a concrete data-curation target: push a corpus above the frequency-reliability boundary to encourage induction-head formation, or below it to suppress it.
  • Since raw update count and raw token count both fail as universal clocks, comparisons of phase transitions across runs should be reported in token-weighted updates U√(BC).
  • The sufficiency of local dependency plus high bigram repetition suggests that a minimal synthetic curriculum for inducing in-context learning can be built from a second-order Markov chain, without categories or Zipfian marginals.

Where Pith is reading between the lines

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

  • A natural next test is whether the inverse-square-root form is really separable: train on a grid of batch and context sizes not covered by the paper's one-dimensional sweeps and check whether the emergence time remains T/√(BC) or gains an interaction correction.
  • The model-size agnosticism claim rests partly on comparisons with other pretrained checkpoints; training the same 2-layer transformer at 4 or 12 layers would directly test whether the constant T is truly independent of depth and width.
  • The frequency-reliability frontier could be used to predict IH emergence in corpora from other domains or languages, rather than only the single English sample studied here.
  • The two knobs are defined over surface bigrams; extending them to higher-order n-gram dependencies may reveal whether the Pareto frontier shifts when longer repeated sequences are the relevant signal.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies when induction heads (IHs) emerge during language model pretraining, using a 2-layer, 8-head GPT2 trained on natural and synthetic data. The central claim is a simple law, UPT = T/√(BC), where UPT is the number of updates at which IHs form, B is batch size, C is context size, and T is a fitted constant. The authors report that this law predicts the emergence point with r = 0.98, and that it is agnostic to model size, citing Pythia models as additional evidence. The paper also explores how bigram repetition frequency and reliability affect IH formation, finding a Pareto frontier, and uses synthetic Markov processes to argue that local dependency is necessary and that categoriality and marginal distribution shape matter near the frontier.

Significance. If the scaling law UPT = T/√(BC) were validated, it would be a practically useful result: it would allow predicting phase transition timing in pretraining from only batch size and context size, independent of model size. The paper's systematic sweeps over batch size, context size, and repetition statistics are a strength, and the use of Pythia as an external model family is a promising design. However, the current evidence for the law is substantially weaker than reported: the correlation is computed on the same data used to fit the parameters, the functional form is underdetermined by the two one-dimensional sweeps, and the Pythia validation is not documented in sufficient detail. The contributions on the role of bigram frequency/reliability and the Pareto frontier are more solid and could stand independently, but the headline prediction claim needs re-validation.

major comments (4)
  1. [Appendix D, Fig. 6] The reported prediction correlation r = 0.98 is an in-sample goodness-of-fit. The parameters α, β, γ in Eq. (2) are estimated by OLS from the same GPT2 training runs that are then plotted as 'predicted vs observed' in Fig. 6. This is circular. The only potentially external evidence, the Pythia points, is not separated in the analysis: the paper never states whether Pythia points were held out from the fit, how their emergence times were measured, or what the per-point residuals are. Please report a proper hold-out test (e.g., fit on GPT2 only and evaluate on Pythia, or a train/test split across configurations) and give the correlation for held-out points separately.
  2. [Section 3.1, Eq. (1)] The joint dependence UPT = T/√(BC) is underdetermined by the experiments. Batch size is varied only with C=1024, and context size only with B=16. If the true relation is UPT = T·B^{-a}·C^{-b}·(B/C)^δ, then the B-sweep slope is -a+δ and the C-sweep slope is -b-δ. The observed slopes (-0.51, -0.56) are compatible with many (a,b,δ), including a≠b and δ≠0. A joint sweep with at least one off-diagonal (B,C) pair is needed to identify the separable form. Pythia does not resolve this because all Pythia models share the same B and C.
  3. [Section 2.2, Fig. 6] The 'agnostic to model size' claim is not adequately tested. All in-house models are the same 2-layer, 8-head GPT2. The only cross-size evidence is the Pythia points in Fig. 6, but the manuscript does not describe how IHs were identified in Pythia (checkpoint schedule, PS threshold, PWLF fit), nor whether these points were included in the regression. Without this, the claim that T is constant across model size is unsupported. Please specify the Pythia analysis explicitly, and if possible, train the same architecture at different widths/depths to test T invariance.
  4. [Section 3.1, Appendix B] The law is fit only on configurations where IHs actually form (e.g., context sizes ≥32). This is reasonable, but the paper should state this limitation clearly: the equation predicts the timing only for configurations above the formation threshold; it does not predict whether IHs will form at all. The current phrasing in the abstract and conclusion ('predicts the point at which IHs form') may overstate the scope.
minor comments (5)
  1. [Appendix A] Typo: 'aganin' should be 'again'.
  2. [Table 1] Typo: 'Inducion' should be 'Induction'.
  3. [Appendix D] Typo: 'piece-wise liner' should be 'piece-wise linear'.
  4. [Section 4.2] The grid search is described as {0.1,0.3,0.5,0.7,0.8}; later the text mentions a second search over {0.01,0.03,0.05,0.07,0.09}, giving a total of 50 models. The relation between these sets and the axes in Fig. 2 could be clearer.
  5. [Appendix D] The constant T is said to be e^α = 750000, but α = 13.5 gives e^13.5 ≈ 7.29×10^5. Also, after approximating β and γ to -0.5, it is unclear whether T is re-estimated or taken directly from the intercept. Please clarify the exact fitting procedure.

Circularity Check

2 steps flagged

The B–C emergence law is fit to the same UPT values that are then reported as 'predicted vs. actual' (r=0.98); the NPT 'verification' is an algebraic rewrite of the same fitted relation, leaving the central prediction in-sample.

specific steps
  1. fitted input called prediction [Section 3.1, Eq. (1); Appendix D, Eqs. (2)-(4) and Figure 6]
    "Given these observations, we fit a regression model that predicts the IH formation point in training steps and obtain a fitted linear model: UPT = T/√(BC) (1) ... We find a strong correlation between the predicted and actual emergence points (r=0.98, p < .0001). ... Fitting this model, we obtain α=13.5, β=−0.51, and γ=−0.56. We approximate these two parameters as β=γ=−0.5 ... let us call e^α = 750000 a constant T."

    The UPT values used to estimate α, β, and γ by OLS in log space are the same values that appear on the 'actual' axis when computing r=0.98. The reported correlation is therefore an in-sample goodness-of-fit, not an out-of-sample prediction. Because the two sweeps are one-dimensional (B varied at C=1024; C varied at B=16), the OLS fit is essentially interpolating those marginal sweeps, so it cannot by itself identify the separable √(BC) form or validate the fitted T. The Pythia points in Figure 6 could provide an external check, but the paper never states whether they were held out from the fit or how their emergence times were measured.

  2. self definitional [Appendix D, Eqs. (5)-(6)]
    "To further verify that this simple law indeed predicts the phase transition point of LMs trained with various training configurations, we can reformulate Equation (4) to predict the number of tokens N: T=UPT√BC; T√BC=UPT BC (5). Because N=U BC by definition, RHS is N, and we get: NPT = T√BC (6)."

    This is presented as further verification, but it is a definitional rewrite, not an independent test. T was defined as UPT√BC and N is defined as UPT·BC, so NPT = T√BC is exactly the quantity from which T was fitted, transformed by the identity N = U·B·C. Consequently, Figure 6's 'predicted vs. observed tokens' plot is the same in-sample fit redrawn in token units, and its correlation cannot provide independent confirmation of the law.

full rationale

The central circularity is in Section 3.1 / Appendix D: the paper fits log UPT as a linear function of log B and log C, rounds the exponents to −0.5, defines T from the fitted intercept, and then reports r=0.98 between 'predicted' and 'actual' emergence points. Since the same configurations supplied both the fit and the correlation, this is an in-sample goodness-of-fit, and the 'prediction' claim reduces to the fitted regression. The subsequent NPT = T√BC 'verification' is algebraically equivalent to N = UPT·BC with T = UPT√BC, so it adds no independent evidence. The frequency/reliability and synthetic-data experiments are more self-contained and are not circular in the same way, though their 'sufficient for IH formation' conclusions rest on single seeds and a single 2-layer/8-head architecture. The Pythia points in Figure 6 could have supplied external validation, but the paper does not explicitly state that they were held out from the fit, so they cannot be credited as an independent test. The self-citation to Aoyama & Wilcox [1] is present but is mainly used for prior phase-transition measurements rather than as the sole justification of the B–C law; that weakens the model-size-agnostic claim but is not itself the main circular step. Overall, the load-bearing 'prediction' is fitted input called prediction, giving a partial circularity score of 6.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central law rests on fitted exponents and a fitted constant, plus domain assumptions about the architecture and PS metric. The frequency/reliability metrics are new constructs but are defined directly from data, not postulated entities. No new physical or theoretical entities are introduced.

free parameters (6)
  • alpha (intercept in log regression) = 13.5 (estimated via OLS in log space)
    Fitted to the GPT2 training runs in Appendix D; determines the constant T in the central law.
  • beta (batch-size exponent) = -0.51, approximated to -0.5
    Fitted via OLS to the same data, then rounded for interpretability; final law uses -0.5.
  • gamma (context-size exponent) = -0.56, approximated to -0.5
    Fitted via OLS to the same data, then rounded for interpretability; final law uses -0.5.
  • T (law constant) = ≈750,000 (e^α); also written as 10^5.7
    Derived from fitted α; the central equation UPT = T/√(BC) depends on this fitted constant.
  • Context size 64 in Experiments 2 and 3 = 64
    Hand-chosen in Appendix C as balancing enough context and a good number of chunks with/without bigram repetitions; affects the frequency/reliability Pareto frontier and synthetic results.
  • Category similarity targets and entropy target = within-category 0.4, across-category 0.1; entropy 6.2
    Hand-set in Appendix H to construct synthetic Markov matrices for Experiment 3; influence the categoriality and entropy comparisons.
axioms (4)
  • domain assumption Prefix-matching score (PS) correctly identifies induction heads
    Section 2.1 adopts the PS metric from Olsson et al. without independent validation that PS corresponds to functionally relevant induction heads in this setting.
  • domain assumption The 2-layer, 8-head GPT2 architecture is representative of larger LMs for IH emergence
    All primary experiments use this single architecture; the claim of model-size agnosticism is largely inherited from prior work or the Pythia follow-up, not tested with own models of different sizes.
  • domain assumption Natural language is adequately approximated by a first-order (bigram) Markov chain for Experiments 2 and 3
    Section 4.1 builds a token-to-token transition matrix from CC100 bigram statistics; higher-order dependencies are ignored, which could affect bigram repetition statistics.
  • domain assumption The 3-segment piecewise-linear fit's knot reliably estimates the emergence point
    Appendix D uses a PWLF with 3 segments to define the phase transition point; the segment specification and knot selection are modeling choices that directly determine the fitted values.

pith-pipeline@v1.3.0-alltime-deepseek · 14260 in / 15203 out tokens · 125553 ms · 2026-08-03T21:01:40.456180+00:00 · methodology

0 comments
read the original abstract

Specialized attention heads dubbed induction heads (IHs) have been argued to underlie the remarkable in-context learning capabilities of modern language models; yet, a precise characterization of their emergence, especially in the context of language modeling, remains wanting. In this study, we investigate the relationship between statistical properties of the training data and IH formation in both natural and synthetic training data settings. We show that: (1) a simple equation combining batch size and context size predicts the point at which IHs form and that this emergence point is agnostic to model size; (2) surface bigram repetition frequency and reliability strongly affect the formation of IHs, and we find an effective decision boundary in terms of these two values; (3) local dependency with high bigram repetition frequency and reliability is sufficient for IH formation, but categoriality and the shape of the marginal distribution appear to modulate IH formation near the decision boundary.

Figures

Figures reproduced from arXiv: 2511.16893 by Ethan Gotlieb Wilcox, Nathan Schneider, Tatsuya Aoyama.

Figure 1
Figure 1. Figure 1: Developmental trajectories of PS of LMs with various batch sizes (left), context sizes (center), and repetitions (right) over the course of 1B tokens of pretraining. BS, CS, %NR stands for batch size, context size, and the proportion of chunks with no repetitions, respectively. number of updates, and (2) the smaller the context size, the flatter the slope once IHs start forming, and the extreme case (conte… view at source ↗
Figure 2
Figure 2. Figure 2: Best PS across all heads at the end of the training for each frequency reliability combination. Scores are represented in colors, with brighter colors representing higher scores. P(AB...A) P(B|AB...A) 10-10 10-30 90-90 Zipf[+D+C] ✗ ✓ ✓ Zipf[+D−C] ✗ ✗ ✓ Zipf[−D−C] ✗ ✗ ✗ Unif[+D+C] ✗ ✗ ✓ Unif[+D−C] ✗ ✗ ✓ Unif[−D−C] ✗ ✗ ✗ Gaus[+D+C] ✗ ✗ ✓ Gaus[+D−C] ✗ ✗ ✓ Gaus[−D−C] ✗ ✗ ✗ [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 3
Figure 3. Figure 3: Developmental trajectories of PS of LMs with various batch sizes (left), context sizes (center), and repetitions (right) over the course of 1B tokens of pretraining. BS, CS, %NR stands for batch size, context size, and the proportion of chunks with no repetitions, respectively. A [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Smoothed distribution of chunks with various numbers of bigram repetitions. Context sizes 1024 and 2048 were rendered invisible, and hence removed from the plot. The plot is truncated at y=0.6 for readability, but context sizes of 4, 8, and 16 had >95% of chunks with no bigram repetitions. D Predicting the emergence points We have seen that context size and batch size affect the phase transition point (Sec… view at source ↗
Figure 5
Figure 5. Figure 5: Development of the highest PS score across all heads at each checkpoint plotted against the number of updates (left column) and TWUs (right column). Different line colors represent different batch sizes (top row) and context sizes (bottom row). All four plots share the scales of the x/y-axes. The x-axis is in log-scale, since both batch and context sizes increase exponentially, and hence the number of upda… view at source ↗
Figure 6
Figure 6. Figure 6: Predicted and observed number of pretraining tokens at which phase transition occurs. x-axis and y-axis represent predicted and observed points of phase transition, expressed in the number of pretraining tokens, respectively. A strong correlation of r = .98 (p < .001) is found. E Frequency P(A, B, . . . , A) and Reliability P(B | A, B, . . . , A) Recall that Elhage et al. [7], Olsson et al. [11] define IHs… view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of chunks with various P(A,B, . . . , A) and P(B | A, B, . . . ,A) for each context size. Only the quartile box, median (center line in each box), mean (diamond), and whiskers are shown, and Outliers are not shown for readability. of these two knobs. To this end, we approximate natural language by first tokenizing texts from the English subcorpus of the Common Crawl Corpus (CC100; 5, 13) and c… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence

    stat.ML 2026-06 unverdicted novelty 7.0

    Bayesian reduction of attention posterior on copy task predicts first-order phase transition for softmax attention and second-order followed by crossover for linear attention.

  2. The Kinetics of Training: A Driven-Nucleation Rate Law for Emergence, Plasticity Loss, and Circuit Control in Language Models

    cs.LG 2026-07 conditional novelty 6.0

    Capability formation in small transformers is claimed to obey a driven-nucleation rate law J = Nνσ(c)e^{−βK} − D, read forward as emergence, backward as plasticity loss, and completed as circuit control.

  3. Features have life history. And we should care

    q-bio.NC 2026-05 unverdicted novelty 5.0

    Language model features form an early stable carrier scaffold of about 50 sparse features that is load-bearing, predictable from onset firing, and recruits most later features.

Reference graph

Works this paper leans on

19 extracted references · 3 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Language models grow less humanlike beyond phase transition

    Tatsuya Aoyama and Ethan Wilcox. Language models grow less humanlike beyond phase transition. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors,Proceedings of the 63rd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 24938–24958, Vienna, Austria, July 2025. Association f...

    2025

  2. [2]

    Pythia: A suite for analyzing large language models across training and scaling.arXiv preprint arXiv:2304.01373, 2023

    Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar van der Wal. Pythia: A suite for analyzing large language models across training and scaling.arXiv preprint arXiv:2304.01373, 2023. URL https:...

    Pith/arXiv arXiv 2023

  3. [3]

    Chan, Adam Santoro, Andrew Kyle Lampinen, Jane X Wang, Aaditya K Singh, Pierre Harvey Richemond, James McClelland, and Felix Hill

    Stephanie C.Y . Chan, Adam Santoro, Andrew Kyle Lampinen, Jane X Wang, Aaditya K Singh, Pierre Harvey Richemond, James McClelland, and Felix Hill. Data distributional properties drive emergent in-context learning in transformers. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems,

  4. [4]

    Sudden drops in the loss: Syntax acquisition, phase transitions, and simplicity bias in MLMs

    Angelica Chen, Ravid Shwartz-Ziv, Kyunghyun Cho, Matthew L Leavitt, and Naomi Saphra. Sudden drops in the loss: Syntax acquisition, phase transitions, and simplicity bias in MLMs. InThe Twelfth International Conference on Learning Representations, 2024. URL https: //openreview.net/forum?id=MO5PiKHELW

    2024

  5. [5]

    Unsupervised cross-lingual representation learning at scale

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wen- zek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoy- anov. Unsupervised cross-lingual representation learning at scale. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors,Proceedings of the 58th Annual Meet- ing of the...

    doi:10.18653/v1/202 2020

  6. [6]

    Edelman, eran malach, and Surbhi Goel

    Ezra Edelman, Nikolaos Tsilivis, Benjamin L. Edelman, eran malach, and Surbhi Goel. The evolution of statistical induction heads: In-context learning markov chains. InThe Thirty- eighth Annual Conference on Neural Information Processing Systems, 2024. URL https: //openreview.net/forum?id=qaRT6QTIqJ

    2024

  7. [7]

    A mathematical framework for transformer circuits.Transformer Circuits Thread,

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah. A...

  8. [8]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020. URL https://arxiv.org/abs/2001.08361

    Pith/arXiv arXiv 2001

  9. [9]

    Transformerlens

    Neel Nanda and Joseph Bloom. Transformerlens. https://github.com/ TransformerLensOrg/TransformerLens, 2022

    2022

  10. [10]

    Byung-Doh Oh and William Schuler. Why does surprisal from larger transformer-based language models provide a poorer fit to human reading times?Transactions of the Association for Computational Linguistics, 11:336–350, 2023. doi: 10.1162/tacl_a_00548. URL https: //aclanthology.org/2023.tacl-1.20/

    doi:10.1162/tacl_a_00548 2023

  11. [11]

    In-context learning and induction heads

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, a...

    2022

  12. [12]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8): 9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8): 9, 2019. URL https://cdn.openai.com/better-language-models/language_models_ are_unsupervised_multitask_learners.pdf. 6

    2019

  13. [13]

    CCNet: Extracting high quality monolingual datasets from web crawl data

    Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. CCNet: Extracting high quality monolingual datasets from web crawl data. In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Hitoshi Isahara, Bente Maegaard, Jos...

    2020

  14. [14]

    An explanation of in-context learning as implicit bayesian inference

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit bayesian inference. InInternational Conference on Learning Representations, 2022. URLhttps://openreview.net/forum?id=RdJVFCHjUMI

    2022

  15. [15]

    above random,

    Kayo Yin and Jacob Steinhardt. Which attention heads matter for in-context learning?arXiv preprint arXiv:2502.14010, 2025. URLhttps://arxiv.org/abs/2502.14010. 7 104 105 106 107 108 109 10100.0 0.2 0.4 0.6 0.8 Best PS BS 4 8 16 32 64 128 256 512 104 105 106 107 108 109 1010 # Tokens 0.0 0.2 0.4 0.6 0.8 CS 4 8 16 32 64 128 256 512 1024 2048 104 105 106 107...

    Pith/arXiv arXiv 2025

  16. [18]

    ,A⟩sequence, 2 of them are followed by B, hence 2 4 = 1

    It might make more intuitive sense to compute this directly without using the chain rule: of all the 4 tokens that complete the ⟨A,B, . . . ,A⟩sequence, 2 of them are followed by B, hence 2 4 = 1

  17. [19]

    frequency

    Equivalently, of the tokens where RU = 1in Table 2, half of them also have RB = 1. However, for the computational purpose, the chain rule is much simpler, which is the reason we introduced the chain rule based calculation above. In this study, “frequency” and P(A,B, . . . ,A)are used interchangeably, and so are “reliability” andP(B|A,B, . . . ,A). Note th...

    1988

  18. [2021]

    URLhttps://transformer-circuits.pub/2021/framework/index.html

    2021

  19. [2022]

    URLhttps://openreview.net/forum?id=lHj-q9BSRjF