Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

The paper argues that a training sample's influence on a model is not a fixed property: it can change sign and spike sharply at developmental phase transitions, so data attribution must be stagewise.

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

2026-08-04 09:59 UTC pith:C5VKRPZL

load-bearing objection Real conceptual contribution and a solid toy model, but the BIF experiments do not actually test the paper's phase-transition mechanism—the local sampler cannot see the between-phase term. the 3 major comments →

arxiv 2510.12071 v2 pith:C5VKRPZL submitted 2025-10-14 cs.LG

Influence Dynamics and Stagewise Data Attribution

Jin Hwa Lee , Matthew Smith , Maxwell Adam , Jesse Hoogland This is my paper
classification cs.LG
keywords training data attributioninfluence functionssingular learning theorystagewise developmentphase transitionsBayesian influence functionsusceptibilitylanguage model development
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.

Current training-data attribution treats the influence one sample has on another as a single static score, but neural networks learn in distinct stages and the paper argues this picture is fundamentally incomplete. Drawing on singular learning theory, it predicts that influence is dynamic—non-monotonic, capable of sign flips, and sharply peaking at the phase transitions that define stagewise learning. The paper validates this prediction analytically and empirically in a toy hierarchical model, where influence peaks align with the model's progressive acquisition of semantic distinctions, and in language models, where token-level influence changes align with known developmental events. If correct, the field should move from pointwise to stagewise data attribution: the question is not only which samples matter, but when and why they matter during training.

Core claim

The central claim is that the Bayesian influence function (BIF)—the derivative of a posterior expectation of an observable with respect to a sample's training weight—serves as a generalized susceptibility that registers stagewise phase transitions. Decomposing the BIF by the law of total covariance into within-phase and between-phase contributions shows that the between-phase term, πUπV(μi,U−μi,V)(μj,U−μj,V), peaks when the posterior is evenly split across two phases and is largest for samples on which the phases disagree most. This yields two predictions: influence can flip sign, and influence peaks at transitions. The paper confirms these predictions with an analytical derivation in a deep

What carries the argument

The Bayesian Influence Function (BIF), defined as the negative posterior covariance between a sample's loss and an observable, replaces the Hessian-inverse classical influence function and stays well-defined even on degenerate loss landscapes. Its role is to make influence a distributional, checkpoint-local quantity, and its decomposition via conditioning on the posterior's phase yields the between-phase covariance term—the mechanism that produces sign flips and sharp peaks. It is estimated in practice with a stochastic-gradient Markov-chain Monte Carlo sampler localized to each checkpoint.

Load-bearing premise

The load-bearing premise is that SGD-trained neural networks can be modeled as a Bayesian learning process whose posterior undergoes the phase transitions predicted by singular learning theory—a correspondence the paper's own limitations section describes as the primary theoretical gap.

What would settle it

Train a neural network in a regime with no degeneracy and no expected phase transition (for example, a well-conditioned model where the Bernstein–von Mises theorem should hold) and measure BIF over training: if influence still shows sharp peaks and sign flips, the phase-transition mechanism is not what drives them; if influence stays smooth and monotonic, the mechanism is supported.

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

If this is right

  • Static end-of-training attribution scores can be actively misleading: a sample that helps early may harm later, and vice versa.
  • Sharp peaks in Bayesian influence localize phase transitions and single out the samples that characterize them.
  • Token-level influence in language models is observable and non-monotonic, with timing that lines up with known developmental milestones such as induction-circuit formation.
  • Ablation timing matters for interveners: retraining experiments show that briefly removing a sample at its peak-influence stage produces the largest loss change.
  • The framework calls for stagewise data attribution as a new default object of study, with influence as a trajectory rather than a point estimate.

Where Pith is reading between the lines

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

  • If the picture holds, attribution benchmarks should stop reporting single final scores and instead score methods on whether they reproduce influence trajectories, sign flips, and transition timing.
  • The between-phase term offers a way to cluster samples by which phase they favor, turning influence spike patterns into a data-structure discovery tool instead of just a debugger.
  • Peaks in influence could be used as an online, training-time detector of phase transitions—and, if transitions are steerable, as a hook for intervening in model development.
  • The framework gives a mechanistic gloss on implicit curricula: rather than being imposed, the curriculum emerges because different data become maximally influential at different stages, which might explain why explicit curricula only sometimes help.

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

3 major / 5 minor

Summary. The paper argues that training data influence is not static but evolves during learning, and proposes a stagewise data attribution framework grounded in singular learning theory (SLT). The theoretical core (Section 2.3) models a first-order phase transition as a two-phase posterior mixture, decomposes the Bayesian influence function (BIF) into within-phase and between-phase covariance terms, and predicts that influence can change sign and peak sharply when the posterior mass is split evenly between phases. Empirical support is sought in two settings: a deep linear network on a hierarchical dataset (Section 3), where BIF trajectories are compared with an analytic influence computation and leave-one-out retraining; and Pythia language models (Section 4), where per-token BIFs across structural token classes are claimed to show non-monotonic dynamics, including sign flips and an induction-related peak. The paper concludes that data attribution methods should track influence as a function of training time rather than returning a single static score.

Significance. If the central claim holds, the paper challenges a foundational assumption of most TDA methods and offers a new diagnostic—BIF as a generalized susceptibility—for detecting developmental transitions in neural networks. The strongest part of the paper is the toy model: the analytic derivation in Section C.6 and the leave-one-out retraining experiments are independent of the BIF estimator and agree qualitatively with the BIF trajectory, providing credible evidence that influence can vary non-monotonically and that peaks align with hierarchy-learning stages. The paper is also transparent about its main gap, the SGD-to-Bayes correspondence, in the Limitations. However, the specific theoretical mechanism advanced in Section 2.3—that peaks arise from the between-phase covariance of a bimodal posterior—is not actually probed by the local BIF experiments, because the estimator is deliberately localized around a single checkpoint. The LM results are suggestive but share this limitation and additionally rely on a normalized BIF whose sign interpretation differs from the theoretical covariance. The empirical phenomenon is therefore more secure than the proposed causal explanation.

major comments (3)
  1. [§2.3, §B (Eq. 15, Algorithm 1)] The central prediction in Section 2.3 attributes sign flips and peaks to the between-phase term πUπV(μi,U−μi,V)(μj,U−μj,V), which requires the posterior p(w|D) to be bimodal with πU≈πV≈0.5. But the local BIF estimator in Eq. (15) and Algorithm 1 is an RMSProp-SGLD sampler initialized at checkpoint w* with a strong localization drift γ(w−w*). Such a sampler stays in a single basin and estimates a unimodal posterior around w*; it structurally cannot estimate the between-phase covariance. Consequently, the BIF peaks in Figures 3–5 and 15 measure only within-phase covariances and do not test the Section 2.3 mechanism. The independent analytic IF and LOO results support non-monotonic influence, but not the specific claim that peaks are susceptibility divergences at phase transitions. Please either revise the causal claims to what the local BIF can actually measure, or add an experiment that c
  2. [§3, §C.2] The toy-model validation is weakened by the hyperparameter selection procedure. Section C.2 states that the BIF hyperparameters shown in Figure 3 (β=1000, ε=1e-3, γ=5e+3) were chosen as the point with the highest correlation between the BIF trace and the LOO loss-difference trace. This makes the reported BIF–LOO agreement in Figure 3 an in-sample fit, not a blind confirmation. The grid in Figure 6 shows that the correlation varies smoothly and substantially with the hyperparameters. Please report the BIF–LOO agreement for a hyperparameter set chosen by an independent criterion (e.g., the LLC-calibration procedure described in C.2) or show the distribution of BIF–LOO correlations over the whole grid, and justify that the chosen point is not cherry-picked.
  3. [§B, §4, Fig. 5] The language-model experiments use the normalized BIF, defined as the Pearson correlation between losses (Section B), whereas the theoretical predictions in Section 2.3 concern the covariance in Eq. (2). A sign flip in Pearson correlation is not equivalent to a sign flip in covariance, since changes in the marginal variances of the losses can induce sign changes in the correlation even when the covariance does not change sign. Claims in Section 4 about “sign flips” (e.g., the left/right delimiter relationship) and “negative influence” are therefore not directly supported by the theory. Please either report the unnormalized BIF for the sign-flip claims, or explicitly analyze how standardization affects the signs and peaks, or rephrase the claims in terms of normalized BIF and state that the theoretical sign-flip prediction is not being tested at LM scale.
minor comments (5)
  1. [§A.1, Eqs. (7)–(8)] There is a sign inconsistency. For Fβn = −log∫e^{−β·ℓ}φ dμ, one has ∂F/∂βi = −E[ℓi], so E[ℓi] = −∂F/∂βi and BIF = ∂E[ℓj]/∂βi = −∂²F/∂βi∂βj. Equations (7)–(8) omit this minus sign and therefore define the BIF with the opposite sign, contradicting Eq. (2). Please correct the derivation or the sign convention.
  2. [§A.2, Eq. (14)] The text says the log-sum-exp is a smooth approximation for the max function, but Eq. (14) shows Fn ≈ mini Fn(Wi). It should say min, not max.
  3. [§B] The sentence “The full algorithm is described in Section B” is self-referential; it should reference Algorithm 1, where the RMSPropSGLD procedure is defined.
  4. [§C.1] Minor typo: “N∼(0, σ2)” should read “N(0, σ²)” or “𝒩(0, σ²)”. Also, the main text does not state which values of C and T from Table 1 were used for the BIF plots in Figure 3; please report them.
  5. [§C.2 and Fig. 3 caption] The Figure 3 caption lists the BIF hyperparameters but does not say that they were selected to maximize BIF–LOO correlation. That fact is in Section C.2 but should be stated prominently wherever the agreement is presented as validation.

Circularity Check

1 steps flagged

Theoretical peak/sign-flip predictions follow algebraically from the assumed bimodal posterior, but independent toy/LM measurements keep the central phenomenon grounded.

specific steps
  1. self definitional [Section 2.3, 'Decomposing influence with the law of total covariance' and 'Predicting stagewise changes in influence']
    "A first-order phase transition is characterized by the posterior distribution p(w| D) having significant mass in two distinct neighborhoods, which we label U and V. ... At the peak of a phase transition, πU ≈π V ≈0.5. ... The between-phase influence term is maximized when the posterior mass is evenly split (πU ≈π V ≈0.5 ), causing a sharp peak in total influence at the critical point of a transition."

    The 'prediction' is obtained by inserting the paper's own definition of a phase-transition peak (posterior mass evenly split, πU≈πV≈0.5) into the law of total covariance applied to the BIF definition (Eq. 2). The between-phase term πUπV(µi,U−µi,V)(µj,U−µj,V) is maximized at πU=πV by construction, so 'influence peaks at transitions' is a property of the assumed mixture model rather than an independent constraint. Sign flips are likewise a direct algebraic consequence of phase-dependent within-phase covariances. This is partial built-in-ness of the theoretical core, not a fitting of the empirical data.

full rationale

Section 2.3's central prediction is a formal consequence of the model setup: once the posterior is written as a two-phase mixture with πU+πV=1, the law of total covariance forces the between-phase contribution πUπV(µi,U−µi,V)(µj,U−µj,V), and this quadratic term is maximal at πU=πV. Because the paper defines the peak of a phase transition as πU≈πV≈0.5, the predicted 'sharp peak in total influence at the critical point' is largely built into the definitions. The same holds for sign flips if within-phase covariances differ. This is the one genuinely definitional element. It does not make the whole paper circular: the toy-model phenomenon is also established by leave-one-out retraining (Eq. 3) and by an analytic perturbation calculation in Section C.6 that does not use the mixture decomposition; the language-model results are compared to external developmental markers (induction-head timing from Olsson et al. 2022; Tigges et al. 2024). A separate validity concern, not a circularity, is that the implemented local BIF (Eq. 15, Algorithm 1) samples a posterior localized to w* and therefore cannot cross free-energy barriers, so the between-phase covariance term is not directly measured by the BIF experiments; this limits how strongly Figures 3–5 and 15 test the Section 2.3 mechanism. In addition, the main toy BIF plot used hyperparameters selected for the highest BIF–LOO correlation (Section C.2), so that particular agreement is partly selection-optimized; the analytic IF and LOO remain independent checks. Self-citations (Kreer et al. 2025, Hoogland et al. 2024/2025, Baker et al. 2025, Lehalleur et al. 2025) supply methods and framing, but the load-bearing SLT results are cited to Watanabe (2009, 2018), not to a self-citation chain, and no uniqueness theorem is imported. Overall, partial definitional circularity in the theoretical core, with independent empirical grounding, warrants a 4.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The theoretical framework relies on standard statistical identities and on SLT results from the literature, plus the paper-specific assumption of a two-phase posterior. The toy model adds a domain assumption about alignment of singular vectors. No new physical entities are introduced; the main free parameters are BIF hyperparameters that were tuned against the validation signal.

free parameters (6)
  • BIF inverse temperature β (toy) = 1000
    Selected as the value maximizing BIF-LOO correlation in Figure 6 (Section C.2).
  • BIF localization strength γ (toy) = 5e+3
    Selected as the value maximizing BIF-LOO correlation in Figure 6 (Section C.2).
  • BIF step size ε (toy) = 1e-3
    Selected as the value maximizing BIF-LOO correlation in Figure 6 (Section C.2).
  • Perturbation strength ε (analytic toy) = -0.1
    Used to approximate the ablation effect in the analytic influence derivation (Section C.6).
  • BIF localization strength γ (LM) = 500 or 1000
    Selected based on KNN class recall in Section D.1.2.
  • BIF inverse temperature β (LM) = 256 or 1024
    Selected based on KNN class recall in Section D.1.2.
axioms (6)
  • standard math Law of total covariance
    Used in Section 2.3 to decompose the BIF into within-phase and between-phase terms.
  • standard math Watanabe's singular learning theory free energy formula
    Used in Section 2.1 and Appendix A to justify phase transitions via competition between loss and complexity.
  • ad hoc to paper Two-phase mixture posterior at phase transitions
    Section 2.3 assumes the posterior is p(w|D)=πU p(w|U)+πV p(w|V); the peak prediction follows from this modeling assumption.
  • domain assumption Deep linear network singular value dynamics with balanced initialization
    Appendix C.6 assumes W=U G V^T with closed-form G(t), which holds under small initialization and small learning rate.
  • domain assumption SGD trajectory approximated by a Bayesian posterior
    Section 2.1 follows the Developmental Interpretability recipe; the paper acknowledges this is an open gap in the Limitations.
  • domain assumption RMSPropSGLD samples approximate the true local posterior
    Section B relies on SGLD estimates of the BIF as covariance over the posterior.

pith-pipeline@v1.3.0-alltime-deepseek · 20697 in / 11817 out tokens · 97891 ms · 2026-08-04T09:59:46.057321+00:00 · methodology

0 comments
read the original abstract

Current training data attribution (TDA) methods treat the influence one sample has on another as static, but neural networks learn in distinct stages that exhibit changing patterns of influence. In this work, we introduce a framework for stagewise data attribution grounded in singular learning theory. We predict that influence can change non-monotonically, including sign flips and sharp peaks at developmental transitions. We first validate these predictions analytically and empirically in a toy model, showing that dynamic shifts in influence directly map to the model's progressive learning of a semantic hierarchy. Finally, we demonstrate these phenomena at scale in language models, where token-level influence changes align with known developmental stages.

Figures

Figures reproduced from arXiv: 2510.12071 by Jesse Hoogland, Jin Hwa Lee, Matthew Smith, Maxwell Adam.

Figure 1
Figure 1. Figure 1: Stagewise learning requires stagewise data attribution. (a) In regular models, development is a uniform, monotonic process of posterior concentration around a single solution (Bernstein–von Mises). (b) In singular models, development is a stagewise process where the posterior undergoes phase transitions (Watanabe’s singular learning theory). This stagewise development means the influence one sample has on … view at source ↗
Figure 2
Figure 2. Figure 2: A toy model of hierarchical semantic knowledge acquisition. (a) A toy dataset adopted from Saxe et al. (2019a). Each object maps to a feature vector that describes the hierarchical structure of semantic knowledge, e.g., ‘penguin’ and cat are all living animals, but ‘penguin’ is a bird and cat is a mammal. (b) The correlation matrix of the feature output shows a hierarchical structure. (c) The hierarchical … view at source ↗
Figure 3
Figure 3. Figure 3: Influence over time on a hierarchical semantic dataset. We measure the influence of dog sample on other query samples j with the following: (Left) BIF (β = 1000, ϵ = 1e-3, γ = 5e + 3). (Center) Analytical IF (see full derivation in Section C.6). (Right) Loss difference from Leave-One-Out (LOO) retraining experiment. All three measures agree that the influence one sample has on another can vary non-monotoni… view at source ↗
Figure 4
Figure 4. Figure 4: BIF captures developmental influence. (a) The peak positive influence from dog to different data points is noted with ⃝1 , ⃝2 , and ⃝3 . (b) MDS of the hidden representations in the network over the course of learning. The peaks in influence match the branching points of the MDS trajectory, where each hierarchical category develops (black points). losses that represent the standard for LLM pretraining: ℓi(… view at source ↗
Figure 5
Figure 5. Figure 5: Token–class relationships. BIF between structural classes through training. We observe structural rela￾tionships between tokens reflected in influence patterns between classes, including strong intra-class relationships, development of induction, and relationships between word elements and corresponding delimiter token classes. Dashed lines indicate major inflection points in the BIF. Opacity captures stro… view at source ↗
Figure 6
Figure 6. Figure 6: BIF–LOO trace correlation with varying inverse temperature β, localization strength γ and step size ϵ. On a single grid point between two hyperparameters (e.g. γ, ϵ in the leftmost), multiple points refer to different values of the remaining hyperparameter (β in the leftmost). In general, high γ and low ϵ gave the highest correlation with the gold standard LOO experiment, and β was less significant in the … view at source ↗
Figure 7
Figure 7. Figure 7: BIF for perturbing each data point. The subtitle indicates the perturbed data point. Analytic Loss Difference query pine sparrow penguin cat dog oak rose lily [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Leave-One-Out (LOO) loss difference for perturbing each data point. The title indicates the ablated data point. we take the approximation methods as in Koh & Liang (2020); Grosse et al. (2023). First is to add a constant dampening 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Influence with damped Hessian approximation. We measure the classical influence of the dog sample with damped Hessian approximation H˜ with varying dampening constant γ. C.5 Time-Specific Ablation Experiment Our observation that influence is a time-dependent function proposes another perspective - perturbing the same data point but at different training times would lead to non-identical influence. That is,… view at source ↗
Figure 10
Figure 10. Figure 10: Influence with damped Gauss-Newton-Hessian approximation. We measure the classical influence of the dog sample with damped Gauss-Newton-Hessian approximation H˜ with varying dampening constant γ. Integrated Ablation Ablation [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Brief ablation of dog. For each experiment, we ablate dog for duration D = 100 epochs starting from timepoint t. Starting time t is sampled uniformly with an interval of 200 epochs. For each experiment, we integrate the loss difference to the baseline during D. We show the integrated loss of query (a) cat and (b) sparrow over different ablation windows. The most critical period of influence from the dog o… view at source ↗
Figure 12
Figure 12. Figure 12: Analytical loss difference from perturbing each data point. The title indicates the perturbed data point (ε = −0.1). D.1 Structural Token Classification Here, we present additional details and further experiments conducted to investigate the development of influence with respect to how tokens structure text. These experiments were conducted using the 14 million parameter Pythia model. Tokens are generated… view at source ↗
Figure 13
Figure 13. Figure 13: Examples of structural classification of tokens. Bold outline indicates class membership. The classes are as follows: 25 [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: KNN BIF predictions change over time. Using the BIF in order to select the top 30 highest influence tokens and taking the majority class from among them (adjusted for class rates), we plot the recall minus the class rate for each class. This means that 0 represents the random baseline. We see that the BIF selects for in-class tokens above random for all categories by the end of training, indicating that i… view at source ↗
Figure 15
Figure 15. Figure 15: Influence dynamics during induction head formation. We trained a small transformer from scratch with high-frequency checkpointing to capture the formation of induction heads. (Top) The normalized BIF between corresponding tokens in repeated sequences (blue) shows a sharp increase and peak, while correlations between non-repeated samples (red) or between non-repeated and repeated segments (green) remain lo… 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. Dead Directions: Geometric Singular Learning

    cs.LG 2026-06 unverdicted novelty 7.0

    Dead directions recover Watanabe's RLCT contribution and triple (λ, m, ν) from directional Fisher curvature decay rates in original parameter space for singular models, extended via K-FAC to networks and gauge-equivar...

  2. Mapping data sensitivities in global QCD analysis with linear response and influence functions

    hep-ph 2026-04 unverdicted novelty 7.0

    A framework based on linear response and influence functions maps data sensitivities in global QCD analyses to show how experiments determine central values, uncertainties, and correlations of non-perturbative functions.

  3. Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units

    cs.CL 2026-01 conditional novelty 6.0

    Influence scores computed against a circuit's behavior identify training samples whose deletion delays—and whose duplication accelerates—induction-head formation in Pythia models.

Reference graph

Works this paper leans on

57 extracted references · 2 canonical work pages · cited by 3 Pith papers · 1 internal anchor

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    The loss kernel: A geometric probe for deep learning interpretability, 2025

    Maxwell Adam, Zach Furman, Wilson Wu, Philipp Alexander Kreer, and Jesse Hoogland. The loss kernel: A geometric probe for deep learning interpretability, 2025

    2025

  3. [3]

    Consideration on the learning efficiency of multiple-layered neural networks with linear units

    Miki Aoyagi. Consideration on the learning efficiency of multiple-layered neural networks with linear units. Neural Networks, 172: 0 106132, 2024

    2024

  4. [4]

    If Influence Functions are the Answer , Then What is the Question ? 2022

    Juhan Bae, Nathan Ng, Alston Lo, Marzyeh Ghassemi, and Roger Grosse. If Influence Functions are the Answer , Then What is the Question ? 2022. CitationKey: pbrf-if

    2022

  5. [5]

    Juhan Bae, Wu Lin, Jonathan Lorraine, and Roger B. Grosse. Training data attribution via approximate unrolling. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang (eds.), Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, Neu...

    2024

  6. [6]

    Structural Inference : Studying Small Language Models with Susceptibilities , April 2025

    Garrett Baker, George Wang, Jesse Hoogland, and Daniel Murfet. Structural Inference : Studying Small Language Models with Susceptibilities , April 2025. URL http://arxiv.org/abs/2504.18274. arXiv:2504.18274 [cs]

    arXiv 2025

  7. [7]

    Curriculum learning

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pp.\ 41--48, 2009

    2009

  8. [8]

    Pythia: a suite for analyzing large language models across training and scaling

    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. In Proceedings of the 40th International Confere...

    2023

  9. [9]

    Dynamics of Transient Structure in In - Context Linear Regression Transformers , January 2025

    Liam Carroll, Jesse Hoogland, Matthew Farrugia-Roberts, and Daniel Murfet. Dynamics of Transient Structure in In - Context Linear Regression Transformers , January 2025. URL http://arxiv.org/abs/2501.17745. arXiv:2501.17745 [cs]

    Pith/arXiv arXiv 2025

  10. [10]

    HyDRA : Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks

    Yuanyuan Chen, Boyang Li, Han Yu, Pengcheng Wu, and Chunyan Miao. HyDRA : Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks . Proceedings of the AAAI Conference on Artificial Intelligence, 35 0 (8): 0 7081--7089, May 2021. ISSN 2374-3468. doi:10.1609/aaai.v35i8.16871. URL https://ojs.aaai.org/index.php/AAAI/article/view/16871. Nu...

    work page doi:10.1609/aaai.v35i8.16871 2021

  11. [11]

    Dynamical versus bayesian phase transitions in a toy model of superposition

    Zhongtian Chen, Edmund Lau, Jake Mendel, Susan Wei, and Daniel Murfet. Dynamical versus bayesian phase transitions in a toy model of superposition. arXiv preprint arXiv:2310.06301, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Training Data Attribution ( TDA ): Examining Its Adoption & Use Cases , January 2025

    Deric Cheng, Juhan Bae, Justin Bullock, and David Kristofferson. Training Data Attribution ( TDA ): Examining Its Adoption & Use Cases , January 2025. URL http://arxiv.org/abs/2501.12642. arXiv:2501.12642 [cs]

    Pith/arXiv arXiv 2025

  13. [13]

    Dennis Cook

    R. Dennis Cook. Detection of influential observation in linear regression. Technometrics : a journal of statistics for the physical, chemical, and engineering sciences, February 1977. ISSN 0040-1706. URL https://www.tandfonline.com/doi/abs/10.1080/00401706.1977.10489493. tex.copyright: Copyright Taylor and Francis Group, LLC

    arXiv 1977

  14. [14]

    Characterizations of an empirical influence function for detecting influential cases in regression

    R Dennis Cook and Sanford Weisberg. Characterizations of an empirical influence function for detecting influential cases in regression. Technometrics, 22 0 (4): 0 495--508, 1980

    1980

  15. [15]

    Dennis Cook and Sanford Weisberg

    R. Dennis Cook and Sanford Weisberg. Residuals and influence in regression. Monographs on statistics and applied probability. Chapman and Hall, New York, 1982. ISBN 0-412-24280-0. URL https://hdl.handle.net/11299/37076

    1982

  16. [16]

    Multidimensional scaling

    Michael AA Cox and Trevor F Cox. Multidimensional scaling. In Handbook of data visualization, pp.\ 315--347. Springer, 2008

    2008

  17. [17]

    The P ile: An 800gb dataset of diverse text for language modeling

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The P ile: An 800gb dataset of diverse text for language modeling. CoRR, abs/2101.00027, 2021. URL https://arxiv.org/abs/2101.00027

    Pith/arXiv arXiv 2021

  18. [18]

    Ryan Giordano, Tamara Broderick, and Michael I. Jordan. Covariances, robustness, and variational Bayes . Journal of Machine Learning Research, 19: 0 51:1--51:49, 2017. URL https://api.semanticscholar.org/CorpusID:53238793

    2017

  19. [19]

    Automated curriculum learning for neural networks

    Alex Graves, Marc G Bellemare, Jacob Menick, Remi Munos, and Koray Kavukcuoglu. Automated curriculum learning for neural networks. In international conference on machine learning, pp.\ 1311--1320. Pmlr, 2017

    2017

  20. [20]

    Roger Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, Evan Hubinger, Kamilė Lukošiūtė, Karina Nguyen, Nicholas Joseph, Sam McCandlish, Jared Kaplan, and Samuel R. Bowman. Studying Large Language Model Generalization with Influence Functions , August 2023. URL http://arxiv.o...

    Pith/arXiv arXiv 2023

  21. [21]

    Learning distributed representations of concepts

    Geoffrey E Hinton. Learning distributed representations of concepts. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 8, 1986

    1986

  22. [22]

    Emergence of computational structure in a neural network physics simulator

    Rohan Hitchcock, Gary W Delaney, Jonathan H Manton, Richard Scalzo, and Jingge Zhu. Emergence of computational structure in a neural network physics simulator. arXiv preprint arXiv:2504.11830, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    The developmental landscape of in-context learning

    Jesse Hoogland, George Wang, Matthew Farrugia-Roberts, Liam Carroll, Susan Wei, and Daniel Murfet. The developmental landscape of in-context learning. arXiv preprint arXiv:2402.02364, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Loss landscape degeneracy and stagewise development in transformers

    Jesse Hoogland, George Wang, Matthew Farrugia-Roberts, Liam Carroll, Susan Wei, and Daniel Murfet. Loss landscape degeneracy and stagewise development in transformers. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=45qJyBG8Oj

    2025

  25. [25]

    The growth of logical thinking from childhood to adolescence: An essay on the construction of formal operational structures

    B \"a rbel Inhelder and Jean Piaget. The growth of logical thinking from childhood to adolescence: An essay on the construction of formal operational structures. Routledge, 1958

    1958

  26. [26]

    Semantic and conceptual development: An ontological perspective

    Frank C Keil. Semantic and conceptual development: An ontological perspective. Harvard University Press, 1979

    1979

  27. [27]

    Understanding black-box predictions via influence functions, December 2020

    Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions, December 2020. URL http://arxiv.org/abs/1703.04730. arXiv:1703.04730 [stat] CitationKey: deep-influence-functions

    Pith/arXiv arXiv 2020

  28. [28]

    Bayesian influence functions for hessian-free data attribution, 2025

    Philipp Alexander Kreer, Wilson Wu, Maxwell Adam, Zach Furman, and Jesse Hoogland. Bayesian influence functions for hessian-free data attribution, 2025

    2025

  29. [29]

    The local learning coefficient: a singularity-aware complexity measure

    Edmund Lau, Zach Furman, George Wang, Daniel Murfet, and Susan Wei. The local learning coefficient: a singularity-aware complexity measure. In The 28th international conference on artificial intelligence and statistics, 2025. URL https://openreview.net/forum?id=1av51ZlsuL

    2025

  30. [30]

    Why do animals need shaping? a theory of task composition and curriculum learning

    Jin Hwa Lee, Stefano Sarao Mannelli, and Andrew Saxe. Why do animals need shaping? a theory of task composition and curriculum learning. arXiv preprint arXiv:2402.18361, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    You Are What You Eat -- AI Alignment Requires Understanding How Data Shapes Structure and Generalisation , February 2025

    Simon Pepin Lehalleur, Jesse Hoogland, Matthew Farrugia-Roberts, Susan Wei, Alexander Gietelink Oldenziel, George Wang, Liam Carroll, and Daniel Murfet. You Are What You Eat -- AI Alignment Requires Understanding How Data Shapes Structure and Generalisation , February 2025. URL http://arxiv.org/abs/2502.05475. arXiv:2502.05475 [cs]

    Pith/arXiv arXiv 2025

  32. [32]

    Preconditioned stochastic gradient langevin dynamics for deep neural networks

    Chunyuan Li, Changyou Chen, David Carlson, and Lawrence Carin. Preconditioned stochastic gradient langevin dynamics for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016

    2016

  33. [33]

    Tilting the odds at the lottery: the interplay of overparameterisation and curricula in neural networks

    Stefano Sarao Mannelli, Yaraslau Ivashynka, Andrew Saxe, and Luca Saglietti. Tilting the odds at the lottery: the interplay of overparameterisation and curricula in neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2024 0 (11): 0 114001, 2024

    2024

  34. [34]

    Optimizing neural networks with Kronecker -factored approximate curvature

    James Martens and Roger Grosse. Optimizing neural networks with Kronecker -factored approximate curvature. In Francis Bach and David Blei (eds.), Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pp.\ 2408--2417, Lille, France, 07--09 Jul 2015. PMLR. URL https://proceedings.mlr.pre...

    2015

  35. [35]

    A connectionist perspective on knowledge and development

    JL McClelland. A connectionist perspective on knowledge and development. In Developing cognitive competence, pp.\ 157--204. Psychology Press, 1995

    1995

  36. [36]

    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

  37. [37]

    Estimating Training Data Influence by Tracing Gradient Descent

    Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. Estimating Training Data Influence by Tracing Gradient Descent . In Advances in Neural Information Processing Systems , volume 33, pp.\ 19920--19930. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/hash/e6385d39ec9394f2f3a354d9d2b88eec-Abstract.html. CitationK...

    2020

  38. [38]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pp.\ 5301--5310. PMLR, 2019

    2019

  39. [39]

    Semantic cognition: A parallel distributed processing approach

    Timothy T Rogers and James L McClelland. Semantic cognition: A parallel distributed processing approach. MIT press, 2004

    2004

  40. [40]

    1 learning and connectionist

    David E Rumelhart and Peter M Todd. 1 learning and connectionist. Attention and performance XIV: Synergies in experimental psychology, artificial intelligence, and cognitive neuroscience, 14: 0 3, 1993

    1993

  41. [41]

    Exact solutions to the nonlinear dynamics of learning in deep linear neural networks

    Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013

    Pith/arXiv arXiv 2013

  42. [43]

    Saxe, James L

    Andrew M. Saxe, James L. McClelland, and Surya Ganguli. A mathematical theory of semantic development in deep neural networks. Proceedings of the National Academy of Sciences, 116 0 (23): 0 11537--11546, June 2019 b . ISSN 0027-8424, 1091-6490. doi:10.1073/pnas.1820226116. URL https://pnas.org/doi/full/10.1073/pnas.1820226116

    doi:10.1073/pnas.1820226116 2019

  43. [44]

    LLM Circuit Analyses Are Consistent Across Training and Scale

    Curt Tigges, Michael Hanna, Qinan Yu, and Stella Biderman. LLM Circuit Analyses Are Consistent Across Training and Scale . In The Thirty-eighth Annual Conference on Neural Information Processing Systems , November 2024

    2024

  44. [45]

    Multidimensional scaling: I

    Warren S Torgerson. Multidimensional scaling: I. theory and method. Psychometrika, 17 0 (4): 0 401--419, 1952

    1952

  45. [46]

    Structure development in list-sorting transformers

    Einar Urdshals and Jasmina Urdshals. Structure development in list-sorting transformers. arXiv preprint arXiv:2501.18666, 2025

    Pith/arXiv arXiv 2025

  46. [47]

    Deep learning generalizes because the parameter-function map is biased towards simple functions

    Guillermo Valle-Perez, Chico Q Camargo, and Ard A Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions. arXiv preprint arXiv:1805.08522, 2018

    Pith/arXiv arXiv 2018

  47. [48]

    Asymptotic statistics, volume 3

    Aad W Van der Vaart. Asymptotic statistics, volume 3. Cambridge university press, 2000

    2000

  48. [49]

    Embryology of a Language Model , August 2025 a

    George Wang, Garrett Baker, Andrew Gordon, and Daniel Murfet. Embryology of a Language Model , August 2025 a . URL http://arxiv.org/abs/2508.00331. arXiv:2508.00331 [cs]

    Pith/arXiv arXiv 2025

  49. [50]

    Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient

    George Wang, Jesse Hoogland, Stan van Wingerden, Zach Furman, and Daniel Murfet. Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient . In Proceedings of The 13th International Conference on Learning Representations , 2025 b . URL https://openreview.net/forum?id=SUc1UOWndp&noteId=MCoFYhi7ZE

    2025

  50. [51]

    A survey on curriculum learning

    Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44 0 (9): 0 4555--4576, 2021

    2021

  51. [52]

    Algebraic Geometry and Statistical Learning Theory

    Sumio Watanabe. Algebraic Geometry and Statistical Learning Theory. Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press, 2009

    2009

  52. [53]

    Mathematical theory of Bayesian statistics

    Sumio Watanabe. Mathematical theory of Bayesian statistics . Chapman and Hall, 2018

    2018

  53. [54]

    Bayesian learning via stochastic gradient Langevin dynamics

    Max Welling and Yee Whye Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th International Conference on International Conference on Machine Learning, ICML'11, pp.\ 681–688, Madison, WI, USA, 2011. Omnipress. ISBN 9781450306195

    2011

  54. [55]

    When do curricula work? arXiv preprint arXiv:2012.03107, 2020

    Xiaoxia Wu, Ethan Dyer, and Behnam Neyshabur. When do curricula work? arXiv preprint arXiv:2012.03107, 2020

    Pith/arXiv arXiv 2012

  55. [56]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  56. [57]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  57. [58]

    Gradient Similarity: An Explainable Approach to Detect Adversarial Attacks against Deep Learning

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1806.10707 2025