Pith. sign in

REVIEW 5 major objections 5 minor 22 references

The temporal structure of a drifting data stream is recoverable from the weights of the models trained on it, and that structure predicts how well those models transfer.

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

2026-08-01 07:08 UTC pith:TU72MQY7

load-bearing objection The core idea is genuinely new—HMM on aligned per-window weights with a within-state transfer test—but the 'beyond the data distribution' claim outruns the controls, and the HMM's edge over a contiguous partition is small. the 5 major comments →

arxiv 2607.27482 v1 pith:TU72MQY7 submitted 2026-07-29 cs.LG cs.AIcs.CL

Latent States in Neural Networks: Recovering the Temporal Structure of Drifting Data from Model Weights

Kevin Guan This is my paper
classification cs.LG cs.AIcs.CL
keywords latent stateshidden Markov modeltemporal driftmodel weightsevent segmentationtransfer generalizationlabel shiftstreaming data
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.

The paper tests whether a data stream that drifts over time passes through discrete regimes, and whether those regimes can be read off from the weights of classifiers trained on successive windows of the stream. Fitting a hidden Markov model to the aligned weight trajectory of models trained on two drifting datasets—misinformation detection and sentiment analysis—yields latent states that partition each timeline into coherent phases. Models generalize better to data from windows sharing their training window's state than across state boundaries, and this within-state advantage survives controls for temporal proximity, class divergence, and a naive equal-size partition. The authors argue this parallels how humans segment continuous experience into discrete events and offers an annotation-free way to detect regime shifts in streaming data.

Core claim

The paper's central claim is that the temporal structure of a drifting data stream is recoverable from the weights of the models it gives rise to, and membership in that structure predicts generalization in a manner that parallels the within-event coherence of human memory. Concretely, across both Fakeddit and Yelp, the within-state transfer advantage—the difference in column-centered macro F1 between same-state and cross-state window pairs—is positive, remains after stratifying by temporal lag and after residualizing out per-pair class divergence, and exceeds a contiguous equal-size partition with the same number of states. The states are estimated entirely from model weights, yet they corr

What carries the argument

The central mechanism is a left-to-right hidden Markov model fit to the chronological trajectory of per-window weight centroids. Model weights from independently trained classifiers are aligned across seeds via iterative linear-assignment matching on unit signatures to remove permutation symmetry, projected with PCA, z-score normalized, and averaged per window; the HMM's Gaussian emissions then locate boundaries between latent regimes, with the Viterbi decode assigning each window to a state. The left-to-right transition constraint encodes the assumption that a regime, once left, does not recur, mirroring the event-segmentation framework from human perception. The within-state advantage is a

Load-bearing premise

The claim that the within-state advantage reflects latent regimes beyond the data distribution rests on the assumption that per-pair Jensen-Shannon divergence of class counts plus temporal lag fully controls for 'data distribution'; if true drift includes covariate shift or label shift beyond marginal class fractions, the residual signal could be an artifact.

What would settle it

Train the same pipeline on a synthetic drifting stream where regime boundaries are known and drift is covariate-only (class margins held constant). If the HMM fails to recover the boundaries—or the within-state advantage vanishes once class divergence and lag are residualized—the claim that weights encode latent regimes beyond label shift is falsified. Alternatively, segment the timeline directly from class distributions and compare transfer advantage; if class-only segmentation matches or beats the HMM, the weight-derived states add nothing beyond label shift.

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

If this is right

  • A model's weight trajectory can serve as an unsupervised probe for regime shifts in streaming data, without access to labels or data statistics.
  • The within-state generalization advantage suggests retraining or adaptation decisions can be timed to HMM-detected boundaries rather than fixed calendar schedules.
  • The method is most informative when label-marginal drift is present; domains with pure covariate shift may require a richer measure than class JSD.
  • The framework extends to any sequence of model-internal representations elicited by ordered inputs, not just weights.
  • The equal-size baseline reproducing most of the advantage implies the timeline's dominant regime shifts are few and large; the HMM's value lies in refining boundary placement.

Where Pith is reading between the lines

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

  • A controlled synthetic stream with known ground-truth regime boundaries and orthogonal covariate vs. label shift could directly test whether weight-space boundaries align with true boundaries or merely with label-shift discontinuities.
  • If the weight-derived states are genuinely about event structure, a similar within-state transfer advantage should appear in continual learning benchmarks with known task boundaries, offering a bridge to task segmentation.
  • The weak Yelp effect hints that the HMM-on-weights pipeline may be blind to covariate shift such as lexical polarity changes; a covariate-shift-only synthetic stream would isolate this limitation.
  • A stricter baseline than equal-size contiguous blocks would be a partition optimized on class labels alone; if the weight-derived segmentation does not beat it, the 'beyond the data distribution' claim would need qualification.

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

5 major / 5 minor

Summary. The paper asks whether discrete regimes in a drifting data stream can be recovered from the weights of classifiers trained on successive temporal windows. For Fakeddit and Yelp, the authors train MLPs on non-overlapping windows, align weights across seeds by Hungarian matching, reduce dimensionality with PCA, fit a left-to-right HMM to the per-window weight centroids, and decode a state sequence. They then test whether transfer macro-F1 is higher for window pairs sharing a state than for pairs crossing a boundary. They report a within-state advantage on both datasets, survival of the advantage after controlling for temporal lag and class-label Jensen-Shannon divergence, a modest but significant advantage over an equal-size contiguous segmentation, and a stronger correlation of state-pair transfer performance with class-distribution divergence than with weight-space centroid distance. The central interpretation is that latent states recover transfer-relevant temporal structure 'beyond the data distribution.'

Significance. If the central claim holds, the paper makes a novel methodological contribution: unsupervised recovery of drifting-data regimes from model weights, with a functional generalization test analogous to event segmentation in human memory research. The pipeline is transparent, the code is promised, and the use of two datasets with explicit permutation controls and an equal-size baseline are strengths. However, the statistical evidence currently has unresolved validity issues: the lag-stratified permutation test appears to treat per-seed observations as independent, the 'beyond the data distribution' claim controls only label-marginal shift, and the Yelp combined-model estimate is negative despite being reported as significant. These issues must be addressed before the headline claims can be accepted.

major comments (5)
  1. [§3.6, §4.3] The lag-stratified permutation test is described as permuting within/across indicators over per-seed observations, with each window pair i→j evaluated by 10 seeds sharing the same pair-level indicator. This treats 10 dependent replicates as independent and inflates the effective sample size by roughly a factor of 10. The reported p<0.0001 values are therefore not trustworthy as written. Please permute at the window-pair level (e.g., average F1 over seeds before testing) or use a cluster/block permutation that keeps all seeds of a pair together, and report the resulting p-values.
  2. [§3.10, §4.7, Abstract] The residualization in Eq. (2) controls only per-pair class-label JSD and lag. As the paper's own Limitations 5.2 acknowledges, covariate shift is part of the data distribution and is not controlled; e.g., lexical polarity drift in review sentiment (Lukes & Søgaard 2018) would be invisible to the JSD measure. Thus the abstract's 'beyond the data distribution' claim is not established: HMM boundaries aligned with covariate shift would survive JSD+lag residualization while being entirely within-distribution. Add a feature-space shift control, or explicitly restrict the claim to label-marginal shift.
  3. [§4.7, Yelp] The combined-model estimate for Yelp is β_state = −0.0003, a negative coefficient. Reporting a one-sided p = 0.0234 against a Freedman-Lane null that is itself centered below zero does not demonstrate a positive within-state advantage; the point estimate is negative. The two-stage residualization gives a positive gap (+0.0054), so the two analyses are in tension. Reconcile the two tests or temper the claim that the Yelp combined model supports a positive within-state effect. As reported, the Yelp replication is weaker than the text suggests.
  4. [§3.5, §4.2] K and the number of PCA components are selected on the same time series used to evaluate the state partition (BIC, leave-one-seed-out CV, scree elbow), yet the permutation tests condition on the selected state sequence as fixed. This selection uncertainty is not reflected in the null distributions. With K=11 on 35 windows and K=16 on 56 windows, there is also a risk of overfitting the weight trajectory. Please report how the main results vary with K and with the number of retained components, or explicitly frame the inference as exploratory rather than confirmatory.
  5. [§3.4, §4.3] The HMM is fit to the same weight vectors whose transfer performance defines the within-state advantage, so part of the advantage is expected by construction: windows with similar weights are grouped together, and weight-similar models tend to transfer better. The equal-size baseline and the JSD+lag controls do not control for pairwise weight similarity between source and target. Residualizing transfer F1 on per-pair weight distance (or including it in Eq. (2)) would directly address this circularity concern; without such a control, the 'recoverable from weights' claim is partly circular.
minor comments (5)
  1. [§4.3] Clarify whether the reported pair counts (114 within, 1,076 across for Fakeddit) are window-pair counts or per-seed observation counts. The text currently mixes 'pairs' with the per-seed evaluation described in §3.6.
  2. [§4.1] Only ~11% of weight variance is retained in the PCA projection. Please justify more explicitly that the retained components carry the temporal signal, e.g., by showing the fraction of drift variance captured or the stability of results with additional components.
  3. [§4.2 / Figure 2] For Yelp, K=16 is selected but state 15 is unvisited. State consistently whether the effective number of states is 15 when reporting pair counts and state-pair analyses.
  4. [§5.2] The limitation concerning covariate shift should be stated next to the 'beyond the data distribution' claim in the abstract and §4.7, not only in Limitations; otherwise the abstract overstates the current evidence.
  5. [Throughout] Minor typographical issues: 'Y elp' at the start of §3.3, 'dshift' in §5.2, and inconsistent naming of 'distance-conditioned' versus 'distance-conditioned gap statistic' in places.

Circularity Check

0 steps flagged

No significant circularity: the HMM is fit to weights without using transfer F1, so the within-state transfer advantage is an empirical association, not a construction.

full rationale

The derivation chain is not circular. The HMM in Section 3.5 is fit by Baum-Welch to per-window PCA weight centroids, optimizing likelihood of the weight trajectory; the transfer matrices in Section 3.6 are computed from the same trained models, but the HMM never sees F1, K is selected by BIC and held-out weight log-likelihood, and no parameter is fit to the within/across labels. The within-state advantage is therefore an empirical association between an unsupervised weight-based partition and an independently measured generalization outcome, not an identity. The equal-size baseline and lag-stratified permutation tests are genuine comparisons. The residualization in Section 4.7 controls JSD and lag but, as Limitations 5.2 concedes, does not control covariate shift; that is a construct-validity gap in the 'beyond the data distribution' claim, not a circular reduction. There are no self-citations, imported uniqueness theorems, or ansatz-by-citation that are load-bearing. The absence of a non-temporal weight-clustering control weakens the temporal interpretation, but it does not make the derivation equivalent to its inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central empirical claims rest on a chain of modeling choices: windowing, 3-PC projection, non-recurrent HMM, and JSD-as-distribution-control. No new physical or mechanistic entities are postulated. The most delicate choices are K, PC count, and the sufficiency of JSD as a control.

free parameters (3)
  • Number of HMM states K = Fakeddit 11; Yelp 16 (15 visited)
    Chosen by BIC and leave-one-seed-out likelihood on the same weight trajectories used for the transfer test; this model selection can inflate the apparent structure.
  • PCA components retained = 3 PCs; 11.47% (Fakeddit), 11.10% (Yelp) variance
    Chosen by scree-elbow on the aligned weights; the central claim depends on these three components carrying the temporal signal.
  • Window duration and minimum-count thresholds = 60 days / 9,000 (Fakeddit); 90 days / 10,000 (Yelp)
    Chosen separately per dataset; they determine the number of windows, the sample sizes, and which early/late periods are excluded.
axioms (4)
  • domain assumption Latent regimes are non-recurrent and contiguous (left-to-right HMM with no revisits).
    Section 3.5 imposes a_{ij}=0 unless j in {i,i+1}. If regimes recur seasonally, the model is forced to invent new states; this is acknowledged in Limitations 5.2.
  • domain assumption Permutation alignment removes the relevant hidden-unit symmetries; remaining symmetries (e.g., ReLU rescaling) do not distort the trajectory.
    Section 3.4 and Limitations 5.2; if unremoved symmetries dominate, the PCA/HMM distances are not meaningful weight-space geometry.
  • domain assumption Three PCA components of the aligned weights capture the temporal regime signal.
    Section 4.1 retains only ~11% of weight variance; the entire HMM is fit in this 3-D subspace.
  • ad hoc to paper Jensen-Shannon divergence of per-window class counts plus lag is a sufficient control for 'data distribution' in the residualization.
    Section 3.10 uses JSD of class label fractions only; covariate shift invisible to label marginals is not controlled, yet the abstract claims signal 'beyond the data distribution.'

pith-pipeline@v1.3.0-daily-deepseek · 16160 in / 14703 out tokens · 156530 ms · 2026-08-01T07:08:14.629838+00:00 · methodology

0 comments
read the original abstract

A temporally drifting data stream may pass through discrete regimes rather than changing continuously. We ask whether such regimes are recoverable from the weights of models trained on the stream, using a hidden Markov model (HMM) fit to the chronologically ordered trajectory of those weights. We study this question in two domains known to drift over time: multimodal misinformation detection, using the Fakeddit dataset; and sentiment analysis, using the Yelp dataset. We train classifiers on consecutive temporal windows and fit an HMM to the trajectory of their aligned weights, recovering latent states that partition each timeline into coherent phases. On both datasets, classifiers generalize better to data from windows sharing the state of their training window than to windows across state boundaries. This within-state transfer advantage survives a control for temporal proximity and modestly exceeds the advantage recovered by a naive partition into contiguous states of equal size. Although the states are estimated solely from model weights, they correlate more strongly with shifts in the data's class distribution than with the weight-space geometry used to estimate them. After class divergence and lag are residualized out, the within-state advantage exceeds its permutation null on both tasks, indicating that the states recover structure relevant to transfer beyond the data distribution. Every effect replicates on both tasks but is attenuated on Yelp, whose label distribution is more temporally stable.

Figures

Figures reproduced from arXiv: 2607.27482 by Kevin Guan.

Figure 1
Figure 1. Figure 1: PCA projections of aligned MLP head weights (PC1 vs. PC2) for (a) Fakeddit and (b) Yelp. After [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Viterbi-decoded HMM state sequences; color indicates the assigned state. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean column-centered macro F1 for within- versus across-state window pairs as a function of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fakeddit: dissimilarity between state pairs versus mean cross-window macro F1. The data-side [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Yelp: dissimilarity between state pairs versus mean cross-window macro F1. Both associations [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The 6-way class distribution across the 35 temporal windows in the Fakeddit dataset (January [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Null distribution of the distance-conditioned gap statistic under 10,000 label-shuffle permutations. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Null distribution of the HMM−equal-size gap difference under 10,000 per-window label permuta￾tions. The observed difference falls outside the null distribution for both datasets. C.3 Correlation Analyses [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Fakeddit: permutation null distributions (10,000 iterations) for the two Spearman correlations [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Yelp: permutation null distributions (10,000 iterations) for the two Spearman correlations between [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Null distribution of the difference in mean pairwise JSD across- versus within-states under [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Freedman-Lane null distribution for the same-state coefficient [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Null distribution of the lag-stratified, harmonic-weighted distance-conditioned gap statistic ap [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: PCA projections of the flattened MLP head weights (PC1 vs. PC2) before permutation alignment, [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Per-component explained variance ratio for the first 50 principal components of the aligned weight [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: HMM model selection for the Fakeddit dataset ( [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: HMM model selection for the Yelp dataset ( [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Fakeddit: within- and across-group macro F1 as a function of temporal lag for the HMM ( [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Yelp: within- and across-group macro F1 as a function of temporal lag for the HMM ( [PITH_FULL_IMAGE:figures/full_fig_p023_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: The 5-way class distribution across the 56 temporal windows in the Yelp dataset (May 2008– [PITH_FULL_IMAGE:figures/full_fig_p024_20.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa

    Samuel K. Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries.arXiv preprint arXiv:2209.04836,

    Pith/arXiv arXiv

  2. [2]

    For 2-way classification, micro F1 fell by 3.3 percentage points (0.886 to 0.853) and macro F1 by 6.1 points (0.881 to 0.820)

    Moving from the original to the temporal split produced consistent degradation across both granularities. For 2-way classification, micro F1 fell by 3.3 percentage points (0.886 to 0.853) and macro F1 by 6.1 points (0.881 to 0.820). For the 6-way task, micro F1 dropped by 1.9 points (0.849 to 0.830) and macro F1 by 6.4 points (0.737 to 0.673). The larger ...

    2023

  3. [3]

    B.1 Permutation Alignment This expands the alignment procedure of Section 3.4

    Each subsection expands a step summarized in the main text. B.1 Permutation Alignment This expands the alignment procedure of Section 3.4. Hidden units in a feed-forward network are symmetric under permutation (Entezari et al., 2022), so weight vectors must be aligned across seeds before they can be compared. Each hidden unitkis assigned an alignment sign...

    2022

  4. [10]

    doi: 10.18653/v1/ P18-2110

    Association for Computational Linguistics. doi: 10.18653/v1/ P18-2110. URLhttps://aclanthology.org/P18-2110/. Kokil Jaidka, Niyati Chhaya, and Lyle Ungar. Diachronic degradation of language models: Insights from social media. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 195–200, Me...

    doi:10.18653/v1/

  5. [12]

    Jan Ole Krugmann and Jochen Hartmann

    URLhttps://arxiv.org/abs/ 1606.05320. Jan Ole Krugmann and Jochen Hartmann. Sentiment analysis in the age of generative ai.Customer Needs and Solutions, 11(1):3,

    Pith/arXiv arXiv

  6. [14]

    doi: 10.18653/v1/W18-6210

    Association for Computational Linguistics. doi: 10.18653/v1/W18-6210. URL https://aclanthology.org/W18-6210/. Kai Nakamura, Sharon Levy, and William Yang Wang. Fakeddit: A new multimodal benchmark dataset for fine-grained fake news detection. In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sar...

    work page doi:10.18653/v1/w18-6210

  7. [15]

    ISBN 979-10-95546-34-4

    European Language Resources Association. ISBN 979-10-95546-34-4. URLhttps://aclanthology.org/2020.lrec-1.755/. Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 Conference on Empirical Methods in Natural Lan- guage Processing and the 9th International ...

    2020

  8. [18]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108,

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108,

    Pith/arXiv arXiv 1910

  9. [19]

    doi: 10.18653/v1/2023.genbench-1.6

    Association for Computational Linguistics. doi: 10.18653/v1/2023.genbench-1.6. URLhttps://aclanthology.org/ 2023.genbench-1.6/. Daegun Won, Peter J. Jansen, and Jaime G. Carbonell. Temporal transfer learning for drift adaptation. In Proceedings of the 26th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (E...

    work page doi:10.18653/v1/2023.genbench-1.6 2023

  10. [20]

    Liu Ziyin and Masahito Ueda

    Accessed: 2026-07-01. Liu Ziyin and Masahito Ueda. Exact phase transitions in deep learning.arXiv preprint arXiv:2205.12510,

    Pith/arXiv arXiv 2026

  11. [1983]

    Mario Geiger, Stefano Spigler, Stéphane d’Ascoli, Levent Sagun, Marco Baity-Jesi, Giulio Biroli, and Matthieu Wyart

    doi: 10.1080/07350015.1983.10509354. Mario Geiger, Stefano Spigler, Stéphane d’Ascoli, Levent Sagun, Marco Baity-Jesi, Giulio Biroli, and Matthieu Wyart. The jamming transition as a paradigm to understand the loss landscape of deep neural networks.arXiv preprint arXiv:1809.09349,

    arXiv 1983

  12. [2013]

    Rahim Entezari, Hanie Sedghi, Olga Saukh, and Behnam Neyshabur

    doi: 10.1037/a0034024. Rahim Entezari, Hanie Sedghi, Olga Saukh, and Behnam Neyshabur. The role of permutation invariance in linear mode connectivity of neural networks.arXiv preprint arXiv:2110.06296,

    Pith/arXiv arXiv

  13. [2014]

    Komal Florio, Valerio Basile, Marco Polignano, Pierpaolo Basile, and Viviana Patti

    doi: 10.1016/j.neuron.2014.01.042. Komal Florio, Valerio Basile, Marco Polignano, Pierpaolo Basile, and Viviana Patti. Time of your hate: The challenge of time in hate speech detection on social media.Applied Sciences, 10(12):4180,

    work page doi:10.1016/j.neuron.2014.01.042 2014

  14. [2016]

    Latent state models of training dynamics

    Mingyuan Hu, Annie Chen, Naomi Saphra, and Kyunghyun Cho. Latent state models of training dynamics. arXiv preprint arXiv:2308.09543,

    Pith/arXiv arXiv

  15. [2017]

    doi: 10.1016/j.neuron.2017.06.041. Erica L. Busch, Jessie Huang, Andrew Benz, Tom Wallenstein, Guillaume Lajoie, Guy Wolf, Smita Krish- naswamy, and Nicholas B. Turk-Browne. Multi-view manifold learning of human brain-state trajectories. Nature Computational Science, 3(3):240–253,

    doi:10.1016/j.neuron.2017.06.041 2017

  16. [2018]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun

    URLhttps://arxiv.org/abs/1809.09349. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778,

    Pith/arXiv arXiv

  17. [2019]

    doi: 10.18653/v1/D19-1018

    Association for Computational Linguistics. doi: 10.18653/v1/D19-1018. URLhttps://aclanthology.org/D19-1018/. Paul Röttger and Janet Pierrehumbert. Temporal adaptation of BERT and performance on downstream document classification: Insights from social media. In Marie-Francine Moens, Xuanjing Huang, Lu- cia Specia, and Scott Wen-tau Yih (eds.),Findings of t...

    doi:10.18653/v1/d19-1018 2021

  18. [2020]

    URLhttps://doi.org/10.3390/app10124180

    doi: 10.3390/app10124180. URLhttps://doi.org/10.3390/app10124180. David Freedman and David Lane. A nonstochastic interpretation of reported significance levels.Journal of Business & Economic Statistics, 1(4):292–298,

    work page doi:10.3390/app10124180

  19. [2021]

    doi: 10.18653/v1/2021.findings-emnlp.206

    Association for Com- putational Linguistics. doi: 10.18653/v1/2021.findings-emnlp.206. URLhttps://aclanthology.org/ 2021.findings-emnlp.206/. Noa Rubin, Inbar Seroussi, and Zohar Ringel. Grokking as a first order phase transition in two layer networks. InInternational Conference on Learning Representations (ICLR),

    doi:10.18653/v1/2021.findings-emnlp.206 2021

  20. [2022]

    Exact Phase Transitions in Deep Learning

    doi: 10.48550/arXiv.2205.12510. A Replication of Stepanova and Ross (2023) This appendix collects the architecture, training protocol, and full results of our replication of Stepanova & Ross (2023). Their study examined how the performance of multimodal misinformation classifiers degrades when evaluated on temporally out-of-domain data. 14 A.1 Classifier ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2205.12510 2023

  21. [2023]

    Latent State Models of Training Dynamics

    doi: 10.48550/arXiv.2308.09543. Xiaolei Huang and Michael J. Paul. Examining temporality in document classification. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 694–699, Melbourne, Australia, July

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.09543

  22. [2024]

    doi: 10.1007/s40547-024-00143-4

    ISSN 2196-291X. doi: 10.1007/s40547-024-00143-4. URLhttps: //doi.org/10.1007/s40547-024-00143-4. X. Liu, P. Li, H. Huang, Z. Li, X. Cui, J. Liang, and Z. He. Fka-owl: Advancing multimodal fake news detection through knowledge-augmented lvlms. InProceedings of the 32nd ACM International Conference on Multimedia, pp. 10154–10163, October

    doi:10.1007/s40547-024-00143-4