Pith. sign in

REVIEW 5 major objections 5 minor 31 references

Multimodal Non-Semantic Feature Fusion for Predicting Segment Access Frequency in Lecture Archives

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Non-semantic teacher motion, audio, and slide cues predict which lecture segments students revisit, reaching a Pearson correlation of 0.5143.

desk verdict An incomplete preprint with a genuinely plausible setup and no supporting experiments; the headline correlation and accuracy are promissory numbers. read the letter →

arxiv 2504.14927 v1 pith:QJBIYLXN submitted 2025-04-21 cs.HC

Ruozhu Sheng , Jinghong Li , Shinobu Hasegawa This is my paper · ORCID
classification cs.HC
keywords lecturearchivessegmentaccessfrequencymultimodalfusionnon-semanticfeaturesstudentengagementactionaudiospectrogramslideprogression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that segment-level student engagement in unedited lecture archives can be predicted from non-semantic signals alone: what the teacher does, how the voice sounds, and when slides change. It defines engagement as segment access frequency, the number of times each one-minute segment is played back by students in a small real classroom archive. The reported best model, built by stacking action, voice, and slide features into the RGB channels of a single image and running a residual convolutional network, reaches a Pearson correlation of 0.5143 in seven-fold lesson-level cross-validation and 69.32% accuracy in a three-class attention-zone task. If these numbers hold, lecture platforms could highlight likely high-engagement segments automatically, without transcripts, manual annotation, or extra hardware.

What carries the argument

The load-bearing mechanism is feature-level fusion by channel stacking: the action map (motion trajectories of the instructor's pose), the slide-progression map, and the audio spectrogram are each resized and packed into the three RGB channels of one 480×320 image, then processed by a single convolutional network trained with MSE loss. This design lets the model reuse well-established image backbones, keeps training cheap, and deliberately avoids semantics, so the pipeline is language-independent. Labels are produced in the same automated pipeline: one-minute access counts from aggregated playback logs, filtered, smoothed, and normalized per lecture.

What would settle it

Compute the same features and labels for a lecture with at least fifty valid viewers, train the identical early-fusion model, and measure Pearson correlation on held-out segments; if the correlation falls to near zero or fails to beat a baseline that predicts the per-lecture mean, the reported result is an artifact of the tiny, self-selected viewer pool rather than a robust engagement signal.

Watch

Extended reading notes

Core claim

The central claim is that the fine-grained pattern of which one-minute segments students revisit in a real lecture archive is predictable from how the instructor moves, how the audio spectrum evolves, and how the slide progresses—none of which requires understanding the lecture's content. The authors ground this in an automatically generated label: segment access frequency, computed from playback logs after filtering to valid viewers and smoothed with a five-segment centered moving average, then normalized per lecture to $[0,1]$. Their best configuration, early feature fusion through a residual convolutional backbone, achieves 0.5143 Pearson correlation and 69.32% three-class accuracy, and ablation results point to the action modality as the main driver.

Load-bearing premise

The prediction target itself—segment access frequency averaged over roughly nine valid viewers per lecture—must faithfully represent true student engagement; if those few viewers are atypical, then even a high correlation only means the model learned the habits of a handful of students.

Editorial extensions

If this is right

  • Lecture platforms could auto-generate engagement heatmaps or 'most replayed' markers from raw recordings, with no transcripts or manual annotation.
  • The non-semantic design should carry across languages and lecture domains, as long as teacher motion, audio, and slide changes are present.
  • Early fusion is enough: the single-stream RGB-stacking model is reported to beat parallel per-modality networks, so deployment remains lightweight.
  • Instructors could use segment-level replay predictions as feedback about which parts of a lecture students found worth revisiting.

Reading between the lines

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

  • The manuscript's own experiment, discussion, and conclusion sections are marked as to be completed, so the headline numbers come from the abstract and methodology rather than a fully written empirical section; those quantitative claims should be treated as provisional until the full write-up appears.
  • Because the labels are averaged over only about 8.71 valid viewers per lecture, the reported 0.5143 correlation is probably an upper bound for what the features can achieve with cleaner, larger-cohort labels.
  • The five-minute moving-average smoothing of labels means the model is trained to predict broad engagement zones, not exact moments; finer-grained logs could reveal whether the features align with conceptual boundaries such as new definitions or examples.
  • A practical follow-up is to use the predicted engagement curve to power automatic skip-and-seek recommendations, an application the paper states as motivation but does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes a multimodal neural-network framework to predict per-minute segment access frequency in lecture archives using deliberately non-semantic features: instructor action maps (OpenPose plus optical flow), audio spectrograms, and slide-progression features. The authors describe label generation from JAIST LMS viewing logs for seven lectures, yielding 665 one-minute segments after trimming/padding, with access frequencies smoothed and normalized per lecture. They outline two fusion strategies (feature stacking into RGB images and model stacking), a set of backbone comparisons (VGG, ResNet, ViT, CNN+LSTM), and a 7-fold lesson-level cross-validation protocol. The abstract reports a best Pearson correlation of 0.5143 and 69.32% three-class accuracy. However, Section 4, which is supposed to report the experiments, consists only of the sentence "This section is currently under development and will be completed in a subsequent version," and Sections 5 and 6 are marked "To be completed." Consequently, none of the claimed numerical results are inspectable in the manuscript.

Significance. If substantiated, the contribution would be practically valuable: a language-independent, hardware-free method for predicting engagement-relevant segments in unedited, low-resource lecture archives would address a real need in institutions that record face-to-face lectures without transcripts or eye-tracking. The feature choices are well grounded in prior literature, the automatic label-generation pipeline is a sensible way to avoid manual annotation, and the paper honestly acknowledges the very small dataset. The main difficulty is that the central empirical claims appear only in the abstract and have no supporting experimental content in the body; until the results are actually reported, the significance of the work cannot be assessed.

major comments (5)
  1. [Section 4; Abstract] The entire empirical basis for the paper's central claim is missing: Section 4 consists only of "This section is currently under development and will be completed in a subsequent version," and Sections 5 and 6 are "To be completed." No table, figure, per-fold result, error bar, ablation result, or evaluation detail is provided. The abstract's Pearson correlation of 0.5143 and 69.32% accuracy are therefore unsupported assertions rather than reported findings. This is load-bearing because the paper's feasibility claim (RQ1) and the fusion and backbone comparisons (RQ2 and RQ3) depend entirely on experiments that are not present.
  2. [§3.1.2, §3.1.3, §3.2.4] The metric definitions are ambiguous with respect to smoothing. Labels are smoothed with a centered five-segment moving average and normalized within each lecture, while predictions are post-processed with moving-average, Savitzky-Golay, or Kalman smoothing before thresholding. The abstract does not state whether the reported PCC and three-class accuracy are computed on raw or smoothed sequences. Since smoothing on both sides can inflate correlation and threshold agreement, the headline numbers may overstate true per-segment predictive skill. The authors should report raw and smoothed metrics separately.
  3. [Table 1; §3.1.2] The target variable is derived from an average of only 8.71 valid viewers per lecture after applying filters that exclude viewing records shorter than one minute and students with less than five minutes of total viewing. With so few contributing viewers, segment access frequency is highly sensitive to individual review behavior, and no stability or agreement analysis is provided. The paper itself acknowledges the small dataset, but the validity of the engagement proxy is load-bearing: if these few viewers are not representative, both the reported prediction performance and its interpretation as engagement prediction are called into question. The authors should report per-segment viewer counts or a leave-one-viewer-out label-stability analysis.
  4. [§3.3.2] The text states that "After confirming the superiority of feature-level fusion over model-level fusion in earlier experiments," all subsequent evaluations use feature-level fusion, but no such earlier experiments are reported anywhere in the manuscript; Section 4 is empty. This unsupported assertion determines the experimental design for RQ2 and RQ3 and must either be backed by reported experiments or removed. As written, this is a circular reference to results that do not appear.
  5. [§3.1.3] The three-class thresholds (high > 0.5, low < 0.2, medium in between) are arbitrary and are applied after per-lecture normalization and temporal smoothing, so the class balance and the 69.32% accuracy figure depend on these choices. No sensitivity analysis or class-distribution table is provided. Because this classification accuracy is one of only two headline results, the threshold dependence should be quantified.
minor comments (5)
  1. [§2.1] The reference to Kim et al. is rendered as "[ ?]" and should be completed.
  2. [Introduction] In the research question list, "RQ2: RQ2:" is duplicated and should be corrected.
  3. [§3.2.1] The optical flow citation [19] (Burton and Radford, 1978) appears mismatched; the Lucas-Kanade method [21] is the relevant citation. The relation between masking the students' seating area and suppressing slide corner points should also be clarified.
  4. [Table 1] The caption "Table 1 shows the statistics of valid viewers and total valid viewing time per lecture" repeats the surrounding text; a simpler descriptive caption would be clearer.
  5. [§3.2.3] The definition of Praw[i] = max{xi} - max{xi-1} is not fully specified: the set over which the maximum is taken should be stated explicitly (e.g., all slide-page values within a five-minute segment). Without this, the feature extraction is not reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the access-frequency label is an independent operationalization from viewing logs, and the only self-citation (Ref. [28]) is a non-load-bearing dataset provenance reference.

full rationale

The paper's target variable (segment access frequency) is computed directly from aggregated playback logs (Section 3.1.2), while the input features are instructor pose/optical flow, audio spectrograms, and slide progression (Sections 3.2.1-3.2.3). There is no equation in which an input is defined in terms of the label, nor any fitted parameter that is later renamed as a prediction. The label smoothing (centered 5-segment moving average) and the post-hoc prediction smoothing (Section 3.2.4) are separate operations applied to ground truth and predictions respectively; while this could inflate correlation if not disclosed, it is not a circular derivation. The only self-citation is reference [28] (Hasegawa, a coauthor) used to describe the JAIST archive's existence and recording practice; that fact is externally grounded and does not carry the paper's central claim. No uniqueness theorem or ansatz is imported from the authors' prior work. The central experimental claim (PCC=0.5143, 69.32%) is currently unsupported because Section 4 states "This section is currently under development and will be completed in a subsequent version," but this is a completeness/verifiability problem rather than a circularity problem. Therefore no circular step is exhibited.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim depends on the reliability of access-frequency labels, which are derived from a very small number of viewers using several hand-chosen filters and smoothing parameters. The modeling also assumes that the three non-semantic feature types are predictive for engagement, an assumption that remains unverified due to the missing experiment section.

free parameters (7)
  • three-class thresholds = 0.5 (high), 0.2 (low)
    Used to convert continuous predictions into high/medium/low attention zones; chosen ad hoc in Section 3.1.3 and directly determine the reported 69.32% accuracy metric.
  • label smoothing window = 5 (centered moving average)
    Applied to raw access-frequency labels in Section 3.1.2; affects all downstream regression and classification scores.
  • valid viewer filters = 1 minute per record, 5 minutes total
    Data cleaning thresholds in Section 3.1.2 that define the label set and exclude most viewing logs; no sensitivity analysis.
  • slide feature aggregation window = 5 minutes, repeated to 1-minute resolution
    The five-step pipeline in Section 3.2.3 hand-selects a 5-minute net progression window and smoothing window, shaping the slide modality input.
  • prediction post-smoothing parameters = MA window 5; SG window 7, order 2; Kalman R=0.05, Q=1e-4
    Post-hoc smoothing choices in Section 3.2.4 that alter the final predictions used for classification.
  • segment duration = 1 minute
    Fixed time interval for labels and features; arbitrary choice made in Section 3.1.2.
  • archive trimming and padding length = 95 minutes
    Standardizes all lectures to a fixed length, as described in Section 3.1.2; original lectures were about 100 minutes.
assumptions (5)
  • domain assumption Segment access frequency is a valid proxy for student engagement.
    The entire label design relies on this, stated in Section 1 and Section 3.1.2; no validation against a direct engagement measure is provided.
  • domain assumption The filtered viewing logs are reliable and representative.
    Section 3.1.2; with an average of 8.71 valid viewers per lecture, the aggregate access counts are noisy and may not generalize.
  • domain assumption Non-semantic features (instructor action, audio spectrogram, slide progression) carry sufficient predictive signal for access frequency.
    Core modeling assumption in Sections 3.2 and 3.3; not established by any experiment in this version.
  • domain assumption Standard CNN architectures trained on stacked-channel images can extract meaningful patterns from these heterogeneous feature maps.
    Section 3.3.1; the feature-stacking approach conflates three different modalities into RGB channels, yet the paper assumes this retains usable information.
  • domain assumption OpenPose pose estimates and optical flow correctly capture instructor movement in this specific camera setup.
    Section 3.2.1; ceiling-mounted camera, masked seating area, and slide overlay may degrade pose detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Non-Semantic Feature Fusion for Predicting Segment Access Frequency in Lecture Archives." pith.science (2026). https://pith.science/paper/QJBIYLXN

@misc{pith2026250414927,
  author       = {Pith},
  title        = {Pith review of: Multimodal Non-Semantic Feature Fusion for Predicting Segment Access Frequency in Lecture Archives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJBIYLXN}},
  note         = {Machine review of arXiv:2504.14927}
}
read the original abstract

This study proposes a multimodal neural network-based approach to predict segment access frequency in lecture archives. These archives, widely used as supplementary resources in modern education, often consist of long, unedited recordings that make it difficult to keep students engaged. Captured directly from face-to-face lectures without post-processing, they lack visual appeal. Meanwhile, the increasing volume of recorded material renders manual editing and annotation impractical. Automatically detecting high-engagement segments is thus crucial for improving accessibility and maintaining learning effectiveness. Our research focuses on real classroom lecture archives, characterized by unedited footage, no additional hardware (e.g., eye-tracking), and limited student numbers. We approximate student engagement using segment access frequency as a proxy. Our model integrates multimodal features from teachers' actions (via OpenPose and optical flow), audio spectrograms, and slide page progression. These features are deliberately chosen for their non-semantic nature, making the approach applicable regardless of lecture language. Experiments show that our best model achieves a Pearson correlation of 0.5143 in 7-fold cross-validation and 69.32 percent average accuracy in a downstream three-class classification task. The results, obtained with high computational efficiency and a small dataset, demonstrate the practical feasibility of our system in real-world educational contexts.

Figures

Figures reproduced from arXiv: 2504.14927 by the authors.

Figure 1
Figure 1. System pipeline for predicting segment engagement from lecture archives. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Original Lecture Archive, I239 Machine Learning. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Action Feature [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Feature fusion input example combining action, voice, and slide features into RGB [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Process of feature-level fusion (example based on ResNet). [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Process of model-level fusion (example based on ResNet). [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

  1. [1]

    Abstractive video lecture summarization: applications and future prospects

    Benedetto, I., La Quatra, M., Cagliero, L., Canale, L., and Farinetti, L. Abstractive video lecture summarization: applications and future prospects. Education and Information Tech- nologies, 29(3):2951–2971, 2024

    work page 2024

  2. [2]

    Video-based learning (VBL)—past, present and future

    Sabli´ c, M., Mirosavljevi´ c, A., andˇSkugor, A. Video-based learning (VBL)—past, present and future. Technology, Knowledge and Learning , 26(4):1061–1077, 2021

    work page 2021

  3. [3]

    Video Summarization Techniques: A Comprehensive Review

    Alaa, T., Mongy, A., Bakr, A., Diab, M., and Gomaa, W. Video Summarization Techniques: A Comprehensive Review. arXiv preprint arXiv:2410.04449 , 2024

    arXiv 2024

  4. [4]

    A review of deep learning models for time series prediction

    Han, Z., Zhao, J., Leung, H., Ma, K.F., and Wang, W. A review of deep learning models for time series prediction. IEEE Sensors Journal , 21(6):7833–7848, 2019

    work page 2019

  5. [5]

    Bayesian fuzzy clustering and deep CNN-based automatic video summarization

    Singh, A., and Kumar, M. Bayesian fuzzy clustering and deep CNN-based automatic video summarization. Multimedia Tools and Applications , 83(1):963–1000, 2024

    work page 2024

  6. [6]

    Large Model based Sequential Keyframe Extraction for Video Summarization

    Tan, K., Zhou, Y., Xia, Q., Liu, R., and Chen, Y. Large Model based Sequential Keyframe Extraction for Video Summarization. In Proceedings of the International Conference on Com- puting, Machine Learning and Data Science (CMLDS ’24) , pp. 52:1–5, 2024

    work page 2024

  7. [7]

    Deep multi-scale pyrami- dal features network for supervised video summarization

    Khan, H., Hussain, T., Khan, S.U., Khan, Z.A., and Baik, S.W. Deep multi-scale pyrami- dal features network for supervised video summarization. Expert Systems with Applications , 237:121288, 2024

    work page 2024

  8. [8]

    Distance education: Definitions, generations and key concepts and future directions

    Saykili, A. Distance education: Definitions, generations and key concepts and future directions. International Journal of Contemporary Educational Research , 5(1):2–17, 2018. 14

    work page 2018

Show all 31 references
  1. [9]

    Promoting student engagement in online education: Online learning experiences of Dutch university students.Technology, Knowledge and Learning, 29(2):941–961, 2024

    Vermeulen, E.J., and Volman, M.L.L. Promoting student engagement in online education: Online learning experiences of Dutch university students.Technology, Knowledge and Learning, 29(2):941–961, 2024

    2024

  2. [10]

    Effects of embedded questions in pre-class videos on learner perceptions, video engagement, and learning performance

    Deng, R., and Gao, Y. Effects of embedded questions in pre-class videos on learner perceptions, video engagement, and learning performance. Active Learning in Higher Education , 25(3):473– 487, 2024

    2024

  3. [11]

    Understanding in-video dropouts and interaction peaks in online lecture videos

    Kim, J., Guo, P.J., Seaton, D.T., Mitros, P., Gajos, K.Z., and Miller, R.C. Understanding in-video dropouts and interaction peaks in online lecture videos. In Proceedings of the First ACM Conference on Learning@Scale , pp. 31–40, 2014

    2014

  4. [12]

    How video production affects student engagement: An em- pirical study of MOOC videos

    Guo, P.J., Kim, J., and Rubin, R. How video production affects student engagement: An em- pirical study of MOOC videos. InProceedings of the First ACM Conference on Learning@Scale, pp. 41–50, 2014

    2014

  5. [13]

    VLEngagement: A dataset of scientific video lectures for evaluating population-based engagement

    Bulathwela, S., Perez-Ortiz, M., Yilmaz, E., and Shawe-Taylor, J. VLEngagement: A dataset of scientific video lectures for evaluating population-based engagement. arXiv preprint arXiv:2011.02273, 2020

    2011 arXiv

  6. [14]

    Context and memory in multimedia content analysis.IEEE Multimedia, 11(3):7– 11, 2004

    Dimitrova, N. Context and memory in multimedia content analysis.IEEE Multimedia, 11(3):7– 11, 2004

    2004

  7. [15]

    Soccer video summarization using deep learning

    Agyeman, R., Muhammad, R., and Choi, G.S. Soccer video summarization using deep learning. In Proc. 2019 IEEE MIPR , pp. 270–273, 2019

    2019

  8. [16]

    Automatic lecture video content summarization with attention- based recurrent neural network

    Andra, M.B., and Usagawa, T. Automatic lecture video content summarization with attention- based recurrent neural network. In Proc. 2019 ICAIIT , pp. 54–59, 2019

    2019

  9. [17]

    Effects of different video lecture types on attention, emotion, cognitive load, and learning performance

    Chen, C.M., and Wu, C.H. Effects of different video lecture types on attention, emotion, cognitive load, and learning performance. Computers & Education , 80:108–121, 2015

    2015

  10. [18]

    Effects of video instructor’s body language on students’ visual attention: An eye-tracking study

    Zhang, J., Bourguet, M.L., and Venture, G. Effects of video instructor’s body language on students’ visual attention: An eye-tracking study. In Proc. 32nd International BCS HCI Conference, pp. 1–5, 2018

    2018

  11. [19]

    Thinking in Perspective: Critical Essays in the Study of Thought Processes

    Burton, A., and Radford, J. Thinking in Perspective: Critical Essays in the Study of Thought Processes. Routledge, 1978

    1978

  12. [20]

    OpenPose: Real-time multi-person 2D pose estimation using Part Affinity Fields

    Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., and Sheikh, Y. OpenPose: Real-time multi-person 2D pose estimation using Part Affinity Fields. IEEE Transactions on PAMI , 43(1):172–186, 2019

    2019

  13. [21]

    An iterative image registration technique with an application to stereo vision

    Lucas, B.D., and Kanade, T. An iterative image registration technique with an application to stereo vision. In Proc. IJCAI, 1981

    1981

  14. [22]

    Audio spectrogram representations for processing with convolutional neural networks

    Wyse, L. Audio spectrogram representations for processing with convolutional neural networks. arXiv preprint arXiv:1706.09559 , 2017

    2017 arXiv

  15. [23]

    Speech synthesis

    Flanagan, J.L. Speech synthesis. In Speech Analysis Synthesis and Perception , pp. 204–276, Springer, 1972. 15

    1972

  16. [24]

    Eye-tracking students’ attention to Pow- erPoint photographs in science education

    Slykhuis, D.A., Wiebe, E.N., and Annetta, L.A. Eye-tracking students’ attention to Pow- erPoint photographs in science education. Journal of Science Education and Technology , 14(5):509–520, 2005

    2005

  17. [25]

    Importance of input data normalization for neural networks in industrial problems

    Sola, J., and Sevilla, J. Importance of input data normalization for neural networks in industrial problems. IEEE Trans. on Nuclear Science , 44(3):1464–1468, 1997

    1997

  18. [26]

    Smoothing and differentiation of data by simplified least squares procedures

    Savitzky, A., and Golay, M.J.E. Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36(8):1627–1639, 1964

    1964

  19. [27]

    Barriers to distance learning during the COVID-19 outbreak: A qualitative review

    Abuhammad, S. Barriers to distance learning during the COVID-19 outbreak: A qualitative review. Heliyon, 6(11), 2020

    2020

  20. [28]

    Case studies for self- directed learning using lecture archives

    Hasegawa, S., Tajima, Y., Matou, M., Futatsudera, M., and Ando, T. Case studies for self- directed learning using lecture archives. In Proc. WBE 2007 , pp. 299–304, 2007

    2007

  21. [29]

    Automated summarization of lecture videos

    Vimalaksha, A., Vinay, S., Prekash, A., and Kumar, N.S. Automated summarization of lecture videos. In Proc. IEEE T4E , pp. 126–129, 2018

    2018

  22. [30]

    The impacts of instructor’s visual attention and lecture type on learning performance

    Shi, Y., Wang, M., Chen, Z., Hou, G., Wang, Z., Zheng, Q., and Sun, J. The impacts of instructor’s visual attention and lecture type on learning performance. Education and Information Technologies, 2024

    2024

  23. [31]

    Very deep convolutional networks for large-scale image recognition

    Simonyan, K. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 16

    2014 arXiv

Pith tools

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