Pith. sign in

REVIEW 3 major objections 6 minor 298 references

Multimodal image-tabular rankings flip with the tabular encoder; a simple fusion with a strong encoder matches a heavy engineered method.

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 · grok-4.5

2026-07-10 19:00 UTC pith:BZJBWIGP

load-bearing objection Solid empirical re-ranking study: tabular encoder choice is a real confound in image-tabular fusion, and a simple bilinear baseline with strong ICL features matches a heavy engineered method. the 3 major comments →

arxiv 2607.07756 v1 pith:BZJBWIGP submitted 2026-07-08 cs.LG

The Importance of Encoder Choice:A Tabular-Image Study

Ilia Koloiarov , Diego Coello de Portugal Mecke , Vijaya Krishna Yalavarthi , Tom Hanika , Lars Schmidt-Thieme This is my paper
classification cs.LG
keywords multimodal learningtabular encodersimage-tabular fusionin-context learningrepresentation shiftfeature extractionbilinear fusion
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.

Most image-tabular pipelines still feed tables through a plain shallow MLP, even though that is known to be a weak standalone tabular model. This paper instead freezes modern state-of-the-art tabular models as encoders and fuses their embeddings with a fixed image encoder. The central finding is that method rankings are unstable across encoders: a carefully engineered multimodal pipeline can look superior with a weak table encoder and be matched by a simple bilinear product once the table encoder is strong. In-context learning tabular foundation models, which need labels for context, also produce a context-query representation shift that breaks the i.i.d. assumption for any downstream head; the authors quantify the shift and show that non-partitioned query extraction largely removes it. The practical message is that encoder choice is a first-order confound in multimodal benchmarks, not an implementation detail.

Core claim

On likely-multimodal image-tabular datasets, multimodal rankings are not stable across tabular encoders, so conclusions drawn from a single encoder do not generalize. A simple bilinear fusion paired with a strong tabular encoder (especially non-partitioned extraction from TabPFN) performs on par with a carefully engineered multimodal method that uses on average 13.2 times more parameters and a dedicated pretraining stage. The observed fusion lift itself shrinks as the unimodal tabular baseline improves, and the context-query representation shift of in-context learning tabular models is not limited to TabPFN.

What carries the argument

Three ICL feature-extraction schemes (Vanilla context-role, leave-one-fold-out query-role, and non-partitioned query-role on the full training set) plus Excess Distribution Shift (EDS), an MMD-based gap that isolates context-query divergence from ordinary data split variation; these let the authors place train and test embeddings in a shared latent space and measure whether that space is usable for fusion.

Load-bearing premise

The paper assumes that a stronger unimodal tabular predictor also yields better frozen representations for multimodal fusion; if predictive accuracy and representation quality diverge, the encoder-sensitivity story collapses.

What would settle it

Re-run the same bilinear-versus-TIP comparison on the same likely-multimodal datasets with an additional strong tabular encoder whose unimodal F1 is high but whose frozen embeddings yield no multimodal lift, or show that Vanilla extraction no longer degrades fusion once a different image encoder or fusion head is used.

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

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 / 6 minor

Summary. The paper argues that tabular encoder choice is a critical and previously neglected confound in image-tabular multimodal learning. Prior work largely used plain MLPs; the authors systematically replace them with SOTA tabular models (TabPFNv2, TabDPT, TabICLv2 under three extraction schemes, TARTE, TabM/TabM-SSL, TabVec) while freezing a ViT-B/16 image encoder and comparing a simple bilinear fusion baseline against the engineered TIP pipeline and unimodal linear heads. They introduce Relative Percentile Rank (RPR) and Excess Distribution Shift (EDS via multi-bandwidth MMD) and, on seven datasets, show that (i) method rankings are unstable across encoders, (ii) fusion can hurt on modality-dominated data, (iii) the multimodal lift shrinks with stronger unimodal tabular performance (OLS slope β̂=0.434 on likely-multimodal sets), (iv) bilinear+strong encoder matches TIP at ~13 imes fewer parameters, and (v) the context-query representation shift of ICL TFMs is not limited to TabPFNv2 and is best mitigated by non-partitioned (NP) extraction.

Significance. If the empirical pattern holds, it is a useful methodological correction for multimodal tabular-image work: many published fusion gains may be partly encoder artifacts, and simple bilinear fusion with a modern tabular foundation model can be competitive with carefully engineered pipelines. The paper also supplies the first multi-model quantification of the ICL context-query shift (EDS, Table 2, Fig. 5) and a practical recommendation (prefer NP over Vanilla). Strengths include nested 4-inner/2-outer CV, TPE HPO, 95% CIs, critical-difference diagrams, an explicit OLS null of constant lift (β=1), and released code/per-trial logs. The contribution is primarily empirical and cautionary rather than a new architecture; its value lies in re-calibrating how the community benchmarks multimodal methods when a tabular modality is present.

major comments (3)
  1. [Introduction / §4.3] The central premise stated in the Introduction—that a stronger unimodal tabular predictor necessarily yields better frozen representations for multimodal fusion—is used to motivate the encoder sweep but is never tested directly. Representation quality and predictive accuracy can diverge (e.g., via label leakage into the embedding or via features that are predictive alone but redundant with the image). A short diagnostic that correlates unimodal F1 with a pure representation metric (linear probe on held-out tabular labels after freezing, or mutual information with the target after residualizing the image) would make the premise falsifiable and would strengthen the interpretation of the OLS slope in §4.3 / Fig. 3.
  2. [§3 / §5 Limitations] All image representations come from a single frozen ViT-B/16 CLS token (§3). The authors correctly flag this in Limitations, yet the dataset typology (image-dominated / tabular-dominated / likely multimodal) and the claim that fusion lift shrinks with encoder quality are both conditioned on this choice. A stronger or architecturally different image encoder could reclassify datasets and alter the OLS slope. At minimum the paper should report a sensitivity check on one likely-multimodal dataset with a second image backbone (e.g., ResNet-50 or a larger ViT) so that the encoder-sensitivity conclusion is not itself image-encoder-specific.
  3. [§4.3 / Fig. 3] The OLS analysis in §4.3 treats each (encoder, dataset) point as independent when fitting Y=βX+b on likely-multimodal data and reports a one-sided p=2.5e-57 for β=1. With only three datasets and multiple extraction schemes of the same ICL model, the effective sample size is small and points are correlated. A mixed-effects or clustered bootstrap that respects dataset and model family would give a more credible confidence interval on β and on the claimed 0.566 reduction in multimodal advantage per unit unimodal F1.
minor comments (6)
  1. [Abstract] Abstract and Introduction list five numbered findings while the body has five correspondingly numbered subsections; the abstract’s fifth finding is the ICL extraction recommendation, which is fine, but the numbering in the abstract jumps from 4 to 5 without an explicit “5.” in the printed text of some renderings—please check consistency.
  2. [Fig. 2] Figure 2 is dense (filled vs hatched bars, two CI styles, many encoders). A short legend sentence in the caption that explicitly maps star/dagger/wrenches to Vanilla/LOFO/NP would help readers who do not keep §2.3 open.
  3. [§3.1 / Eq. (4)] Eq. (4) for bilinear fusion uses W∈R^{d imes C imes d}; a one-line note that this is the standard multi-class bilinear form of Jayakumar et al. (and that setting one modality to zero recovers the unimodal linear head) would make the reduction claim fully self-contained.
  4. [Table 1] Table 1 reports image size quantiles as √(W imes H); the notation “Img √W imes H” is slightly ambiguous at first glance—consider “√(area)” or “side length of equivalent square”.
  5. [§4.1] In §4.1 the win/tie/loss counts for TIP vs baselines are given as (W/T/L) without stating whether ties are defined by overlapping 95% CIs or by exact F1 equality; a one-sentence definition would remove ambiguity.
  6. [References / header] References: TabICLv2 is cited as Qu et al. 2026 in the text and 2025 in the bibliography in places; unify the year. Also, the arXiv ID in the header (2607.07756) is future-dated relative to the content—confirm the intended citation key.

Circularity Check

0 steps flagged

No significant circularity: pure empirical encoder sweep with independently defined metrics (RPR, EDS, F1 ranks, OLS slope) that do not reduce to inputs by construction.

full rationale

The paper is an empirical study of tabular encoder choice in image-tabular fusion. Its load-bearing claims (encoder-dependent re-ranking of TIP vs. bilinear, fusion lift shrinking with unimodal strength, context-query shift across ICL TFMs) are established by nested CV F1, RPR relative to a fixed image baseline, EDS (MMD gaps), critical-difference diagrams, and an OLS regression of multimodal vs. unimodal F1 whose null is the external hypothesis eta=1. None of these quantities is defined in terms of the result being claimed, fitted to the target and then re-presented as a prediction, or forced by a self-citation uniqueness theorem. Extraction schemes (Vanilla/LOFO/NP) follow Ye et al. and are evaluated, not assumed. Self-citations are ordinary background (TabPFN, TabDPT, TIP, TabM) and do not underwrite the encoder-sensitivity finding. The study is therefore self-contained against its own experimental protocol; circularity score is zero.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard experimental assumptions of multimodal learning plus two paper-specific modeling choices (frozen encoders + linear projection, and the bilinear fusion as the naïve baseline). Free parameters are the usual tuned hyperparameters; no new physical constants or ad-hoc scales are introduced. Invented entities are the two evaluation metrics RPR and EDS, both operationally defined and falsifiable on any embedding set.

free parameters (3)
  • shared embedding dimension d = selected per trial from the discrete set
    Chosen from {192,256,512,768} by average inner-fold F1; directly affects parameter count and fusion capacity.
  • learning rate, weight decay, label smoothing, dropout rates = per-dataset, per-encoder via TPE
    TPE-tuned over 50 trials; standard but free parameters that influence absolute F1 numbers used in rankings.
  • LOFO fold count K = 10
    Fixed to 10 following prior work; changes the amount of context available under the leave-one-fold-out scheme.
axioms (4)
  • ad hoc to paper A stronger unimodal tabular predictor yields better frozen representations for multimodal fusion.
    Stated in the introduction and used to justify sweeping SOTA tabular models; not independently verified outside the paper’s own rankings.
  • domain assumption Frozen pretrained encoders plus a linear LayerNorm projection isolate the effect of the downstream fusion module.
    Standard practice in multimodal ablations; invoked in Section 3 to keep unimodal and multimodal baselines comparable.
  • domain assumption Bilinear product (Jayakumar et al.) is a fair naïve multimodal baseline that reduces to the unimodal linear classifier when one modality is zero.
    Used throughout Section 3.1 and 4 to claim that simple fusion matches TIP.
  • standard math Maximum Mean Discrepancy with multi-bandwidth RBF kernel quantifies representation shift beyond ordinary train-val variation.
    Gretton et al. MMD is standard; the excess construction (EDS) is a straightforward difference of two MMD terms.
invented entities (2)
  • Relative Percentile Rank (RPR) independent evidence
    purpose: Normalize F1 differences across datasets by ranking methods relative to the image-only baseline.
    Defined in Section 3.4; purely evaluative, no ontological claim.
  • Excess Distribution Shift (EDS) independent evidence
    purpose: Measure how much train embeddings deviate from val/test beyond ordinary held-out variation.
    Defined via MMD differences in Eq. 5; used to quantify context-query shift.

pith-pipeline@v1.1.0-grok45 · 19654 in / 2927 out tokens · 33671 ms · 2026-07-10T19:00:43.943501+00:00 · methodology

0 comments
read the original abstract

Multimodal learning usually requires a dedicated encoder per modality. When a tabular modality is involved, prior work has been mostly using a \emph{plain MLP} as the encoder. Yet if it were a strong encoder, the tabular domain would not be ``the last unconquered castle for deep learning''. This study evaluates state-of-the-art tabular models as encoders in the image-tabular setting for the first time. An obstacle stands out. In-Context Learning models, among the best performing methods in the tabular domain, require labels to process instances, making it non-trivial to embed training and test instances the same way. We addressed this problem across multiple models of this family. With this study, we would like to highlight the importance of encoder factor in the multimodal learning.

Figures

Figures reproduced from arXiv: 2607.07756 by Diego Coello de Portugal Mecke, Ilia Koloiarov, Lars Schmidt-Thieme, Tom Hanika, Vijaya Krishna Yalavarthi.

Figure 1
Figure 1. Figure 1: Context (dark circles) and query (bright triangles) encodings of the same instance (connected by gray lines) do not coincide in latent space. 2D PCA of TabPFNv2 on DVM. The neglect of the tabular encoder has a historical explanation. For years, gradient-boosted decision trees (GBDTs) dominated tabular benchmarks [Grinsztajn et al., 2022, Borisov et al., 2022, Shwartz-Ziv and Armon, 2022], leaving no compel… view at source ↗
Figure 2
Figure 2. Figure 2: RPR ↑. See the section 3.4 for details. Filled bars show multimodal performance (color indicates tabular encoder). Hatched bars show the corresponding unimodal tabular performance on the same tabular encoder. Error bars (uncapped gray for multimodal, capped black for unimodal tabular) indicate 95% confidence intervals. ICL features extractions are ⋆ = Vanilla, † = LOFO, and ≀ = NP. 4.1 Evaluation on a sole… view at source ↗
Figure 3
Figure 3. Figure 3: Multimodal F1 against Unimodal F1. Left: Likely multimodal datasets. Right: Tabular￾dominated multimodal datasets. The dashed line shows the OLS fit with its 95% confidence band (reflecting uncertainty in the estimated slope and intercept), the dotted line shows constant gain with a slope of 1, encoder-independent lift. Colors are tabular embeddings following the same coding as on [PITH_FULL_IMAGE:figures… view at source ↗
Figure 4
Figure 4. Figure 4: Critical Difference Diagrams on F1, where horizontal bar indicates the absence of statistical [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Left: EDS(Dtr, Dval) in blue and EDS(Dtr, Dte) in green (Eq. 5). Boxplots are aggregated over datasets over each encoder-scheme pair. Positive values exceed data-intrinsic variation. Right: 2D PCA on COVID (rows top to bottom: Vanilla(⋆), LOFO(†), NP(≀); columns left to right: TabPFN, TabICL, TabDPT). Dark circles and light triangles/stars are train and val/test respectively. Color reflects class. baseline… view at source ↗
Figure 6
Figure 6. Figure 6: CDD on Multimodal Performance for all datasets [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: DVM (left), Wikiart (center), and HAM10000 (right). [PITH_FULL_IMAGE:figures/full_fig_p014_7.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

298 extracted references · 298 canonical work pages · 48 internal anchors

  1. [1]

    [Neo4j-Tut] Jimwebber/Neo4j-Tutorial -

    work page

  2. [2]

    SlidesLive , urldate =

    39024579. SlidesLive , urldate =

    work page

  3. [3]

    Adaptive Boosting for Transfer Learning Using Dynamic Updates

    work page

  4. [4]

    Google Docs , urldate =

    work page

  5. [5]

    Bayesian Dark Knowledge , urldate =

    work page

  6. [6]

    ACM Conferences , doi =

    work page

  7. [7]

    SIGKDD - KDD 2018 , urldate =

    Calibrated. SIGKDD - KDD 2018 , urldate =

    work page 2018

  8. [8]

    www.shortscience.org , urldate =

    Deciding. www.shortscience.org , urldate =

    work page

  9. [9]

    Desmos , urldate =

    Desmos. Desmos , urldate =

    work page

  10. [10]

    Diagrams.Net , urldate =

    work page

  11. [11]

    EUR-Lex - 31995L0046 - EN , urldate =

    Directive 95/46/. EUR-Lex - 31995L0046 - EN , urldate =

    work page

  12. [12]

    Hella Provides Lane-Change Warning for

    work page

  13. [13]

    Index of File:///Home/Lst/Books/ , urldate =

    work page

  14. [14]

    An Integrated Growing-Pruning Method for Feedforward Network Training -

    work page

  15. [15]

    Jimwebber/Neo4j-Tutorial -

    work page

  16. [16]

    Log-Based Predictive Maintenance , urldate =

    work page

  17. [17]

    MathOverflow , urldate =

    Markov Chains -. MathOverflow , urldate =

    work page

  18. [18]

    SIGKDD - KDD 2018 , urldate =

    Multi-. SIGKDD - KDD 2018 , urldate =

    work page 2018

  19. [19]

    GraphLab - CMU , urldate =

    [Pmf]. GraphLab - CMU , urldate =

    work page

  20. [20]

    Random Permutations Fix a Worst Case for Cyclic Coordinate Descent

    work page

  21. [21]

    Chemical & Engineering News , urldate =

    Simulations. Chemical & Engineering News , urldate =

    work page

  22. [22]

    Teaching , urldate =

    work page

  23. [23]

    Ukraine Dekret 117 -

    work page

  24. [24]

    Zotero - Ismllwiki , urldate =

    work page

  25. [25]

    doi:10.1109/ICADIWT.2008.4664413 , urldate =

    Realizing Process Interoperability Using. doi:10.1109/ICADIWT.2008.4664413 , urldate =

    work page doi:10.1109/icadiwt.2008.4664413 2008

  26. [26]

    Statistical

    Wiley. Statistical. doi:10.1002/9781119013563.scard , urldate =

    work page doi:10.1002/9781119013563.scard

  27. [27]

    Partial Sorting , year = 2013, month = may, journal =

    work page 2013

  28. [28]

    Neurips , annotation =

    Supervised. Neurips , annotation =

    work page

  29. [29]

    ICML , annotation =

    The. ICML , annotation =

    work page

  30. [30]

    SAIL Blog , urldate =

    Extrapolating to. SAIL Blog , urldate =

    work page

  31. [31]

    Bundesministerium f\"ur Bildung und Forschung - BMBF , urldate =

    work page

  32. [32]

    Price Elasticity of Demand , year = 2025, month = mar, journal =

    work page 2025

  33. [33]

    SlidesLive , urldate =

    39049897. SlidesLive , urldate =

    work page

  34. [34]

    User Modeling and User-Adapted Interaction , volume =

    Modeling. User Modeling and User-Adapted Interaction , volume =

    work page

  35. [35]

    User Modeling and User-Adapted Interaction , volume =

    Knowledge. User Modeling and User-Adapted Interaction , volume =

    work page

  36. [36]

    Aalst, Wil M. P. and Song, Minseok , editor =. Mining. Business

    work page

  37. [37]

    Brendan and Mironov, Ilya and Talwar, Kunal and Zhang, Li , year = 2016, series =

    Abadi, Martin and Chu, Andy and Goodfellow, Ian and McMahan, H. Brendan and Mironov, Ilya and Talwar, Kunal and Zhang, Li , year = 2016, series =. Deep. Proceedings of the 2016. doi:10.1145/2976749.2978318 , urldate =

    work page doi:10.1145/2976749.2978318 2016

  38. [38]

    TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems

    Abadi, Mart. Tensorflow:. arXiv preprint arXiv:1603.04467 , eprint =

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Large Scale Tag Recommendation Using Different Image Representations , booktitle =

    Abbasi, Rabeeh and Grzegorzek, Marcin and Staab, Steffen , year = 2009, pages =. Large Scale Tag Recommendation Using Different Image Representations , booktitle =

    work page 2009

  40. [40]

    Rajendra and Makarenkov, Vladimir and Nahavandi, Saeid , year = 2021, month = dec, journal =

    Abdar, Moloud and Pourpanah, Farhad and Hussain, Sadiq and Rezazadegan, Dana and Liu, Li and Ghavamzadeh, Mohammad and Fieguth, Paul and Cao, Xiaochun and Khosravi, Abbas and Acharya, U. Rajendra and Makarenkov, Vladimir and Nahavandi, Saeid , year = 2021, month = dec, journal =. A Review of Uncertainty Quantification in Deep Learning:. doi:10.1016/j.inff...

    work page doi:10.1016/j.inffus.2021.05.008 2021

  41. [41]

    Adaptive Optimization of Hyperparameters in

    work page

  42. [42]

    Sparse-posterior Gaussian Processes for general likelihoods

    Sparse-Posterior. arXiv preprint arXiv:1203.3507 , eprint =

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Applying Convolutional Neural Networks Concepts to Hybrid. 2012

    work page 2012

  44. [44]

    , booktitle =

    Exploring Convolutional Neural Network Structures and Optimization Techniques for Speech Recognition. , booktitle =

    work page

  45. [45]

    IEEE/ACM Transactions on audio, speech, and language processing , volume =

    Convolutional Neural Networks for Speech Recognition , author =. IEEE/ACM Transactions on audio, speech, and language processing , volume =

    work page

  46. [46]

    Abdelmalak, Ibram and Madhusudhanan, Kiran and Choi, Jungmin and Kl. Channel. Advances in. doi:10.1007/978-981-92-1462-4_46 , urldate =

    work page doi:10.1007/978-981-92-1462-4_46

  47. [47]

    and Venet, P

    Abdennadher, K. and Venet, P. and Rojat, G. and Retif, J. and Rosset, C. , year = 2010, month = jul, journal =. A. doi:10.1109/TIA.2010.2049972 , abstract =

    work page doi:10.1109/tia.2010.2049972 2010

  48. [48]

    Explainable Matrix Factorization for Collaborative Filtering , booktitle =

    Abdollahi, Behnoush and Nasraoui, Olfa , year = 2016, pages =. Explainable Matrix Factorization for Collaborative Filtering , booktitle =

    work page 2016

  49. [49]

    Using Explainability for Constrained Matrix Factorization , booktitle =

    Abdollahi, Behnoush and Nasraoui, Olfa , year = 2017, pages =. Using Explainability for Constrained Matrix Factorization , booktitle =

    work page 2017

  50. [50]

    Advances in Neural Information Processing Systems , volume =

    A Definition of Continual Reinforcement Learning , author =. Advances in Neural Information Processing Systems , volume =

    work page

  51. [51]

    Abid, Lobna and Masmoudi, Afif and Ghorbel, Sonia Zouari , year = 2016, journal =. The. doi:10.18488/journal.aefr/2016.6.1/102.1.27.42 , urldate =

    work page doi:10.18488/journal.aefr/2016.6.1/102.1.27.42 2016

  52. [52]

    Learning a

    Abid, Abubakar and Zou, James Y , editor =. Learning a. Advances in

    work page

  53. [53]

    Abouzeid, Azza and Pawlikowski, Kamil Bajda and Abadi, Daneil , year = 2009, month = aug, journal =

    work page 2009

  54. [54]

    Abouzeid, A. and. Proceedings of the VLDB Endowment , volume =

    work page

  55. [55]

    DNF-Net: A Neural Architecture for Tabular Data

    Abutbul, Ami and Elidan, Gal and Katzir, Liran and. doi:10.48550/arXiv.2006.06465 , urldate =. arXiv , keywords =:2006.06465 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2006.06465 2006

  56. [56]

    and Dunlavy, D

    Acar, E. and Dunlavy, D. M. and Kolda, T. G. , year = 2009, pages =. Link Prediction on Evolving Data Using Matrix and Tensor Factorizations , booktitle =

    work page 2009

  57. [57]

    All-at-once Optimization for Coupled Matrix and Tensor Factorizations

    Acar, Evrim and Kolda, Tamara G. and Dunlavy, Daniel M. , year = 2011, month = may, journal =. All-at-Once. arXiv , keywords =:1105.3422 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  58. [58]

    and Vigo, D

    Accorsi, L. and Vigo, D. , year = 2021, urldate =. A

    work page 2021

  59. [59]

    International Journal of Production Research , volume =

    Solving Inventory Routing with Transshipment and Substitution under Dynamic and Stochastic Demands Using Genetic Algorithm and Deep Reinforcement Learning , author =. International Journal of Production Research , volume =. doi:10.1080/00207543.2021.1987549 , urldate =

    work page doi:10.1080/00207543.2021.1987549 2021

  60. [60]

    Acharya, Deepak Bhaskar and Kuppan, Karthigeyan and Bhaskaracharya, Divya , year = 2025, journal =. Agentic. doi:10.1109/ACCESS.2025.3532853 , file =

    work page doi:10.1109/access.2025.3532853 2025

  61. [61]

    GPT-4 Technical Report

    Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and Aleman, Florencia Leoni and Almeida, Diogo and Altenschmidt, Janko and Altman, Sam and Anadkat, Shyamal and Avila, Red and Babuschkin, Igor and Balaji, Suchir and Balcom, Valerie and Baltescu, Paul and Bao, Haiming and Bavarian, Mohammad and Belgum, Jeff and Bello, I...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774

  62. [62]

    and Gomes, Carla P

    Acikalin, Utku Umur and Ferber, Aaron M. and Gomes, Carla P. , year = 2024, month = oct, urldate =. Learning to. The

    work page 2024

  63. [63]

    Information Diffusion in Online Social Networks , author =

    work page

  64. [64]

    and Bershad, Brian N

    Adar, Eytan and Weld, Daniel S. and Bershad, Brian N. and Gribble, Steven S. , year = 2007, pages =. Why We Search , booktitle =. doi:10.1145/1242572.1242595 , urldate =

    work page doi:10.1145/1242572.1242595 2007

  65. [65]

    Multichannel Sound Event Detection Using 3D Convolutional Neural Networks for Learning Inter-channel Features

    Adavanne, Sharath and Politis, Archontis and Virtanen, Tuomas , year = 2018, month = jan, journal =. Multichannel. arXiv , keywords =:1801.09522 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  66. [66]

    Addanki, Ravichandra and Nair, Vinod and Alizadeh, Mohammad , langid =. Neural

    work page

  67. [67]

    Unpublished doctoral dissertation, Global Optimization Laboratory, University of Florence , urldate =

    Global Optimization Using Local Searches , author =. Unpublished doctoral dissertation, Global Optimization Laboratory, University of Florence , urldate =

    work page

  68. [68]

    and Tuzhilin, A

    Adomavicius, G. and Tuzhilin, A. , year = 2001, journal =. Multidimensional Recommender Systems: A Data Warehousing Approach , shorttitle =

    work page 2001

  69. [69]

    Computer , volume =

    Using Data Mining Methods to Build Customer Profiles , author =. Computer , volume =. doi:10.1109/2.901170 , urldate =

    work page doi:10.1109/2.901170

  70. [70]

    ACM Transactions on Information Systems (TOIS) , volume =

    Incorporating Contextual Information in Recommender Systems Using a Multidimensional Approach , author =. ACM Transactions on Information Systems (TOIS) , volume =

    work page

  71. [71]

    and Tuzhilin, A

    Adomavicius, G. and Tuzhilin, A. , year = 2005, journal =. Toward the next Generation of Recommender Systems:

    work page 2005

  72. [72]

    IEEE Intelligent Systems , pages =

    New Recommendation Techniques for Multicriteria Rating Systems , author =. IEEE Intelligent Systems , pages =

    work page

  73. [73]

    and Tuzhilin, A

    Adomavicius, G. and Tuzhilin, A. , year = 2008, volume =. Context-Aware Recommender Systems , booktitle =

    work page 2008

  74. [74]

    Recommender Systems Handbook , pages =

    Multi-Criteria Recommender Systems , author =. Recommender Systems Handbook , pages =

    work page

  75. [75]

    Efficient

    Adriaensen, Steven and Rakotoarison, Herilalaina and M. Efficient. Advances in Neural Information Processing Systems , volume =

    work page

  76. [76]

    Proceedings of the VLDB Endowment , volume =

    Building a High-Level Dataflow System on Top of. Proceedings of the VLDB Endowment , volume =

    work page

  77. [77]

    and Ullman, Jeffrey D

    Afrati, Foto N. and Ullman, Jeffrey D. , year = 2010, pages =. Optimizing Joins in a Map-Reduce Environment , booktitle =

    work page 2010

  78. [78]

    doi:10.5445/KSP/1000058749/23 , urldate =

    Aga, Rosa Tsegaye and Wartena, Christian and Lange, Otto and Aders, Nelleke , year = 2017, journal =. doi:10.5445/KSP/1000058749/23 , urldate =

    work page doi:10.5445/ksp/1000058749/23 2017

  79. [79]

    Explore/

    Agarwal, Deepak and Chen, Bee-Chung and Elango, Pradheep , year = 2009, month = dec, pages =. Explore/. Proceedings of. doi:10.1109/ICDM.2009.52 , urldate =

    work page doi:10.1109/icdm.2009.52 2009

  80. [80]

    and Chen, B

    Agarwal, D. and Chen, B. C , year = 2009, pages =. Regression-Based Latent Factor Models , booktitle =

    work page 2009

Showing first 80 references.