Pith. sign in

REVIEW 3 major objections 5 minor 63 references

RamanPFN claims that re-coding each Raman spectrum as global composition plus local vibration before a frozen tabular model reduces regression error by 19.6% and classification error by 9.0%, because feature coverage alone is not context.

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 13:33 UTC pith:X6XGNAYE

load-bearing objection Credible benchmark result for frozen TabPFN on Raman spectra; needs PCA control and code to nail the mechanism. the 3 major comments →

arxiv 2608.02157 v1 pith:X6XGNAYE submitted 2026-08-03 cs.LG cs.AI

RamanPFN: learning from Raman spectral structure with a tabular foundation model

Xingyu Pan , Huan Wang , Jinjia Guo , Zhenlin Zhao , Siming Dong , Jixi Lu This is my paper
classification cs.LG cs.AI
keywords Raman spectroscopytabular foundation modelsin-context learningnon-negative matrix factorizationspectral representationsmall-sample predictionvibrational spectroscopyfeature subsampling
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 argues that tabular foundation models are already strong base learners for Raman prediction, but their handling of very wide inputs creates a specific failure: every wavenumber gets seen across multiple feature-subsampled forward passes, while related bands rarely get seen together. Its diagnosis is that coverage is not context. To restore context, RamanPFN replaces raw intensities with two coordinate systems fitted on the training spectra—non-negative full-spectrum components that let distant bands share one axis, and multi-mode encodings of contiguous wavenumber windows that keep local peak shape, intensity, and shifts—and feeds both to the same frozen model, combining the outputs with a signed triplet rule. Across 150 tasks from 74 public datasets, the paper reports a 19.6% average reduction in root-mean-square error on 129 regression targets relative to the frozen baseline and a 9.0% reduction in the remaining classification error on 21 tasks. If true, the practical payoff is large: small-sample Raman analysis can gain substantially without training a new spectral network, and the same input-coordinate strategy may transfer to other ordered measurement modalities.

Core claim

The paper's central claim is that the reason a tabular foundation model underperforms on Raman spectra is not lack of coverage but lack of context. The reference model's balanced feature subsampling ensures each wavenumber appears in some forward pass, yet the probability that two or three related wavenumbers appear together falls with spectral width: at the median width of 1,901 channels, 45.4% of random pairs and 13.6% of random triplets co-occur; in the five widest tasks (11,567–11,689 channels) only 4.4% of pairs and 0.19% of triplets do. RamanPFN replaces raw intensities with two coordinate systems—Global Compositional Unmixing, whose non-negative components span the full wavenumber axi

What carries the argument

The load-bearing object is a dual-scale input representation fitted from training spectra and applied to test spectra before inference. Global Compositional Unmixing (GCU) is non-negative matrix factorization of the full spectrum: each basis component is defined across the complete wavenumber axis, so separated bands driven by a shared latent source land on the same abundance-like coordinate, and the factorization trajectory acts as an implicit regularizer. Local Vibrational Subspace Encoding (LVSE) splits the ordered axis into contiguous regions and retains several singular modes per region, so independent changes in peak intensity, shape, shoulders, and small shifts are preserved within a

Load-bearing premise

The load-bearing premise is that the representation fitted on each training split continues to help on the data the model will meet, because the evaluation uses random within-dataset splits and does not test cross-instrument transfer, batch drift, calibration extrapolation, or prospective deployment.

What would settle it

Run a train-on-instrument-A, test-on-instrument-B split on any dataset with multi-instrument metadata. If the 19.6% average RMSE reduction against the frozen baseline shrinks toward zero or reverses on held-out instruments, the in-distribution gain does not survive real Raman practice. A simpler first check is to re-evaluate with spectra recorded weeks apart and compare the GCU/LVSE gain with the same-split result.

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

If this is right

  • On Raman datasets with few labelled spectra and thousands of wavenumbers, re-expressing spectra as GCU/LVSE coordinates yields a 19.6% average RMSE reduction relative to feeding raw channels to the same frozen model.
  • The regression gain is broad rather than driven by a few targets: 109 of 129 targets and 49 of 53 source datasets improve, and the method beats the best of 26 rerun baselines on 93 targets.
  • The two representations carry distinct information: 73 tasks improve with both, 20 only with GCU, 24 only with LVSE, and only 12 improve with neither; combining the two branches beats either alone.
  • The same configuration ordering appears in classification, supporting the interpretation that the gain lives in the spectral representation rather than in a regression-specific trick.
  • The authors argue the same principle—explicit local structure plus long-range covariation—may extend to near-infrared, infrared, hyperspectral, chromatography, and diffraction data, and they list cross-modality transfer as an open question.

Where Pith is reading between the lines

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

  • If the gain comes mainly from restoring joint visibility, then applying the same dual-scale encoding to a tabular foundation model with a wider context window should shrink the improvement: as joint visibility rises, the representation's advantage should fall. This is a testable consequence the paper does not run.
  • The paper's own open questions—cross-instrument transfer, batch drift, calibration extrapolation, and prospective deployment—are the natural falsification window, because GCU/LVSE maps are fitted per training split and a deployment spectrum from a different instrument may not project onto the same components.
  • A natural next step is to couple GCU coordinates to reference spectra or molecular constraints; if the predictive axes then acquire chemical identity, the representation becomes explanatory rather than merely predictive.
  • The 'coverage is not context' principle likely generalizes to other ordered signals with local and long-range structure, but the paper's evidence only establishes it for Raman; tests on near-infrared or hyperspectral data would settle the scope.

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 proposes RamanPFN, a two-branch input representation for feeding Raman spectra to a frozen TabPFN model. Global Compositional Unmixing (GCU) fits a non-negative matrix factorization on training spectra and uses sample coordinates over full-spectrum components; Local Vibrational Subspace Encoding (LVSE) partitions the wavenumber axis into contiguous regions and keeps several SVD modes per region. GCU and LVSE predictions are combined by a signed triplet rule with per-task OOF selection of candidates and weights. On 129 regression tasks from 74 public Raman datasets, the full framework is reported to reduce mean RMSE by 19.6% relative to direct TabPFN inference; on 21 classification tasks it reduces remaining weighted-F1 error by 9.0%. The paper frames the gain as evidence that 'coverage is not context': full marginal coverage of wide spectra does not preserve joint visibility of related bands, and dual-scale representation restores it.

Significance. If the empirical claims hold, the contribution is useful: it shows that a frozen tabular foundation model can be made substantially more accurate on high-dimensional Raman spectra by changing input coordinates only. Strengths include unusually broad evaluation (150 tasks, 74 datasets, 3 seeds, no task filtering), bootstrapped confidence intervals, Holm-corrected 26 contrasts, controlled five-configuration ablations, and transparent handling of the four four-spectrum classification tasks. The mechanistic visualizations and task-level complementarity analysis are also valuable. However, the central attribution to 'explicit spectral structure' rather than generic compression is not yet established because the missing unsupervised-compression control is absent, and the baseline numbers are not fully reconciled. The paper's significance is therefore conditional on the resolution of these two issues.

major comments (3)
  1. [Complementary global and local spectral structure / Table 3] The central claim that explicit dual-scale spectral structure—not mere compression—causes the gain is not established because no control feeds TabPFN a comparably sized generic unsupervised compression (e.g., PCA, random projection, or autoencoder fit on training spectra). GCU yields ~6 coordinates and LVSE ~124 (31 regions × 4 modes), both below TabPFN's 500-feature cap, whereas the raw baseline requires multiple 500-feature estimator views. The Table 3 ablation shows GCU and LVSE each improve over the raw baseline and combine, but this does not rule out the alternative that any dimensionality reduction producing a single-pass, <500-feature view would recover most or all of the 19.6% benefit. If that were true, the conclusion would be 'compression restores context,' not 'explicit spectral representation works.' Please add PCA/random-projection/autoencoder + TabPFN controls under the sam
  2. [Table 3 vs Table 2] The ablation baseline in Table 3 has mean RMSE 9.259, but Table 2 reports TabPFN v2 and v2.5 mean RMSE 9.574 and 9.621 over the same 129 regression tasks. All reported reductions (10.2%, 10.7%, 13.3%, 19.6%) are relative to the Table 3 baseline. It is unclear which configuration is the 'direct TabPFN inference' used in the abstract: if the baseline differs from the external TabPFN runs, the claim of a 19.6% reduction over 'direct TabPFN inference' is not directly backed by Table 2. Please state the exact preprocessing, estimator count, and seed handling for the Table 3 baseline, and reconcile the discrepancy or explicitly identify which Table 2 method corresponds to the baseline.
  3. [Methods, Dual-scale prediction integration / Discussion] The Discussion states that RamanPFN was evaluated 'without dataset-specific hyperparameter search,' but the Methods specify that 'protocol-matched OOF predictions rank the GCU and LVSE candidates' and that α is determined analytically from the training-target scale, with λ ∈ {1, 0.75, 0.5}. This is a per-task model-selection procedure using training labels. It is not leakage, but the 'no search' wording is misleading and the reader cannot tell how many portfolio elements are considered or how often each is selected. Please report the portfolio size, selection frequencies, and a sensitivity check with fixed α/λ (or fixed rank/partition).
minor comments (5)
  1. [General formatting] There are spacing artifacts in the headers: 'T ask' in Table 1b and 'W eighted' in Table 5. Please clean these up.
  2. [Methods, RamanPFN formulation] The exact content of the multiresolution portfolios (ranks ρ, factorization checkpoint t, partition size S, retained modes k_s) is not specified in the paper. Provide at least the ranges considered, or a persistent link to the code that fixes these values.
  3. [Methods, Dual-scale prediction integration] The role-assignment policy for the signed triplet is described only verbally. Include a precise algorithm or pseudocode specifying how the negative, local, and anchor roles are assigned from OOF evidence and how λ is chosen from {1, 0.75, 0.5}.
  4. [Figure 2a] The co-occurrence analysis samples random pairs and triples. To connect more directly to the proposed representations, it would be informative to also report co-occurrence for pairs/triples weighted by GCU loadings or by high-variance regions, rather than only uniformly random channels.
  5. [Implementation and reproducibility] The text says 'The exact numerical configuration is retained in machine-readable form with the implementation,' but no repository URL or access information is given. Please include one.

Circularity Check

0 steps flagged

No circularity: GCU/LVSE representations are fitted without target information, TabPFN weights are frozen, and all headline numbers are measured on held-out test partitions; there is no load-bearing self-citation chain.

full rationale

RamanPFN's derivation chain is self-contained with respect to the reported predictions. The central claim is an empirical error reduction on held-out test partitions. Representation construction is unsupervised: GCU solves min ||X_tr^+ - W_tr H||^2_F on training spectra, LVSE fits per-region SVDs on training spectra, standardization statistics are estimated from training, and test spectra are projected onto fixed training-derived components (W_te = argmin ||X_te^+ - W H^t||^2). The paper states: 'For each candidate representation, decomposition bases and standardization statistics were estimated from the training partition. Sample-wise transformations required no target information.' TabPFN weights remain fixed: 'the model random state matched the repetition seed' and 'RamanPFN therefore changes the spectral coordinates presented to TabPFN while preserving its pretrained architecture and parameters.' The only use of targets in the pipeline is OOF model selection on the training partition ('protocol-matched OOF predictions rank the GCU and LVSE candidates'), which is standard supervised model selection and cannot by itself force a held-out test result. The joint-visibility analysis is also target-independent: 'This analysis depended only on feature indices and the TabPFN sampling procedure and was independent of target values and RamanPFN representations.' Thus the 19.6% RMSE reduction and 9.0% classification-error reduction are not fitted values renamed as predictions. The paper's Discussion explicitly states the main limitation: 'Cross-instrument transfer, batch drift, calibration extrapolation and prospective deployment provide stronger tests of whether the constructed context remains stable beyond the present data distributions.' This is a generality concern, not circularity. A missing PCA/random-projection control for generic compression and the baseline RMSE mismatch between Tables 2 and 3 are validity/clarity concerns, but they do not amount to a definitional equivalence. No load-bearing self-citation was found: the cited RamanBench benchmark, TabPFN, and related methods are external works, not the present authors' prior results.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on two unsupervised linear decompositions plus a frozen tabular transformer; no new physical entities are introduced. The main hidden costs are the hyperparameters of the representation portfolios (NMF rank, trajectory checkpoints, region count, mode-retention, λ, α, preprocessing), whose exact values are not disclosed, and the assumption that within-dataset benchmark splits measure deployable utility.

free parameters (7)
  • GCU rank ρ
    Number of non-negative components in GCU; exact value or range not stated in the preprint, selected via OOF predictions.
  • GCU factorization trajectory checkpoint t
    Intermediate NMF states define the multiresolution family; exact checkpoints are not enumerated.
  • LVSE partition size S = 32-way partition shown in example
    Number of contiguous wavenumber regions; not stated as a fixed protocol across all tasks.
  • LVSE retained modes k_s per region = 4 modes / 98.9% variance in example
    Mode-retention criterion is not specified; example uses 4 modes but the global rule is absent.
  • Triplet weights λ and role-assignment policy = λ ∈ {1, 0.75, 0.5}
    Signed triplet integration uses λ values and a fixed prediction-space role policy; exact assignment rule and selection of retained triplets are under-specified.
  • Integration magnitude α
    α is 'fixed or determined analytically from the training-target scale, the radius of the selected correction and agreement among the retained triplets'—i.e., fitted on training data.
  • Preprocessing choice Ta (SNV or other)
    The text says 'Where used, standard normal variate normalization was applied'; the selection rule for when SNV is used is not specified.
axioms (5)
  • domain assumption Non-negative additive mixture model for Raman spectra
    GCU factorizes transformed spectra as X+ ≈ WH with W,H ≥ 0, assuming spectra are non-negative additive mixtures of latent components (Methods, Global Compositional Unmixing).
  • domain assumption Local low-rank linear structure within contiguous wavenumber intervals
    LVSE fits a centered SVD within each interval and retains k_s singular directions, assuming each region's variation is approximately a low-rank linear subspace (Methods, Local Vibrational Subspace Encoding).
  • domain assumption TabPFN in-context learning transfers to arbitrary coordinate transformations
    RamanPFN feeds GCU/LVSE coordinates to the frozen TabPFN and assumes pretrained in-context inference remains valid on these non-original features (Methods, RamanPFN formulation: p_r,q = f_θ(...)).
  • domain assumption Within-dataset stationarity of train and test spectra
    All evaluations use RamanBench's deterministic 80/20 splits; representation maps fit on training are applied to test spectra from the same dataset. Cross-instrument/batch-drift transfer is untested and listed as an open question.
  • standard math Statistical standard assumptions for bootstrap and rank tests
    Bootstrap resampling at target level, Wilcoxon signed-rank with Pratt treatment, Friedman test, and Holm correction assume appropriate exchangeability and no hidden dependencies beyond task clusters (Methods, Evaluation metrics and statistical analysis).

pith-pipeline@v1.3.0-daily-deepseek · 21157 in / 14004 out tokens · 117532 ms · 2026-08-04T13:33:35.665429+00:00 · methodology

0 comments
read the original abstract

Raman spectroscopy enables non-destructive, label-free molecular characterization across materials science, biomedicine and process monitoring. Predictive Raman datasets often contain few labelled spectra and thousands of ordered wavenumbers, with informative variation within bands and across distant spectral regions. Latent-variable chemometrics accommodates collinear small-sample data but can obscure fine peak morphology, whereas deep spectral networks resolve this structure only after task-specific training. TabPFN avoids task-specific parameter fitting through pretrained in-context inference, but processes very wide inputs as feature-subsampled views that do not preserve joint visibility of related bands. We present RamanPFN, a spectral representation framework that encodes these dependencies before TabPFN inference. Global Compositional Unmixing constructs non-negative coordinates over the complete spectrum so that distant bands with shared latent variation occupy a common predictive axis. Local Vibrational Subspace Encoding represents contiguous wavenumber regions with multiple orthogonal modes that retain independent changes in peak shape, intensity and position. The representations are evaluated separately and combined at the prediction level. Evaluation covered 150 tasks from 74 public Raman datasets. RamanPFN reduced root-mean-square error by 19.6% on average across 129 regression targets relative to direct TabPFN inference and further reduced the remaining classification error by 9.0% across 21 classification tasks. These results establish explicit spectral representation as an effective interface between high-dimensional Raman measurements and reusable tabular inference.

Figures

Figures reproduced from arXiv: 2608.02157 by Huan Wang, Jinjia Guo, Jixi Lu, Siming Dong, Xingyu Pan, Zhenlin Zhao.

Figure 1
Figure 1. Figure 1: RamanPFN constructs a dual-scale spectral interface between Raman spectra and a tabular foundation model. a, Generic feature coverage exposes multiple views of an ordered spectrum but fragments global co-variation and local continuity. b, Global Compositional Unmixing (GCU) encodes full-spectrum compositional structure, whereas Local Vibrational Subspace Encoding (LVSE) preserves multiple modes of variatio… view at source ↗
Figure 2
Figure 2. Figure 2: GCU and LVSE restore complementary spectral structure that generic feature subsampling obscures. a, Fractions of randomly sampled channel pairs and triplets that appeared together in at least one TabPFN forward pass across 129 regression tasks; marginal coverage remains complete as joint visibility declines with spectral width. b, Six high-variance GCU components for a 961-channel fuel-analysis task, each … view at source ↗
Figure 3
Figure 3. Figure 3: Breadth and statistical robustness of RamanPFN regression performance. a, Task-level RMSE change relative to the TabPFN reference across 129 regression targets from 53 source datasets. Bars show the mean of three repetitions, and grey vertical lines span the corresponding repetition-specific values. Targets are ordered by mean change; blue indicates lower RMSE and red indicates higher RMSE. b, Three-repeti… view at source ↗
Figure 4
Figure 4. Figure 4: Complementary task-level contributions of GCU and LVSE. a, Mean RMSE change produced by GCU alone and LVSE alone for each of the 129 regression targets across three repetitions. Negative values indicate lower prediction error relative to the TabPFN reference. Point colours identify targets improved by both representations, by GCU alone, by LVSE alone or by neither representation; dashed lines mark zero cha… view at source ↗
Figure 5
Figure 5. Figure 5: Transfer of the dual-scale representation to Raman classification. a, Mean weighted F1 for the TabPFN reference and RamanPFN across 21 classification tasks, averaged over three repetitions. Open grey circles denote the reference and filled circles denote RamanPFN. Blue connecting segments indicate higher weighted F1 with RamanPFN, whereas red segments indicate lower weighted F1. Tasks are ordered by mean c… 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

63 extracted references · 19 canonical work pages

  1. [1]

    Sokolov, Sergei V

    Guang Yang, Xin Li, Yongqiang Cheng, Mingchao Wang, Dong Ma, Alexei P. Sokolov, Sergei V. Kalinin, Gabriel M. Veith, and Jagjit Nanda. Distilling nanoscale heterogeneity of amorphous silicon using tip-enhanced Raman spectroscopy (TERS) via multiresolution manifold learning. Nature Communications, 12(1):578, 2021. doi: 10.1038/s41467-020-20691-2

    work page doi:10.1038/s41467-020-20691-2 2021

  2. [2]

    Photon-counting Raman spectroscopy at a MHz spectral rate for biochemical imaging of an entire organism.Nature Communications, 16(1):3808, 2025

    Sicheng Li, Haozheng Li, Yiran Li, Qi Zhang, Shuai Wang, Xin Lv, Shuai Yan, Zhiliang Huang, Xingbo Liu, Qipei Zhou, Bi Zhang, Long Xiao, Yage Chen, Zhe Wang, Wanjun Lu, Aiguo Shen, Jianfeng Liu, and Ping Wang. Photon-counting Raman spectroscopy at a MHz spectral rate for biochemical imaging of an entire organism.Nature Communications, 16(1):3808, 2025. do...

    work page doi:10.1038/s41467-025-59030-8 2025

  3. [3]

    Bohndiek

    Marlous Kamp, Jakub Surmacki, Marc Segarra Mondejar, Tim Young, Karolina Chrabaszcz, Fadwa Joud, Vincent Zecchini, Alyson Speed, Christian Frezza, and Sarah E. Bohndiek. Raman micro-spectroscopy reveals the spatial distribution of fumarate in cells and tissues.Nature Communications, 15(1):5386, 2024. doi: 10.1038/s41467-024-49403-w

    work page doi:10.1038/s41467-024-49403-w 2024

  4. [4]

    Rapid, label-free histopathological diagnosis of liver cancer based on Raman spectroscopy and deep learning

    Liping Huang, Hongwei Sun, Liangbin Sun, Keqing Shi, Yuzhe Chen, Xueqian Ren, Yuancai Ge, Danfeng Jiang, Xiaohu Liu, Wolfgang Knoll, Qingwen Zhang, and Yi Wang. Rapid, label-free histopathological diagnosis of liver cancer based on Raman spectroscopy and deep learning. Nature Communications, 14(1):48, 2023. doi: 10.1038/s41467-022-35696-2

    work page doi:10.1038/s41467-022-35696-2 2023

  5. [5]

    Tadesse, Amanda R

    Yirui Zhang, Kai Chang, Babatunde Ogunlade, Liam Herndon, Loza F. Tadesse, Amanda R. Kirane, and Jennifer A. Dionne. From genotype to phenotype: Raman spectroscopy and machine learning for label-free single-cell analysis.ACS Nano, 18(28):18101–18117, 2024. doi: 10.1021/acsnano.4c04282

    work page doi:10.1021/acsnano.4c04282 2024

  6. [6]

    Optics miniaturization strategy for demanding Raman spectroscopy applications

    Oleksii Ilchenko, Yurii Pilhun, Andrii Kutsyk, Denys Slobodianiuk, Yaman Goksel, Elodie Dumont, Lukas Vaut, Chiara Mazzoni, Lidia Morelli, Sofus Boisen, Konstantinos Stergiou, Yaroslav Aulin, Tomas Rindzevicius, Thomas Emil Andersen, Mikael Lassen, Hemanshu Mundhada, Christian Bille Jendresen, Peter Alshede Philipsen, Merete Hædersdal, and Anja Boisen. Op...

    work page doi:10.1038/s41467-024-4 2024

  7. [7]

    A universal and accurate method for easily identifying components in Raman spectroscopy based on deep learning.Analytical Chemistry, 95(11): 4863–4870, 2023

    Xiaqiong Fan, Yue Wang, Chuanxiu Yu, Yuanxia Lv, Hailiang Zhang, Qiong Yang, Ming Wen, Hongmei Lu, and Zhimin Zhang. A universal and accurate method for easily identifying components in Raman spectroscopy based on deep learning.Analytical Chemistry, 95(11): 4863–4870, 2023. doi: 10.1021/acs.analchem.2c03853

    work page doi:10.1021/acs.analchem.2c03853 2023

  8. [8]

    Seitz, David Heinzmann, Katja Schenke-Layland, Patricia B

    Manuel Sigle, Anne-Katrin Rohlfing, Martin Kenny, Sophia Scheuermann, Na Sun, Ulla Graeßner, Verena Haug, Jessica Sudmann, Christian M. Seitz, David Heinzmann, Katja Schenke-Layland, Patricia B. Maguire, Axel Walch, Julia Marzi, and Meinrad Paul Gawaz. 21 Translating genomic tools to Raman spectroscopy analysis enables high-dimensional tissue characteriza...

    work page doi:10.1038/s41467-023-41417-0 2023

  9. [9]

    Cruz Bournazou, Peter Neubauer, Felix Bießmann, and Erik Rodner

    Mario Koddenbrock, Christoph Lange, Robin Legner, Martin Jäger, Martin Kögler, Mariano N. Cruz Bournazou, Peter Neubauer, Felix Bießmann, and Erik Rodner. RamanBench: a large- scale benchmark for machine learning on Raman spectroscopy.arXiv preprint arXiv:2605.02003, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    PLS-regression: a basic tool of chemometrics.Chemometrics and Intelligent Laboratory Systems, 58(2):109–130, 2001

    Svante Wold, Michael Sjöström, and Lennart Eriksson. PLS-regression: a basic tool of chemometrics.Chemometrics and Intelligent Laboratory Systems, 58(2):109–130, 2001. doi: 10.1016/S0169-7439(01)00155-1

    doi:10.1016/s0169-7439(01)00155-1 2001

  11. [11]

    Partial least squares for discrimination.Journal of Chemometrics, 17(3):166–173, 2003

    Matthew Barker and William Rayens. Partial least squares for discrimination.Journal of Chemometrics, 17(3):166–173, 2003. doi: 10.1002/cem.785

    doi:10.1002/cem.785 2003

  12. [12]

    Stevens, and Mauricio Barahona

    Dimitar Georgiev, Álvaro Fernández-Galiana, Simon Vilms Pedersen, Georgios Papadopoulos, Ruoxiao Xie, Molly M. Stevens, and Mauricio Barahona. Hyperspectral unmixing for Raman spectroscopy via physics-constrained autoencoders.Proceedings of the National Academy of Sciences, 121(45):e2407439121, 2024. doi: 10.1073/pnas.2407439121

    doi:10.1073/pnas.2407439121 2024

  13. [13]

    Image processing and machine learning for hyperspectral unmixing: an overview and the HySUPP python package.IEEE Transactions on Geoscience and Remote Sensing, 62:1–31, 2024

    Behnood Rasti, Alexandre Zouaoui, Julien Mairal, and Jocelyn Chanussot. Image processing and machine learning for hyperspectral unmixing: an overview and the HySUPP python package.IEEE Transactions on Geoscience and Remote Sensing, 62:1–31, 2024. doi: 10.1109/ TGRS.2024.3393570

    arXiv 2024

  14. [14]

    MAT-Net: multiscale aggregation transformer network for hyperspectral unmixing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–15, 2024

    Pengrui Wang, Rong Liu, and Liangpei Zhang. MAT-Net: multiscale aggregation transformer network for hyperspectral unmixing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–15, 2024. doi: 10.1109/TGRS.2024.3494795

    arXiv 2024

  15. [15]

    Horgan, Magnus Jensen, Anika Nagelkerke, Jean-Philippe St-Pierre, Tom Vercauteren, Molly M

    Conor C. Horgan, Magnus Jensen, Anika Nagelkerke, Jean-Philippe St-Pierre, Tom Vercauteren, Molly M. Stevens, and Mads S. Bergholt. High-throughput molecular imaging via deep- learning-enabled Raman spectroscopy.Analytical Chemistry, 93(48):15850–15860, 2021. doi: 10.1021/acs.analchem.1c02178

    work page doi:10.1021/acs.analchem.1c02178 2021

  16. [16]

    Dunlop, and Ji-Xin Cheng

    Haonan Lin, Hyeon Jeong Lee, Nathan Tague, Jean-Baptiste Lugagne, Cheng Zong, Fengyuan Deng, Jonghyeon Shin, Lei Tian, Wilson Wong, Mary J. Dunlop, and Ji-Xin Cheng. Microsecond fingerprint stimulated Raman spectroscopic imaging by ultrafast tuning and spatial-spectral learning.Nature Communications, 12(1):3052, 2021. doi: 10.1038/s41467-021-23202-z

    work page doi:10.1038/s41467-021-23202-z 2021

  17. [17]

    Deep learning-decoded Raman spectroscopy for hour-scale iPSC pluripotency assessment via lipid–protein biomarkers

    Jianhui Wan, Yuheng Wang, Weile Zhu, Weina Zhang, and Liyun Zhong. Deep learning-decoded Raman spectroscopy for hour-scale iPSC pluripotency assessment via lipid–protein biomarkers. Analytical Chemistry, 97(46):25488–25499, 2025. doi: 10.1021/acs.analchem.5c03560

    work page doi:10.1021/acs.analchem.5c03560 2025

  18. [18]

    Martin, Peter J

    Longji Zhu, Yunan Yang, Fei Xu, Xinyu Lu, Mingrui Shuai, Zhulin An, Xiaomeng Chen, Hu Li, Francis L. Martin, Peter J. Vikesland, Bin Ren, Zhong-Qun Tian, Yong-Guan Zhu, and Li Cui. Open-set deep learning-enabled single-cell Raman spectroscopy for rapid identification of airborne pathogens in real-world environments.Science Advances, 11(2):eadp7991, 2025. ...

    work page doi:10.1126/sciadv.adp7991 2025

  19. [19]

    Jiaqi Hu, Gina Jinna Chen, Chenlong Xue, Pei Liang, Yanqun Xiang, Chuanlun Zhang, Xiaokeng Chi, Guoying Liu, Yanfang Ye, Dongyu Cui, De Zhang, Xiaojun Yu, Hong Dang, 22 Wen Zhang, Junfan Chen, Quan Tang, Penglai Guo, Ho-Pui Ho, Yuchao Li, Longqing Cong, and Perry Ping Shum. RSPSSL: a novel high-fidelity Raman spectral preprocessing scheme to enhance biome...

    work page doi:10.1038/s41377-024-01394-5 2024

  20. [20]

    Accurate predictions on small data with a tabular foundation model.Nature, 637:319–326, 2025

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model.Nature, 637:319–326, 2025. doi: 10.1038/s41586-024-08328-6

    doi:10.1038/s41586-024-08328-6 2025

  21. [21]

    When do neural nets outperform boosted trees on tabular data? InAdvances in Neural Information Processing Systems, volume 36, pages 76336–76369

    Duncan McElfresh, Sujay Khandagale, Jonathan Valverde, Vishak Prasad C, Ganesh Ramakr- ishnan, Micah Goldblum, and Colin White. When do neural nets outperform boosted trees on tabular data? InAdvances in Neural Information Processing Systems, volume 36, pages 76336–76369. Curran Associates, Inc., 2023. doi: 10.52202/075280-3337

    work page doi:10.52202/075280-3337 2023

  22. [22]

    TabArena: a living benchmark for machine learning on tabular data

    Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzmüller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. TabArena: a living benchmark for machine learning on tabular data. InAdvances in Neural Information Processing Systems, volume 38. Curran Associates, Inc., 2025

    2025

  23. [23]

    TabICL: a tabular foundation model for in-context learning on large data

    Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. TabICL: a tabular foundation model for in-context learning on large data. InProceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, pages 50817–50847. PMLR, 2025

    2025

  24. [24]

    Cresswell, Keyvan Golestan, Guangwei Yu, Anthony L

    Junwei Ma, Valentin Thomas, Rasa Hosseinzadeh, Alex Labach, Jesse C. Cresswell, Keyvan Golestan, Guangwei Yu, Anthony L. Caterini, and Maks Volkovs. TabDPT: scaling tabular foundation models on real data. InAdvances in Neural Information Processing Systems, volume 38, pages 172692–172722. Curran Associates, Inc., 2025

    2025

  25. [25]

    A closer look at TabPFN v2: understanding its strengths and extending its capabilities.arXiv preprint arXiv:2502.17361, 2025

    Han-Jia Ye, Si-Yang Liu, and Wei-Lun Chao. A closer look at TabPFN v2: understanding its strengths and extending its capabilities.arXiv preprint arXiv:2502.17361, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    TabPFN unleashed: a scalable and effective solution to tabular classification problems.arXiv preprint arXiv:2502.02527, 2025

    Si-Yang Liu and Han-Jia Ye. TabPFN unleashed: a scalable and effective solution to tabular classification problems.arXiv preprint arXiv:2502.02527, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    TabPFN-Wide: continued pre-training for extreme feature counts.arXiv preprint arXiv:2510.06162, 2025

    Christopher Kolberg, Jules Kreuer, Jonas Huurdeman, Sofiane Ouaari, Katharina Eggensperger, and Nico Pfeifer. TabPFN-Wide: continued pre-training for extreme feature counts.arXiv preprint arXiv:2510.06162, 2025

    arXiv 2025

  28. [28]

    TuneTables: context optimization for scalable prior-data fitted networks

    Benjamin Feuer, Robin Tibor Schirrmeister, Valeriia Cherepanova, Chinmay Hegde, Frank Hutter, Micah Goldblum, Niv Cohen, and Colin White. TuneTables: context optimization for scalable prior-data fitted networks. InAdvances in Neural Information Processing Systems, volume 37, pages 83430–83464, 2024. doi: 10.52202/079017-2654

    work page doi:10.52202/079017-2654 2024

  29. [29]

    Al Zadid Sultan Bin Habib, Md Younus Ahamed, Prashnna Kumar Gyawali, Gianfranco Doretto, and Donald A. Adjeroh. GOTabPFN: from feature ordering to compact tokenization for tabular foundation models on high-dimensional data. InProceedings of the 43rd International Conference on Machine Learning, 2026. To appear

    2026

  30. [30]

    Tabular foundation models for robust calibration of near-infrared chemical sensing data.arXiv preprint arXiv:2605.21544, 2026

    Robin Reiter, Denis Cornet, Fabien Michel, Lauriane Rouan, and Gregory Beurier. Tabular foundation models for robust calibration of near-infrared chemical sensing data.arXiv preprint arXiv:2605.21544, 2026. 23

    Pith/arXiv arXiv 2026

  31. [31]

    Bartel, and Gerbrand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5:1031–1041, 2023. doi: 10.1038/s42256-023-00716-3

    doi:10.1038/s42256-023-00716-3 2023

  32. [32]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

    Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery.Nature, 624:80–85, 2023. doi: 10.1038/s41586-023-06735-9

    doi:10.1038/s41586-023-06735-9 2023

  33. [33]

    A generative model for inorganic materials design.Nature, 639:624–632,

    Claudio Zeni et al. A generative model for inorganic materials design.Nature, 639:624–632,

  34. [34]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 10148–10167. PMLR, 2024

    2024

  35. [35]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Ali Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Bernie Wang. Chronos: learning the langu...

    2024

  36. [36]

    SpectralGPT: spectral remote sensing foundation model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227–5244, 2024

    Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. SpectralGPT: spectral remote sensing foundation model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227–5244, 2024. doi: 10.1109/TPAMI.202...

    arXiv 2024

  37. [37]

    Deep spectral component filtering as a foundation model for spectral analysis demonstrated in metabolic profiling.Nature Machine Intelligence, 7(5):743–757, 2025

    Bingsen Xue, Xinyuan Bi, Zheyi Dong, Yunzhe Xu, Minghui Liang, Xin Fang, Yizhe Yuan, Ruoxi Wang, Shuyu Liu, Rushi Jiao, Yuze Chen, Weitao Zu, Chengxiang Wang, Jianhao Zhang, Jiang Liu, Qin Zhang, Ye Yuan, Midie Xu, Ya Zhang, Yanfeng Wang, Jian Ye, and Cheng Jin. Deep spectral component filtering as a foundation model for spectral analysis demonstrated in ...

    work page doi:10.1038/s42256-025-01027-5 2025

  38. [38]

    HyperSIGMA: hyperspectral intelligence comprehension foundation model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(8):6427–6444, 2025

    Di Wang, Meiqi Hu, Yao Jin, Yuchun Miao, Jiaqi Yang, Yichu Xu, Xiaolei Qin, Jiaqi Ma, Lingyu Sun, Chenxing Li, Chuan Fu, Hongruixuan Chen, Chengxi Han, Naoto Yokoya, Jing Zhang, Minqiang Xu, Lin Liu, Lefei Zhang, Chen Wu, Bo Du, Dacheng Tao, and Liangpei Zhang. HyperSIGMA: hyperspectral intelligence comprehension foundation model.IEEE Transactions on Patt...

    arXiv 2025

  39. [39]

    A multi-fidelity tabular prior-data fitted network model for accurate prediction and uncertainty quantification.Nature Communications,

    Yan Shi, Cheng Liu, Aodi Yu, Zhenzhou Lu, Said Elias, Kai Cheng, Jiaqing Kou, Xin Chen, Yu Liu, Hong-Zhong Huang, and Michael Beer. A multi-fidelity tabular prior-data fitted network model for accurate prediction and uncertainty quantification.Nature Communications,

  40. [40]

    In context learning foundation models for materials 24 property prediction with small datasets.npj Computational Materials, 12(1):222, 2026

    Qinyang Li, Rongzhi Dong, Nicholas Miklaucic, Jeffrey Hu, Sadman Sadeed Omee, Lai Wei, Sourin Dey, Ming Hu, and Jianjun Hu. In context learning foundation models for materials 24 property prediction with small datasets.npj Computational Materials, 12(1):222, 2026. doi: 10.1038/s41524-026-02089-8

    work page doi:10.1038/s41524-026-02089-8 2026

  41. [41]

    A foundation model for atomistic materials chemistry.The Journal of Chemical Physics, 163:184110, 2025

    Ilyes Batatia et al. A foundation model for atomistic materials chemistry.The Journal of Chemical Physics, 163:184110, 2025. doi: 10.1063/5.0297006

    doi:10.1063/5.0297006 2025

  42. [42]

    A universal graph deep learning interatomic poten- tial for the periodic table.Nature Computational Science, 2(11):718–728, 2022

    Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic poten- tial for the periodic table.Nature Computational Science, 2(11):718–728, 2022. doi: 10.1038/s43588-022-00349-3

    doi:10.1038/s43588-022-00349-3 2022

  43. [43]

    Antoine Loew, Dewen Sun, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Universal machine learning interatomic potentials are ready for phonons.npj Computational Materials, 11(1):178, 2025. doi: 10.1038/s41524-025-01650-1

    doi:10.1038/s41524-025-01650-1 2025

  44. [44]

    MOMENT: a family of open time-series foundation models

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. MOMENT: a family of open time-series foundation models. InProceedings of the 41st Inter- national Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 16115–16152. PMLR, 2024

    2024

  45. [45]

    Timer: generative pre-trained transformers are large time series models

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Timer: generative pre-trained transformers are large time series models. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 32369–32399. PMLR, 2024

    2024

  46. [46]

    Unified training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 53140–53164. PMLR, 2024

    2024

  47. [47]

    A semantic-enhanced multi-modal remote sensing foundation model for earth observation.Nature Machine Intelligence, 7(8):1235–1249, 2025

    Kang Wu, Yingying Zhang, Lixiang Ru, Bo Dang, Jiangwei Lao, Lei Yu, Junwei Luo, Zifan Zhu, Yue Sun, Jiahao Zhang, Qi Zhu, Jian Wang, Ming Yang, Jingdong Chen, Yongjun Zhang, and Yansheng Li. A semantic-enhanced multi-modal remote sensing foundation model for earth observation.Nature Machine Intelligence, 7(8):1235–1249, 2025. doi: 10.1038/ s42256-025-01078-8

    2025

  48. [48]

    RemoteCLIP: a vision-language foundation model for remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024

    Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. RemoteCLIP: a vision-language foundation model for remote sensing.IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024. doi: 10.1109/TGRS.2024. 3390838

    doi:10.1109/tgrs.2024 2024

  49. [49]

    HyperSL: a spectral foundation model for hyperspectral image interpretation.IEEE Transactions on Geoscience and Remote Sensing, 63:1–19, 2025

    Weili Kong, Baisen Liu, Xiaojun Bi, Changdong Yu, Xinyao Li, and Yushi Chen. HyperSL: a spectral foundation model for hyperspectral image interpretation.IEEE Transactions on Geoscience and Remote Sensing, 63:1–19, 2025. doi: 10.1109/TGRS.2025.3566205

    arXiv 2025

  50. [50]

    Foundation model-based multimodal remote sensing data classification.IEEE Transactions on Geoscience and Remote Sensing, 62:1–17, 2024

    Xin He, Yushi Chen, Lingbo Huang, Danfeng Hong, and Qian Du. Foundation model-based multimodal remote sensing data classification.IEEE Transactions on Geoscience and Remote Sensing, 62:1–17, 2024. doi: 10.1109/TGRS.2023.3344698

    arXiv 2024

  51. [51]

    Foundation model-based spectral–spatial transformer for hyperspectral image classification.IEEE Transactions on Geoscience and Remote Sensing, 62:1–25, 2024

    Lingbo Huang, Yushi Chen, and Xin He. Foundation model-based spectral–spatial transformer for hyperspectral image classification.IEEE Transactions on Geoscience and Remote Sensing, 62:1–25, 2024. doi: 10.1109/TGRS.2024.3456129. 25

    arXiv 2024

  52. [52]

    Lee and H

    Daniel D. Lee and H. Sebastian Seung. Learning the parts of objects by non-negative matrix factorization.Nature, 401(6755):788–791, 1999. doi: 10.1038/44565

    doi:10.1038/44565 1999

  53. [53]

    SVD based initialization: a head start for nonnegative matrix factorization.Pattern Recognition, 41(4):1350–1362, 2008

    Christos Boutsidis and Efstratios Gallopoulos. SVD based initialization: a head start for nonnegative matrix factorization.Pattern Recognition, 41(4):1350–1362, 2008. doi: 10.1016/j. patcog.2007.09.010

    doi:10.1016/j 2008

  54. [54]

    Hierarchical ALS algorithms for nonnegative matrix and 3D tensor factorization

    Andrzej Cichocki, Rafal Zdunek, and Shun-ichi Amari. Hierarchical ALS algorithms for nonnegative matrix and 3D tensor factorization. InIndependent Component Analysis and Signal Separation, volume 4666 ofLecture Notes in Computer Science, pages 169–176. Springer Berlin Heidelberg, 2007. doi: 10.1007/978-3-540-74494-8_22

    work page doi:10.1007/978-3-540-74494-8_22 2007

  55. [55]

    R. J. Barnes, M. S. Dhanoa, and Susan J. Lister. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra.Applied Spectroscopy, 43(5):772–777,

  56. [56]

    G. H. Golub and C. Reinsch. Singular value decomposition and least squares solutions. Numerische Mathematik, 14(5):403–420, 1970. doi: 10.1007/BF02163027

    doi:10.1007/bf02163027 1970

  57. [57]

    ReZero is all you need: fast convergence at large depth

    Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. ReZero is all you need: fast convergence at large depth. InProceedings of the 37th Conference on Uncertainty in Artificial Intelligence, volume 161 ofProceedings of Machine Learning Research, pages 1352–1361. PMLR, 2021. URLhttps://proceedings.mlr.press/ v161/ba...

    2021

  58. [58]

    Random forests.Machine Learning, 45(1):5–32, 2001

    Leo Breiman. Random forests.Machine Learning, 45(1):5–32, 2001. doi: 10.1023/A: 1010933404324

    doi:10.1023/a: 2001

  59. [59]

    TabM: advancing tabular deep learning with parameter-efficient ensembling

    Yury Gorishniy, Akim Kotelnikov, and Artem Babenko. TabM: advancing tabular deep learning with parameter-efficient ensembling. InInternational Conference on Learning Representations, 2025

    2025

  60. [60]

    Revisiting nearest neighbor for tabular data: a deep tabular baseline two decades later

    Han-Jia Ye, Huai-Hong Yin, De-Chuan Zhan, and Wei-Lun Chao. Revisiting nearest neighbor for tabular data: a deep tabular baseline two decades later. InInternational Conference on Learning Representations, 2025. 26

    2025

  61. [1989]

    doi: 10.1366/0003702894202201

    doi:10.1366/0003702894202201

  62. [2025]

    doi: 10.1038/s41586-025-08628-5

    doi:10.1038/s41586-025-08628-5

  63. [2026]

    doi: 10.1038/s41467-026-75163-w

    work page doi:10.1038/s41467-026-75163-w