{"id":"436d1639-f1dc-4474-9300-0ebd71c6366a","arxiv_id":"2608.00608","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TabPFN, a tabular foundation model, outperformed PLSR, Cubist, Random Forest, and CNN on 85 soil spectroscopy regression tasks, with PLS-compressed features giving the best results.","lead":"Spectroscopy predicts soil properties from how soil reflects light, but which prediction model works best has been unclear. On 85 soil-spectroscopy datasets, a modern AI model called TabPFN beat classical methods most often, and pairing it with a standard compression step worked even better.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IID random splits may bias rankings under spatial autocorrelation; §4.4 concedes OOD is untested, yet the abstract claims 'consistently best across scales' without this caveat.","rationale":"The reader identified the same weakest assumption: random splitting in spatially autocorrelated soil data. This is the most load-bearing concern because the paper's practical contribution is model-selection guidance, and that guidance currently rests on IID rankings that may not transfer to the spatial deployment setting. The paper explicitly flags this limitation in §4.4, which strengthens the case that the abstract overreaches. Other potential concerns (e.g., lack of significance testing, internal consistency of the 'full spectra surpass all baselines' reading) are secondary because the rankings are generally clear, whereas spatial leakage could systematically favor memory-based models and invert the central ranking. The proposed spatial-CV test is feasible given the public code and data, and would directly settle whether the concern lands. The verdict remains CONDITIONAL rather than REJECT because the evidence is substantial and the authors already acknowledge the limitation; a careful re-analysis with spatial splits could confirm or qualify the central claim.","tokens_in":23585,"tokens_out":2831,"duration_ms":39688,"concrete_test":"Re-run all 14 configurations with spatial block cross-validation: for LimeSoDa, use site- or field-level leave-one-group-out splits based on the provided coordinates; for OSSL, group samples by source library or geographic cluster and perform group-wise CV. Recompute mean ranks and relative RMSE gaps. If TabPFN variants remain first under spatial splits, the concern is resolved; if PLSR/Cubist close the gap or overtake, the random-split IID evaluation was load-bearing and the abstract/conclusions must be qualified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central ranking claim rests entirely on the cross-validation protocol of §2.2.4, where outer folds are 'created by random splitting.' Soil spectra are strongly spatially autocorrelated, and field-scale datasets are collected from a limited number of sites or fields. With random splits, near-duplicate spectra from the same site or pedon can appear in both training and test folds. This leakage is not neutral across models: memory- and attention-based methods (TabPFN with its context set, Cubist with its instance-based nearest-neighbor correction) can exploit spectral similarity far more effectively than PLSR or CNN, potentially inflating their measured advantage. The paper itself acknowledges in §4.4 that the evaluation is IID and that whether the advantages persist under OOD conditions 'requires further evaluation.' Since the abstract and conclusions present unqualified guidance ('consistently delivered the best overall performance across scales', 'strong default choice for soil spectroscopy'), the external validity of the ranking to real deployments—where models are applied to new sites or regions—is the weakest load-bearing point. The stated evidence does not yet support the unqualified claim; it supports an IID-only ranking.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a large benchmark comparing TabPFN v2.5, CNN, Cubist, Random Forest, PLSR, and linear regression, each with full spectra, PCA, or PLS latent variables, on 85 soil-spectroscopy regression tasks (39 field-scale LimeSoDa, 46 global OSSL). Using nested 5-fold CV with random splits, shared folds, and Bayesian hyperparameter optimization, it finds that TabPFN variants dominate, PLS+TabPFN is best, PLS is the most effective dimensionality reduction, and Cubist is the strongest classical baseline. It also reports runtime scaling showing that TabPFN shifts computational cost to the prediction stage.","tokens_in":23758,"tokens_out":6997,"duration_ms":86895,"significance":"If the IID ranking holds, this is a valuable empirical contribution: it would provide evidence that tabular foundation models are a strong default for soil spectroscopy across scales and that supervised PLS feature extraction remains useful even for modern models. The strengths are the breadth of the benchmark (85 tasks across two operational scales), the careful protocol (same folds across configurations, nested CV, inner validation for hyperparameters), and the open code and data. I find no circularity in the PLS-as-features design: PLS is fitted only on training folds within the nested CV. However, the external-validity gap described below is substantial, so the overarching claims currently outrun the evidence.","major_comments":[{"comment":"The abstract's central claim ('TabPFN consistently delivered the best overall performance across scales') and the conclusion's recommendation of TabPFN 'as default ... across operational scales' outrun the evidence. §2.2.4 states outer folds were 'created by random splitting'; §4.4 concedes this is an IID evaluation and that OOD behavior 'requires further evaluation.' Because soil spectra are spatially autocorrelated and the LimeSoDa field datasets are small and site-clustered (32–460 samples, §2.1.1), random splits can place near-duplicate spectra from the same field in both training and test folds. This is not neutral: TabPFN and Cubist are memory/instance-based (context set; nearest-neighbor correction) and can exploit spectral similarity, while PLSR and CNN cannot. The ranking may therefore be inflated in favor of these models for deployment to new sites. Please add site/field-groupe","section":"§2.2.4 and §4.4"},{"comment":"No uncertainty quantification is provided for the ranks that carry the conclusions. Mean/median ranks across 39 and 46 tasks are shown, but there are no confidence intervals, paired tests, or repeated resampling. The text itself notes that on OSSL 'performance across top-ranking configurations was more comparable,' and Figure A1 shows substantial overlap between PLS TabPFN and PCA TabPFN. A Friedman-style or Wilcoxon signed-rank test would substantiate the 'consistently best' claim; otherwise the top-1 vs top-2 separation may be within noise.","section":"§3 (Figures 5–6 and A1)"},{"comment":"The rank computation is under-specified. Eq. (3) defines RMSE with n as the number of samples in a held-out test fold, which suggests a per-fold RMSE; the text then reports 'mean and median ranks' across tasks. It is unclear whether ranks were computed per fold and then averaged, or computed once from task-level RMSE (pooled or averaged across folds). This ambiguity affects the headline ordering. Please state the aggregation explicitly and, if per-fold ranks were used, also report task-level ranks from pooled test predictions.","section":"§2.2.4 and §3"}],"minor_comments":[{"comment":"On OSSL, TabPFN on full spectra has a mean relative gap of 10.63%, while PLS Cubist has 8.74%. The phrase 'surpassed all classical baselines' should be stated as 'ranked higher' or reconciled with the gap metric, since the two criteria do not always coincide.","section":"Appendix A"},{"comment":"A standard gradient-boosted tree baseline (XGBoost/LightGBM) is absent. Given its prevalence in tabular ML, a sentence acknowledging this exclusion would improve completeness.","section":"§2.2.3"},{"comment":"The prediction-time comparison mixes GPU (CNN, TabPFN) and CPU (all others) runtimes. The caption notes this, but the text should explicitly caution that absolute times are not directly comparable across those platforms.","section":"Figure 7"},{"comment":"For CNN, batch_size is a range on LimeSoDa but a single value (300) on OSSL. Please confirm this is intentional and state the reason.","section":"Table B1"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-suited to the journal and the benchmark is carefully constructed. My main concern is the gap between the IID-only evidence and the unqualified 'across operational scales' recommendation. I would encourage the editor to require either a grouped-CV/OOD analysis or a clear reframing of the claims before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: this is the first systematic run of TabPFN v2.5 (and PLS-latent-features) across 85 soil-spectroscopy tasks, field-scale LimeSoDa plus global OSSL. The design is careful: same folds across all configurations, nested CV, inner hyperparameter tuning, public code and data. The ranking it reports—TabPFN variants on top, PLS features helping most models, Cubist the best classical baseline—is credible as an IID result. The reproducibility alone makes this a useful reference for the pedometrics community.\n\nThe main new result is not that TabPFN is magical but that PLS latent variables remain an effective feature extraction step even for a modern tabular foundation model. That is a practical and transferable finding, and the paper does it without data leakage: PLS is fit inside the training folds. The self-citation is heavy but not circular; the prior TabPFN-DSM papers are real and this extends them.\n\nSoft spots. First, the abstract and conclusions claim TabPFN 'consistently delivered the best overall performance across scales' and that raw-spectra TabPFN 'surpassed all classical baselines.' The paper's own Appendix A shows that on OSSL, PLS-Cubist has a lower mean RMSE gap to the best model (8.74%) than raw-spectra TabPFN (10.63%). So the unqualified sentences overshoot the evidence. Second, the evaluation is IID: outer folds are random splits, and soil samples are spatially autocorrelated. The authors acknowledge this in Section 4.4 and say OOD 'requires further evaluation,' but that caveat never reaches the abstract. This matters because practitioners will deploy at new sites; the ranking could shift under spatial block CV, especially for memory-based methods like TabPFN and Cubist. That is a moderate concern, not a fatal one—the IID benchmark is still informative. Third, there are no confidence intervals or significance tests; with 85 tasks the ranking pattern is suggestive but the error bars are unknown. Also minor: no commit hash for the code.\n\nWho is this for? Anyone choosing a regression model and feature pipeline for soil spectroscopy will get concrete, evidence-based guidance. The paper is honest about its limitations in the discussion, but the framing needs to catch up with that honesty. A serious referee should ask for a spatial block CV analysis (or at least a clear IID-only label in the abstract), a significance test on the ranks, and a correction of the 'surpassed all baselines' overstatement.\n\nI'd send it to review, and I'd cite the benchmark if I worked in the area.","headline":"A solid, reproducible IID benchmark showing TabPFN + PLS wins on soil-spectroscopy tasks; the abstract overstates generality before OOD is tested.","tokens_in":24324,"tokens_out":2513,"would_cite":true,"duration_ms":28766,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Tabular foundation model TabPFN beats classical soil-spectroscopy regressors on every scale, and PLS latent variables make it even better.","keywords":["soil spectroscopy","tabular foundation models","TabPFN","partial least squares","digital soil mapping","dimensionality reduction","chemometrics","vis-NIR and MIR spectroscopy"],"falsifier":"A spatially blocked cross-validation that holds out entire fields or regions on the same 85 tasks could settle it: if under such splits PLSR or Cubist matches or beats TabPFN in RMSE ranking, the claimed consistent TabPFN advantage would not generalize to realistic deployment.","tokens_in":23412,"feed_emoji":"🧪","tokens_out":6465,"duration_ms":70933,"temperature":0.7,"pith_summary":"The paper sets out to show that a tabular foundation model, TabPFN, is the best default regressor for soil spectroscopy at every operational scale, and that its performance can be pushed further by feeding it partial least squares (PLS) latent variables instead of raw spectra. It runs a uniform benchmark across 85 regression tasks—39 field-scale tasks from LimeSoDa and 46 large-scale tasks from the Open Soil Spectral Library—comparing TabPFN to a CNN, Cubist, Random Forest, and PLSR, each on full spectra and on PCA and PLS features. The consistent ordering is the same at both scales: TabPFN first, TabPFN on PLS features even better, and Cubist the strongest classical baseline. If true, practitioners get a clear, evidence-based rule: reach for a tabular foundation model, pair it with supervised PLS features when accuracy matters, and skip explicit dimensionality reduction when a simple workflow is preferred.","feed_headline":"TabPFN tops soil spectroscopy at every operational scale","feed_subtitle":"On 85 regression tasks it beats CNN, Cubist, Random Forest, and PLSR; PLS features push it further.","key_machinery":"TabPFN (Tabular Prior-data Fitted Network), an in-context learning transformer pre-trained on synthetic tabular tasks that conditions on a labeled context set at inference time and needs no per-dataset training or hyperparameter search; and PLS latent variables, target-aware components computed by maximizing covariance between spectra and the soil property, used as supervised dimensionality-reduction input features for any regressor. The paper's mechanism is the pairing of these two: PLS compresses the high-dimensional, collinear spectra into a small set of target-aligned features, and TabPFN learns the residual mapping from those features to the property. A second mechanism is the runtime s","core_discovery":"The central claim is that in-context learning tabular foundation models, specifically TabPFN v2.5, deliver the best predictive accuracy for soil property estimation from vis-NIR and MIR spectra across scales, including global spectral libraries with tens of thousands of samples. Directly applying TabPFN to full SNV-preprocessed spectra already outperformed all classical baselines—PLSR, Cubist, Random Forest, and CNN—demonstrating that explicit dimensionality reduction is not required for strong performance. Adding PLS latent variables as inputs improved every model family, and PLS + TabPFN achieved the best predictions overall. PLS was consistently more effective than PCA as a feature-extrac","pith_inferences":["Beyond the paper: because folds were random splits and soil samples are spatially autocorrelated, the IID benchmark may overstate absolute accuracy; a spatially blocked cross-validation could produce different RMSEs and possibly compress the gap between TabPFN and the best classical models.","Beyond the paper: if PLS latent variables capture the target-relevant linear structure, then even larger gains might come from target-aware feature selection or from fine-tuning a spectroscopy-specific foundation model on real spectra, an avenue the paper flags as likely.","Beyond the paper: the quadratic inference cost could be mitigated by distillation into a lightweight student model, which would make TabPFN-level accuracy practical for real-time field sensing.","Beyond the paper: TabPFN's built-in predictive distributions could turn this accuracy advantage into calibrated uncertainty estimates for spectroscopy, an extension the paper mentions but does not test."],"forward_implications":["TabPFN becomes a strong default for soil spectroscopy across operational scales: on full spectra it beats classical baselines without dataset-specific tuning.","When accuracy is the priority, PLS latent variables plus TabPFN is the best configuration; PLS improves Cubist, Random Forest, and TabPFN relative to full spectra and PCA.","PLS is the more effective dimensionality-reduction strategy for spectroscopy than PCA, and PLSR remains the strongest linear model, confirming the lasting value of supervised compression.","Computational burden shifts to prediction: TabPFN inference grows quadratically with test-set size and can take roughly 24 minutes on the largest libraries; PLS cuts that by about 61%, while training time stays minimal.","Whether the advantage holds under out-of-distribution conditions such as new sites, instruments, or laboratories remains open; the paper identifies a systematic out-of-distribution evaluation as the most immediate next step."],"supporting_citations":[{"why":"Introduces TabPFN and the in-context learning paradigm on small tabular tasks that the paper scales to spectroscopy.","marker":"Hollmann et al., 2023"},{"why":"The Nature version of TabPFN; the paper cites it for model architecture and the O(n²+m²) attention complexity.","marker":"Hollmann et al., 2025"},{"why":"Presents TabPFN v2.5, the version evaluated here, with native support for datasets up to 100,000 samples and 2,000 features.","marker":"Grinsztajn et al., 2025"},{"why":"Provides the LimeSoDa benchmark collection with 13 spectroscopy datasets and 39 field-scale regression tasks.","marker":"Schmidinger et al., 2025"},{"why":"Provides the Open Soil Spectral Library (OSSL) with 46 large-scale regression tasks and the harmonized spectral data.","marker":"Safanelli et al., 2025"},{"why":"Earlier work established TabPFN as a strong field-scale digital soil mapping default; the paper extends this to large spectral libraries.","marker":"Barkov et al., 2026"},{"why":"Exploratory TabPFN application to national spectral library subsets, motivating the large-scale evaluation here.","marker":"Huang et al., 2026"},{"why":"Supplies the nested cross-validation protocol used for all model comparisons.","marker":"Varma and Simon, 2006"},{"why":"Standard normal variate (SNV) preprocessing applied to all spectra before modeling.","marker":"Barnes et al., 1989"}],"fun_headline_variants":["TabPFN out-predicts all classical soil spectra models","Soil spectroscopy: TabPFN wins, PLS boosts it further","From field to global scale, TabPFN leads soil spectra","TabPFN beats CNN, Cubist, RF, PLSR on 85 soil tasks"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that randomly splitting samples into training and test folds estimates real-world accuracy fairly; because soil samples close together are similar, random splits can leak spatial structure into training and may inflate accuracy or change which model ranks best.","fun_headline_variants_meta":{"raw":{"variants":["TabPFN out-predicts all classical soil spectra models","Soil spectroscopy: TabPFN wins, PLS boosts it further","From field to global scale, TabPFN leads soil spectra","TabPFN beats CNN, Cubist, RF, PLSR on 85 soil tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3069,"prompt_tokens":786,"completion_tokens":2283,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":530,"tokens_out":2283,"duration_ms":20397,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:31:15.019229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A spatially blocked cross-validation that holds out entire fields or regions on the same 85 tasks could settle it: if under such splits PLSR or Cubist matches or beats TabPFN in RMSE ranking, the claimed consistent TabPFN advantage would not generalize to realistic deployment.","supporting_citations":[],"review_version":1}