# Tabular Foundation Models for the Estimation of Probabilistic Quasar Photometric Redshifts in S-PLUS

- **DOI:** [10.3847/1538-3881/ae9eb6](https://doi.org/10.3847/1538-3881/ae9eb6)
- **Rating:** Strong (3.9 out of 5)
- **Verdict:** A rigorous benchmark establishes TabPFN 2.5 as a strong S-PLUS quasar photo-z default, constrained by uncertain target-population weighting.
- **Framework version:** v5.9
- **Pipeline version:** v0.4
- **Evaluation date:** 2026-10-02
- **Evaluation page:** https://www.nabu.science/paper/10.3847%2F1538-3881%2Fae9eb6

## Key claims

1. TabPFN 2.5 is best or statistically tied for best on all metrics except the unweighted CDE loss, on which the normalizing flow is statistically tied and attains the lowest mean value
   - Support: Supported
   - Type: Primary finding
   - Basis: broad table-level consistency, tempered by a permissive tie rule and uncertainty that excludes training variation
   - Location: Abstract; Sections 4.1–4.2; Tables 2–3
2. its largest gains occur for small training sets and in difficult regimes (very bright and faint sources, high redshift), while retaining near-nominal calibration under a covariate shift
   - Support: Tentative
   - Type: Secondary
   - Basis: faint-tail gains are modest and raw target weighting has very low effective sample size
   - Location: Abstract; Sections 4.1–4.3; Figures 3–6
3. reporting only unweighted spectroscopic metrics would therefore give an optimistic view of deployed performance
   - Support: Supported
   - Type: Secondary
   - Basis: all methods worsen after reweighting, conditional on an untestable covariate-shift assumption
   - Location: Section 5
4. SHAP attributions identify the Wide-field Infrared Survey Explorer W1/W2 as the strongest individual predictors, with UV and optical bands offering nonnegligible refinements
   - Support: Supported
   - Type: Secondary
   - Basis: first-order attributions cover 100 objects with a constrained coalition budget and moderate cross-model agreement
   - Location: Abstract; Section 4.4; Figure 7

## Strengths

- **Broad uncertainty-aware comparative benchmark** Eleven methods are evaluated on common data using density, calibration, point-error, and catastrophic-outlier metrics, with five repetitions and reported standard errors.
  - Dimension: methodological rigour
  - Location: Sections 3.1–3.3; Tables 2–3; Appendix B
- **Shift sensitivity is explicitly tested** The study reports unweighted, tempered, and raw importance-weighted evaluations, discloses the raw-weight ESS collapse, and repeats the analysis with a different density-ratio classifier.
  - Dimension: methodological rigour
  - Location: Sections 3.2 and 4.1; Appendix C
- **Conclusions remain survey-scoped** The recommendation is restricted to S-PLUS-like quasar samples, with covariate-shift assumptions, target-catalog impurity, grid sensitivity, and inference costs connected to interpretation.
  - Dimension: positioning
  - Location: Section 5; Appendix C

## Limitations

- **Missingness handling may confound comparisons** Nondetections are replaced by zero before standardization for non-tree methods, while tree methods retain the sentinel and foundation models apply their own processing. No alternative imputation or missing-indicator comparison tests whether this contributes to the tail-regime advantage.
  - Dimension: methodological rigour
  - Location: Appendix E, Feature handling
- **Raw target weighting is unstable** The α=1 evaluation has an effective sample size of only 216 objects, while the tempered α≈0.36 regime only partially shifts toward the target population. Target-performance conclusions also rely on an untestable covariate-shift assumption.
  - Dimension: methodological rigour
  - Location: Sections 3.2.2 and 4.1; Appendix B
- **Minor reporting inconsistencies remain** Nominally identical unweighted panels differ between Tables 2 and C1, Appendix C slightly understates one sensitivity difference, and the conclusion’s PIT-KS summary excludes the raw-weight value.
  - Dimension: reporting
  - Location: Table 2; Tables C1–C2; Section 5
- **Uncertainty omits training variation** The reported bootstrap standard errors quantify finite-test-set uncertainty after averaging repetitions rather than seed and training-set variability. This makes the broad one-SE tie sets less decisive than the headline wording may suggest.
  - Dimension: methodological rigour
  - Location: Appendix B

## Reliability checks

Caveats (4 of 4 checks).

- **Internal coherence:** Noted. The central tables and conclusions are broadly consistent, but several minor numerical and summary discrepancies are present. They affect precision and reporting clarity rather than the direction of the main benchmark result.
- **Research conduct:** Clean. The paper identifies a public spectroscopic compilation, a named code repository, and the planned public release location for DR6 photometry. The astronomical benchmark does not involve human-subject research requiring ethics approval.
- **Reference integrity:** Clean. 61 of 61 checkable references verified
  - 67 references in manuscript
  - 6 have no canonical index record — counted, but not index-checkable
  - 55 references confirmed by manual review
  - References not verified automatically are resolved manually within 24 hours.

## Post-publication record

- **Status:** Clean
- **Summary:** No retraction or expression of concern found in Crossref (includes Retraction Watch).
- **Last checked:** 2026-10-17
