»Forward
Pass.

Problem

TabPFN predicts well on tabular data without task-specific training, but why it succeeds or fails on geometrically hard datasets is unclear, making it hard to know when to trust its outputs.

Method

Treat TabPFN's 12 transformer-layer representations as evolving point clouds and track their shape across layers using zigzag persistent homology, tested on synthetic tasks (circles, tori, spheres, knots) with known ground truth.

Results

On a warped-circle case study, H0 fragmentation correlated with mean absolute residual (rho=0.92) and Bayes error (rho=0.94); across 180 benchmark runs, harder geometries roughly doubled MAR (0.022 to 0.244) and Bayes error (0.015 to 0.223).

Limitations

Tested only on synthetic data with known topology, not real-world tabular datasets; H0/H1 descriptors are coarse global summaries that can't pinpoint responsible rows or features; results depend on graph-construction choices (k, metric); fine-tuning didn't fix reliability (p=0.79).

Problem

Prior power-law studies of classical ML learning curves on tabular data use one team and a few curves, so it's unclear if scaling trends and data-requirement estimates generalize or reflect one group's implementation choices.

Method

A distributed classroom replication: 127 students each ran a fixed protocol (7 training fractions, seed 42) on 3 of 18 datasets across 6 model families, fitting error(N) = aN^-b + c to 11,536 runs.

Results

77.7% of 1,648 curves hit R^2>0.8; Boosting/RandomForest dominate at full data (50%/33% of datasets); 5/6 families share a near-universal exponent (cross-dataset R^2 gap<0.011); Lasso fails; same-seed re-implementations still differ, CV(b)=0.144.

Limitations

Fixed hyperparameters, single seed, modest dataset sizes (up to ~54k rows), only partial curve collapse (32-58% within 0.5 dex), AIC prefers unconstrained per-dataset fits, and shared-exponent claims are described as compressibility, not proven universality.

Problem

Synthetic tabular data is certified using metrics like logistic-regression C2ST and Trend score, but these are largely blind to destroyed inter-column dependencies -- the structure carrying most discriminative signal for minority classes in imbalanced tasks like fraud and clinical risk.

Method

Decompose a strong XGBoost classifier two-sample test (XGB-C2ST) into marginal, dependency, and numerical-categorical cross components, anchored between a fully-factorized worst-case reference and a real-data oracle; applied to the flow-matching generator TabbyFlow/EF-VFM across five imbalanced tabular datasets.

Results

On adult, a fully-factorized (dependency-destroyed) baseline scores 0.999-1.000 on LR-C2ST (looks perfect) but only 0.102 minority F1 (vs 0.709 oracle, 0.663 TabbyFlow). XGB-C2ST correctly flags it (0.956-0.995). TabbyFlow's residual dependency component is +0.050 (p<1e-3), costing 0.012-0.046 minority F1. A 16x capacity increase doesn't shrink it (p=0.098-0.484).

Limitations

Single-author study on 4-5 tabular datasets (Shoppers excluded as noise-dominated; Default excluded from the capacity experiment). Authors caution the decomposition is a discriminability share in XGB-AUC space, not a formal fraction of statistical divergence, and note tested fixes (copula correction, coupling) may not exhaust the intervention space.

Problem

Supplier lead time forecasts feed inventory and risk planning, but many orders are still in transit when forecasts are made (right-censored), and standard regression/classification models simply discard this partial information.

Method

LeadTime-ICL pairs a transformer (PFN-style in-context learning) with a conditional normalizing-flow head, pretrained on synthetic censored lead-time tasks so it adapts to new supplier datasets without retraining.

Results

Across 24 proprietary datasets (7 industries), LT-ICL had best point-forecast (WMAPE) on 15/24 and best probabilistic (SCRPS) on 14/24, beating runner-ups (avg WMAPE 0.467 vs 0.531; SCRPS 0.344 vs 0.378), with lowest deployment runtime.

Limitations

Datasets restricted to a shared covariate schema for comparability/confidentiality, limiting generality; transformer context scales quadratically; evaluation used retrospective censoring rather than live deployment, so real-world planning impact is untested.

Problem

Filling missing table cells needs exact, not just close, predictions, but frontier reasoning LLMs are overconfident, hallucinate answers, and are too costly to run over every spreadsheet with missing data.

Method

Auto-Fill post-trains three small specialist LMs (world-knowledge, text-reasoning via DeepSeek-R1 distillation, code-based self-validation), then calibrates and ensembles them, picking the most confident specialist or abstaining.

Results

Across 11 benchmarks (2,200 real tables), Auto-Fill beats o3-pro, Gemini 3 Pro, and DeepSeek R1 on quality while costing under 1% as much (Auto-Fill-Qwen is 100x cheaper than o3-pro, 29x cheaper than Gemini 3 Pro).

Limitations

Likely: relies on distillation from a single teacher (DeepSeek-R1) so quality is bounded by it; specialists need per-domain fine-tuning data; code specialist has low recall since programmatic patterns exist in only some tables.

Problem

Time series classification needs to capture shapes, frequencies, and cross-channel patterns, but random convolutional features (MultiRocket, Hydra) are usually paired with simple linear classifiers, potentially wasting their expressiveness.

Method

MASHT concatenates MultiRocket and Hydra convolutional features into a tabular representation, then feeds them directly into TabPFN-3, a pretrained in-context tabular foundation model, with no task-specific training.

Results

On UTF-112 univariate datasets MASHT reaches 0.892 mean accuracy (vs 0.891 for HIVE-COTE 2.0), with better mean rank (3.27 vs 3.41) and wins on 65/112 datasets over MR-Hydra. On 71 multivariate Multiverse datasets it trails HC2 (0.795 vs 0.805 accuracy, rank 7.65 vs 6.51).

Limitations

Inference relies on a foundation model with nontrivial memory/compute cost (A100 GPU, ~59.7s median runtime per dataset); results are confined to UTF-112/Multiverse benchmarks and protocols; multivariate performance is only competitive, not leading.

Problem

Feature-attribution scores explain input-output relationships but not how attention operators actually propagate information across a Transformer's intermediate layers. The authors want a structural, layer-by-layer account of internal organization in tabular Transformers, not just a final attribution map.

Method

Treat attention rollout as a row-stochastic operator and apply classical Doeblin-Dobrushin contraction theory: a small Dobrushin coefficient implies rollout is close to a rank-one matrix set by normalized column sums, giving rollout's propagation profile a structural interpretation.

Results

Dobrushin coefficient shrinks sharply with depth on the metabolomic dataset (1.7e-1 at 3 layers, 7.6e-4 at 11 layers, 1.5e-6 at 20 layers). Rollout vs. GradientExplainer top-6 overlap is 0.50-0.78, but full-ranking Spearman correlations across rollout, PCA, and SHAP stay near zero.

Limitations

Explicitly stated: rollout ignores value/output projections, MLPs, and normalization layers, so scores are not causal effects or guaranteed-faithful attributions. Tests use one biomedical cohort and three random seeds only; predictive relevance of top-ranked variables remains unestablished without intervention-based controls.

Problem

Classifying malware families from memory-forensics data is hard due to high inter-class similarity, class imbalance, and shifting data distributions, especially when distinguishing fine-grained sub-families rather than broad categories.

Method

LFS-FRAME stacks a Kolmogorov-Arnold Network (smooth functional patterns) with XGBoost (sharp rule-based splits), using strict out-of-fold cross-validation so a meta-classifier fuses their probability outputs without training/validation leakage.

Results

On CIC-MalMem-2022/EnhancedVolMem, LFS-FRAME reaches 89.85% accuracy on 4-class major families and 81.74% on 16-class sub-families, beating solo KAN (~70%), solo XGBoost (~76-79%), Random Forest, HyStack, and CNN-BiLSTM baselines.

Limitations

Likely: evaluated only on memory-forensics malware datasets (one partly synthetic via GAN/SMOTE augmentation), adds K-fold training overhead from two base learners, and generalization beyond this domain and against statistical significance testing is untested.

Problem

Retrieved tables can be topically relevant to a query yet lack the actual evidence to answer it ("answerability"). Dense retrievers optimize for semantic relevance, not answerability, so RAG pipelines can confidently return unhelpful tables.

Method

TCR-Bench uses 'sibling tables' (same schema, subtly different content, 209 queries/637 tables from Spider/BIRD) to isolate this gap, then tests Answerability-Aware Reranking (AAR): a second-stage cross-encoder or LLM judge asking directly 'can this table answer the query?'

Results

Retrievers find the right sibling group (GR@1=0.670) but pick the wrong table within it (R@1=0.182); QA F1 falls from 0.755 (oracle table) to 0.330 (top-5 retrieved). AAR reranking raises R@1 to 0.574.

Limitations

Only open-source retrievers/embedders tested, no proprietary systems; benchmark is a controlled diagnostic (sibling tables), not representative of messy real-world retrieval; AAR adds inference cost and mitigates rather than fixes the gap; single-table focus only.