Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.06342
CoRTeC: A Differentially Private Release for Synthetic Tabular Data, Decodable by Any Generator
CoRTeC releases a differentially private file of noisy cohort counts—sizes, class balances, class-conditional histograms, and selected conditional outcome rates—then lets any decoder repeatedly generate synthetic rows without accessing private records or spending additional privacy budget. A maximum-entropy decoder and frozen language models can consume the same release, while a selection and sub-bin redraw step aligns generated rows with released statistics. At ε = 2, the paper reports that the maximum-entropy configuration comes within 0.02 AUC of a same-sized real sample on UCI Adult across three splits, and within 0.05 AUC on NHANES and bank marketing without selection.
How it works
Stage A forms non-overlapping cohorts from public domains and releases noisy cohort sizes, target balances, class-conditional histograms, and conditional outcome rates. Histograms are used because their sensitivity remains 1 regardless of the number of bins. Class blocks can be released in parallel, and conditional cells form partitions whose rates are post-processing of two noisy counts.
What the paper evaluates
The main evaluation uses UCI Adult, NHANES 2017-2018, and bank marketing. The main split uses seed 42; Adult is used for every comparison, controls, ladder, and replication, NHANES is used for the auto-configured clinical evaluation and replication, and bank marketing is used as a third dataset with one split. Comparisons include AIM, MST, DP-2Stage, DP-LLMTGen, and Tab-PE, plus controls for the release without a language model, the language model without a release, and a permuted-target floor. Fidelity is measured with one-way and two-way total variation and conditional total variation on seen cohort groups and held-out groups.
Limit to keep in view
The workload depends on public domains, bin edges, cohort rules, and conditional levels; fixed workloads can miss pairwise structure that adaptive methods capture more effectively.
arXiv:2611.07886
When, Where, How: Active In-Context Learning for Tabular Foundation Models
This paper studies active in-context learning for frozen tabular foundation models, asking which unlabeled rows should be labeled as the context grows. Across twenty OpenML classification benchmarks and two foundation models, it finds that diversity-based acquisition is strongest during the cold start, while target-aware methods become better once predictions are informative. The proposed Expected Predictive Information Gain in Context (EPIC) selects labels by estimating how hypothetical candidate labels would change predictions on target inputs, achieving the best final Cohen’s κ, retaining 80.3% and 87.0% of full-context performance with 5% and 10% of large pools labeled, and producing labels that also improve XGBoost and CatBoost.
How it works
The setting is transductive active in-context learning. An unlabeled pool supplies candidates, a target set represents the inputs whose predictions matter, and an oracle reveals queried labels at unit cost. The context starts with one uniformly drawn row per class and grows in batches of 10 until 100 labels are acquired or half of a small pool is labeled. The main comparison uses TabICL v2 with eight ensemble members and default inference settings.
What the paper evaluates
The main benchmark is the twenty OpenML classification datasets used by Tab-AICL. They range from 145 to 581,012 rows, 4 to 94 features, and 2 to 7 classes; datasets above 10,000 rows are uniformly subsampled to 10,000. The eight large-pool datasets have pools of 1,358 to 7,000 rows, while the remaining twelve have pools of 101 to 700 rows. The paper also reports one indicative out-of-time weather regression task from TabReD, but EPIC itself is evaluated as a classification method. Each seed creates a stratified 70/30 pool/test split, initial context, and acquisition randomness.
Limit to keep in view
The paper observes the value of different acquisition criteria but does not evaluate an adaptive schedule that switches from diversity to target-aware acquisition or detects the transition automatically.
arXiv:2611.05325
CATF: A Manager-Worker Framework for Context-Aware Time-Series Forecasting
CATF is a context-aware time-series forecasting framework that combines a Transformer-based Manager with multiple Worker forecasters. The Manager uses target and contextual histories to hard-route each sample to one Worker during both training and inference, allowing Workers to specialize in different temporal regimes without manual data partitioning. Across five real-world datasets, four horizons, and two normalization schemes, CATF improves over its standalone counterpart in 162 of 200 comparisons; the best CATF-enhanced model beats the best standalone baseline in 19 of 20 settings and beats zero-shot Time-MoE in 15 of 20 settings.
How it works
For each forecast origin, the model receives a target-history window X and a context-history window C, both of length L. The target horizon Y contains the next Ly target values. The context variables are used only as inputs; CATF does not forecast their future values. Inputs are normalized either globally using training-split statistics or per sample using statistics from the input window, and predictions are inverse-transformed before evaluation. The Manager is a K-class Transformer classifier.
What the paper evaluates
The evaluation uses five context-rich datasets: Weather with CO2 as the target and 20 meteorological context variables; PM2.5-LA and PM2.5-RV with air-quality, meteorological, reanalysis, and co-pollutant context; Bike Share with calendar and weather context; and Appliances with room-level, lighting, and outdoor weather context. Samples are split chronologically into 80% training, 10% validation, and 10% testing. The look-back length is L=96 and horizons are Ly∈{96,192,336,720}. Baselines are Informer, Autoformer, FEDformer, TimesNet, and iTransformer.
Limit to keep in view
The evaluation covers five context-rich datasets with one target variable and does not establish performance on jointly forecasted multivariate benchmarks such as ETT or Electricity.
arXiv:2611.06441
Graph Foundation Models Can Become Effective Multivariate Time Series Anomaly Detectors
GFM4MTSAD adapts frozen pretrained graph foundation models (GFMs) for unsupervised multivariate time series anomaly detection by addressing the mismatch between continuous sensor streams and graph-model inputs, the lack of reliable anomaly labels, and the distortion that can occur when sensor-level signals are combined into window-level decisions. Its three modules are Structure-and-Information Contract Alignment (SICA), Mechanism-Aligned Pseudo-Anomaly Screening (MAPS), and Tail-Gated Evidence Calibration (TGEC). On MSL P-15, SMAP T-3, and SMD 1-1, two GFM backbones achieve strong but dataset-dependent results against nine baselines.
How it works
Given a window of sensor observations, a temporal frontend extracts per-sensor features and the Token Bridge projects them into the native input dimension of a frozen GFM. The projection is initialized from leading principal components of normal features. SICA then constructs a message-passing graph from absolute feature correlations regularized by edge frequencies measured on normal training graphs. Equation 1 defines this prior-weighted top-k graph.
What the paper evaluates
The authors evaluate on Mars Science Laboratory telemetry entity MSL P-15, Soil Moisture Active Passive telemetry entity SMAP T-3, and Server Machine Dataset entity SMD 1-1. The reported dataset table lists 55, 25, and 38 variables, respectively, along with train/test counts and window/anomaly counts. Sequences are divided into their original train and test partitions, then segmented into windows of length 60 and stride 1. A test window is anomalous if it contains at least one anomalous timestamp. The anomaly-free training partition is further split: the first 80% is used for model fitting and reference construction, and the remaining 20% is held out for validation.
Limit to keep in view
Evaluation covers three dataset entities—MSL P-15, SMAP T-3, and SMD 1-1—with a three-run main comparison; broader datasets, entities, and operating regimes are not established by these results.
Read the full reviews
- CoRTeC: A Differentially Private Release for Synthetic Tabular Data, Decodable by Any Generator · arXiv:2611.06342
- When, Where, How: Active In-Context Learning for Tabular Foundation Models · arXiv:2611.07886
- CATF: A Manager-Worker Framework for Context-Aware Time-Series Forecasting · arXiv:2611.05325
- Graph Foundation Models Can Become Effective Multivariate Time Series Anomaly Detectors · arXiv:2611.06441
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.