When, Where, How: Active In-Context Learning for Tabular Foundation Models
Organizations: Department of Computer Science & Artificial Intelligence, University of Technology Nuremberg, Dr.-Luise-Herzberg-Straße 4, 90461 Nuremberg, Germany
Abstract
Tabular Foundation Models (TFMs) adapt to a new task only through the labeled rows in their context. However, labels are expensive to acquire in real-world scenarios due to the need for highly specific domain knowledge to ensure high-quality data. Active in-context learning transfers classical acquisition rules to this setting, in which the model that guides acquisition changes with every label added. We ask how the value of diversity, predictive uncertainty and target-aware information gain changes as a frozen TFM acquires more context. For that, we study Target-Aware Nearest Neighbor (TANN) acquisition and the Expected Predictive Information Gain in Context (EPIC) and evaluate them against classical rules. Our study across twenty benchmarks with two well-established TFMs and classical GBDTs highlights three findings: (i) When: diversity-based acquisition is best during the first acquisition iterations; (ii) Where: labeling pool rows that serve the target inputs via a simple target-aware nearest-neighbor rule matters substantially; (iii) How: EPIC yields the best models at the label budget, retains on average 80--87% of the chance-corrected full-data performance with only 5-10% of the pool labeled on the eight benchmarks with the largest pools, and selects labels that also improve the predictive performance of gradient-boosted decision trees.
Figures & tables
| Dataset | Random | Margin | Coreset | Hybrid | Proxy- Hybrid | Target- Random | Target- Coverage | TANN-feat | TANN | EPIC |
| iris | 0.940 0.053 | 0.927 0.051 | 0.920 0.052 | 0.930 0.041 | 0.930 0.053 | 0.933 0.047 | 0.933 0.047 | 0.940 0.036 | 0.930 0.050 | 0.923 0.052 |
| glass | 0.635 0.058 | 0.651 0.084 | 0.594 0.070 | 0.623 0.066 | 0.632 0.090 | 0.676 0.046 | 0.638 0.065 | 0.675 0.052 | 0.667 0.080 | 0.706 0.073 |
| ionosphere | 0.878 0.047 | 0.872 0.028 | 0.740 0.058 | 0.863 0.039 | 0.864 0.031 | 0.878 0.048 | 0.833 0.031 | 0.850 0.051 | 0.880 0.030 | 0.887 0.040 |
| balance-scale | 0.855 0.039 | 0.975 0.013 | 0.899 0.023 | 0.927 0.015 | 0.970 0.015 | 0.904 0.031 | 0.922 0.026 | 0.915 0.036 | 0.897 0.056 | 0.964 0.019 |
| vehicle | 0.954 0.028 | 0.991 0.009 | 0.934 0.034 | 0.965 0.023 | 0.989 0.011 | 0.957 0.014 | 0.937 0.041 | 0.986 0.011 | 0.989 0.006 | 0.996 0.005 |
| page-blocks | 0.695 0.047 | 0.770 0.074 | 0.695 0.080 | 0.798 0.025 | 0.791 0.036 | 0.725 0.030 | 0.804 0.028 | 0.832 0.023 | 0.819 0.067 | 0.812 0.040 |
| AULC | Final | ||||||
| Method | vs. | W/T/L | W/T/L | ||||
| EPIC | Random | 16/1/3 | +0.026 | 0.001 | 16/2/2 | +0.042 | 0.000 |
| Margin | 13/4/3 | +0.012 | 0.017 | 13/4/3 | +0.015 | 0.024 | |
| Coreset | 16/0/4 | +0.055 | 0.000 | 16/2/2 | +0.058 | 0.001 | |
| Hybrid | 7/2/11 | -0.003 | 0.701 | 15/1/4 | +0.030 | 0.011 | |
| Proxy-Hybrid | 10/4/6 | +0.014 | 0.177 | 15/2/3 | +0.028 | 0.001 | |
| Labels acquired by | TabICL | CatBoost | XGBoost |
| Random | 0.536 | 0.500 | 0.487 |
| Margin | 0.563 | 0.523 | 0.514 |
| Hybrid | 0.549 | 0.519 | 0.504 |
| TANN | 0.570 | 0.538 | 0.519 |
| EPIC | 0.579 | 0.543 | 0.530 |
| CatBoost-Margin (own labels) | – | 0.525 | – |
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | OpenML | Rows | |||
| Iris | 61 | 150 | 4 | 3 | 52 |
| Glass | 41 | 214 | 9 | 6 | 74 |
| Ionosphere | 59 | 351 | 34 | 2 | 100 |
| Balance-scale | 997 | 625 | 4 | 2 | 100 |
| Vehicle | 994 | 846 | 18 | 2 | 100 |
| Page-blocks | 1021 | 5,473 | 10 | 2 | 100 |
| Context | Random | Hybrid | TANN | EPIC |
| Global context ( ) | 0.536 | 0.549 | 0.570 | 0.579 |
| Local context, | ∗ 3/0/17 | ∗ 3/2/15 | ∗ 2/2/16 | ∗ 1/2/17 |
| Local context, | 2/4/13 | ∗ 2/4/13 | ∗ 2/5/12 | 2/4/13 |
| + confident pseudo-labels | 12/7/1 | 12/5/3 | ∗ 13/7/0 | 11/8/1 |
| + pseudo-labelled neighbours | 9/6/5 | 10/5/5 | 7/11/2 | 3/13/4 |
| + same neighbours, true labels (oracle) | ∗ 16/2/2 | ∗ 16/4/0 | ∗ 15/4/1 | ∗ 13/6/1 |
| AULC | Final | EPIC vs. | |||
| Method | Mean | Rank | Mean | Rank | W/T/L ( ) |
| EPIC | 0.479 | 3.85 | 0.578 | 3.35 | – |
| TANN | 0.479 | 3.80 | 0.570 | 4.85 | 11/4/5 (0.133) |
| Hybrid | 0.482 | 3.55 | 0.549 | 6.42 | 15/1/4 (0.011) |
| Margin | 0.467 | 5.50 | 0.563 | 5.58 | 13/4/3 (0.024) |
| Random | 0.453 | 6.90 | 0.536 | 7.65 | 16/2/2 (0.000) |
| Rule | Labelled fraction of the pool | |||
| 1% | 2% | 5% | 10% | |
| Random | 55.7 (0/8) | 62.6 (0/8) | 72.5 (0/8) | 78.9 (0/8) |
| Margin | 51.2 (0/8) | 59.0 (0/8) | 73.7 (1/8) | 83.6 (1/8) |
| Hybrid | 60.4 (0/8) | 66.0 (0/8) | 76.0 (0/8) | 82.0 (0/8) |
| TANN | 50.6 (0/8) | 61.9 (0/8) | 75.8 (1/8) | 83.5 (1/8) |
| EPIC | 50.9 (0/8) | 64.8 (0/8) | 80.3 (1/8) | 87.0 (1/8) |
| Labels used | (adaptive) | (fixed fraction) | (full budget) | |
| 0.02 | 39.6% | 0.493 | 0.483 | 0.578 |
| 0.01 | 57.2% | 0.526 | 0.526 | 0.578 |
| 0.005 | 72.5% | 0.554 | 0.550 | 0.578 |
| 0.002 | 91.2% | 0.574 | 0.569 | 0.578 |
| 0.001 | 97.4% | 0.577 | 0.576 | 0.578 |
| Dataset | Random | Margin | Coreset | Hybrid | Proxy- Hybrid | Target- Random | Target- Coverage | TANN-feat | TANN | EPIC |
| iris | 0.884 0.051 | 0.877 0.045 | 0.890 0.039 | 0.897 0.029 | 0.891 0.037 | 0.889 0.043 | 0.887 0.034 | 0.891 0.039 | 0.885 0.048 | 0.884 0.043 |
| glass | 0.499 0.037 | 0.516 0.060 | 0.441 0.051 | 0.524 0.044 | 0.519 0.037 | 0.540 0.046 | 0.525 0.058 | 0.551 0.054 | 0.554 0.062 | 0.586 0.046 |
| ionosphere | 0.734 0.042 | 0.744 0.042 | 0.525 0.089 | 0.788 0.034 | 0.759 0.045 | 0.759 0.061 | 0.706 0.077 | 0.730 0.060 | 0.760 0.044 | 0.719 0.038 |
| balance-scale | 0.711 0.042 | 0.770 0.048 | 0.756 0.029 | 0.780 0.005 | 0.769 0.017 | 0.735 0.048 | 0.753 0.048 | 0.767 0.031 | 0.764 0.042 | 0.772 0.053 |
| vehicle | 0.781 0.037 | 0.769 0.080 | 0.763 0.045 | 0.843 0.032 | 0.770 0.055 | 0.785 0.035 | 0.758 0.057 | 0.802 0.038 | 0.796 0.033 | 0.809 0.036 |
| page-blocks | 0.640 0.044 | 0.663 0.043 | 0.485 0.119 | 0.720 0.024 | 0.700 0.031 | 0.636 0.023 | 0.694 0.061 | 0.711 0.029 | 0.706 0.049 | 0.666 0.054 |
| AULC | Final | |||
| Rule | TabICL | TabPFN | TabICL | TabPFN |
| Random | 0.453 (7.65) | 0.419 (7.10) | 0.536 (6.90) | 0.518 (7.03) |
| Margin | 0.467 (6.40) | 0.437 (5.90) | 0.563 (5.30) | 0.542 (5.15) |
| Coreset | 0.424 (8.35) | 0.382 (7.30) | 0.520 (8.45) | 0.489 (7.88) |
| Hybrid | 0.482 (3.65) | 0.456 (3.75) | 0.549 (5.88) | 0.534 (5.55) |
| Proxy-Hybrid | 0.466 (5.75) | 0.435 (5.70) | 0.550 (6.70) | 0.551 (4.15) |
| Dataset | Random | Margin | Coreset | Hybrid | Proxy-Hybrid | Target- Random | Target- Coverage | TANN-feat | TANN | EPIC |
| iris | 0.957 0.030 | 0.950 0.045 | 0.943 0.042 | 0.957 0.052 | 0.957 0.033 | 0.933 0.039 | 0.950 0.040 | 0.957 0.042 | 0.957 0.033 | 0.940 0.042 |
| glass | 0.609 0.054 | 0.616 0.052 | 0.546 0.090 | 0.593 0.035 | 0.601 0.085 | 0.641 0.065 | 0.590 0.074 | 0.665 0.076 | 0.667 0.061 | 0.680 0.066 |
| ionosphere | 0.869 0.045 | 0.894 0.040 | 0.619 0.102 | 0.877 0.041 | 0.890 0.048 | 0.869 0.037 | 0.871 0.025 | 0.828 0.051 | 0.886 0.040 | 0.874 0.043 |
| balance-scale | 0.847 0.033 | 0.938 0.018 | 0.882 0.028 | 0.919 0.021 | 0.941 0.016 | 0.879 0.025 | 0.868 0.032 | 0.900 0.041 | 0.928 0.021 | 0.940 0.014 |
| vehicle | 0.946 0.019 | 0.995 0.007 | 0.927 0.041 | 0.969 0.019 | 0.994 0.007 | 0.950 0.029 | 0.959 0.022 | 0.989 0.007 | 0.996 0.005 | 0.993 0.008 |
| page-blocks | 0.673 0.061 | 0.774 0.056 | 0.714 0.073 | 0.755 0.054 | 0.782 0.076 | 0.697 0.065 | 0.797 0.037 | 0.833 0.019 | 0.830 0.021 | 0.815 0.022 |
| Dataset | Random | Margin | Coreset | Hybrid | Proxy- Hybrid | Target- Random | Target- Coverage | TANN-feat | TANN | EPIC |
| iris | 0.902 0.045 | 0.902 0.030 | 0.916 0.041 | 0.903 0.040 | 0.909 0.034 | 0.889 0.034 | 0.900 0.035 | 0.908 0.036 | 0.907 0.033 | 0.896 0.038 |
| glass | 0.471 0.039 | 0.478 0.048 | 0.385 0.041 | 0.472 0.035 | 0.449 0.048 | 0.515 0.053 | 0.493 0.060 | 0.529 0.063 | 0.540 0.055 | 0.551 0.044 |
| ionosphere | 0.702 0.072 | 0.727 0.034 | 0.346 0.108 | 0.784 0.036 | 0.771 0.042 | 0.706 0.085 | 0.759 0.038 | 0.705 0.051 | 0.744 0.057 | 0.651 0.056 |
| balance-scale | 0.695 0.052 | 0.703 0.053 | 0.737 0.025 | 0.752 0.024 | 0.714 0.051 | 0.705 0.039 | 0.691 0.069 | 0.714 0.049 | 0.718 0.036 | 0.739 0.030 |
| vehicle | 0.753 0.049 | 0.790 0.045 | 0.719 0.048 | 0.822 0.026 | 0.802 0.039 | 0.766 0.059 | 0.756 0.058 | 0.791 0.038 | 0.796 0.036 | 0.794 0.062 |
| page-blocks | 0.576 0.086 | 0.630 0.060 | 0.447 0.083 | 0.689 0.041 | 0.645 0.063 | 0.588 0.051 | 0.678 0.046 | 0.667 0.059 | 0.693 0.034 | 0.635 0.061 |
| Rule | Labelled fraction of the pool | |||
| 1% | 2% | 5% | 10% | |
| Random | 47.3 (0/8) | 56.6 (0/8) | 70.1 (0/8) | 78.4 (0/8) |
| Margin | 44.0 (0/8) | 57.7 (0/8) | 73.3 (1/8) | 81.9 (2/8) |
| Hybrid | 57.0 (0/8) | 63.4 (0/8) | 74.9 (0/8) | 80.8 (1/8) |
| TANN | 41.3 (0/8) | 59.3 (0/8) | 73.6 (1/8) | 80.7 (2/8) |
| EPIC | 48.2 (0/8) | 59.6 (0/8) | 75.5 (1/8) | 84.5 (2/8) |
| AULC | Final | ||||||
| Method | vs. | W/T/L | W/T/L | ||||
| EPIC | Random | 16/0/4 | +0.027 | 0.004 | 15/2/3 | +0.040 | 0.001 |
| Margin | 10/5/5 | +0.010 | 0.123 | 11/3/6 | +0.016 | 0.064 | |
| Coreset | 13/2/5 | +0.065 | 0.003 | 15/3/2 | +0.070 | 0.000 | |
| Hybrid | 5/2/13 | -0.009 | 0.330 | 14/1/5 | +0.024 | 0.017 | |
| Proxy-Hybrid | 11/1/8 | +0.012 | 0.245 | 7/5/8 | +0.007 | 0.729 | |
| Labels acquired by | TabPFN | CatBoost | XGBoost |
| Random | 0.518 | 0.500 | 0.487 |
| Margin | 0.542 | 0.526 | 0.507 |
| Hybrid | 0.534 | 0.520 | 0.501 |
| TANN | 0.554 | 0.532 | 0.519 |
| EPIC | 0.558 | 0.541 | 0.530 |
| CatBoost-Margin (own labels) | – | 0.525 | – |
| Final | AULC | ||||
| mean | mean | Time | |||
| 16 | 0.742 | (4/1/3, 1.00) | 0.645 | (6/0/2, 0.64) | 15 |
| 32 | 0.733 | (1/4/3, 0.56) | 0.641 | (2/3/3, 0.84) | 24 |
| 64 | 0.743 | – | 0.641 | – | 41 |
| 128 | 0.740 | (2/2/4, 0.46) | 0.627 | (0/4/4, 0.04) | 65 |
| all | 0.737 | (2/3/3, 0.25) | 0.624 | (0/1/7, 0.01) | 108 |
| Final | AULC | ||||
| mean | mean | Time | |||
| 16 | 0.580 | (7/2/11, 0.84) | 0.480 | (7/6/7, 0.96) | 49 |
| 32 | 0.574 | (6/4/10, 0.26) | 0.475 | (3/5/12, 0.08) | 52 |
| 64 | 0.581 | (4/10/6, 0.47) | 0.478 | (5/8/7, 0.53) | 58 |
| 256 | 0.578 | – | 0.479 | – | 87 |
| 1024 | 0.579 | (4/12/4, 0.59) | 0.478 | (4/11/5, 0.59) | 157 |
Active Learning
Active learning selects the examples, labels, features, or experiments most useful for improving a model under limited budgets.
See all 321 papers