acceptodds
Under review as a conference paper at ICLR 2027

DRIP-MPU: Distilled Context Retrieval for Few-Shot Tabular MPU Classification

Abstract

Few-shot tabular multi-positive and unlabeled (MPU) classification requires distinguishing multiple positive classes and an unknown negative class using only a few labeled positives and an unlabeled pool. Scarce supervision limits task-specific PU/MPU learning, while the absence of labeled negatives prevents the direct use of tabular foundation models that require every class to be represented in context. We propose the framework DRIP-MPU, which formulates this task as learning to retrieve a multiclass context for a frozen TabPFN. TabPFN serves both as a teacher guiding context retrieval and as the final in-context classifier. Starting from labeled positives and geometrically selected diverse pseudo-negatives, a Feedback-Guided Context Retriever learns from confidence-weighted out-of-fold TabPFN predictions and geometric supervision. The retrieved pseudo-labeled examples update the context for the next round of teacher feedback, iteratively refining the context without updating TabPFN's parameters. Experiments on 25 tabular datasets show that DRIP-MPU achieves the highest average AUC and positive-class F1 among the evaluated PU/MPU baselines across 5-, 10-, and 15-shot settings. In the 5-shot setting, it improves upon the strongest baselines on these metrics by 7.7% and 46.1%, respectively. Code: https://anonymous.4open.science/r/opal-forest-7c4d.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
tabular data, few-shot, pu-learning, classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AyQSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reformulate the problem of MPU as a retrieval problem, to find the closely related data for negative classes. The solution is to combine the advatages of MPU and TabPFN. The evaluation across 25 datasets shows that the method is competitive.

Critical strengths
  1. The problem reformulation is reasonable and seems to be effective to deal with the problem
  2. The paper is well written and easy to follow
  3. The experiment includes 25 datasets with each results reported in Appendix.
Critical weaknesses
  1. The experimental evaluation is restricted to a single MPU construction protocol, with one held-out negative class per dataset. It remains unclear whether the method generalizes to more unlabeled negative classes. If there are multiple classes included in the negative set, I am not sure whether the designed method can be robust enough, especially for those unlabeled negative classes similar to labeled classes.
  2. A central contribution of the paper is the Feedback-Guided Context Retriever (FGCR), which distills predictions from a frozen TabPFN and iteratively refines the pseudo-labeled context. However, it remains unclear whether training an additional retriever is necessary, since TabPFN already provides class probabilities that could be used directly for pseudo-label selection and iterative context refinement. A natural baseline would initialize the context using the same GeoDNI procedure, obtain out-of-fold predictions from frozen TabPFN, and directly select high-confidence pseudo-labeled examples for each class to update the context over multiple rounds, without training FGCR. This baseline should use the same context budgets and refinement iterations as the proposed method. Although the authors include a w/o FGCR ablation (Table 5(b)), it is unclear whether this variant implements such a direct iterative self-training strategy. A more explicit comparison would help establish whether the performance improvements arise from the proposed learned retriever or could be achieved through simpler iterative context refinement.
Questions for authors

Please refer to weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal Answer

Thank you for carefully reading our paper and for your insightful questions and valuable comments.

W1: Additional Experiments with Multiple Hidden Negative Classes

We thank the reviewer for highlighting this important limitation. In our original experiments on multiclass datasets, each task designated only one original class as the unlabeled negative class. To address this concern, we conducted additional experiments under the most challenging 5-shot setting. For each multiclass dataset, we designated multiple original classes as hidden negative classes: two for datasets with four classes, three for datasets with five to seven classes, and four for datasets with eight or more classes. All other experimental settings remained unchanged. We compared DRIP-MPU with the four strongest baselines under the same multiple-hidden-negative protocol.

Table 1 presents the results. DRIP-MPU achieved an average AUC of 81.2581.25\\% and an average Pos.-F1 of 52.1152.11\\%, outperforming the four baselines. These results support its applicability to the multiple-hidden-negative tasks tested here.

We also acknowledge that introducing multiple hidden negative classes affects performance. While the average AUC remains close to the original result, Pos.-F1 decreases by 5.525.52 percentage points. Addressing this decline is a direction for future work.

Table 1. Average AUC and Pos.-F1 (%) over the ten multiclass datasets in the 5-shot setting with ten seeds. All rows use multiple hidden negative classes, except DRIP-MPU (single negative class).

MethodAUCPos.-F1
DRIP-MPU81.2552.11
DRIP-MPU (single negative class)81.5957.63
LaGAM-PU78.4338.28
PUET (KM2)78.4423.08
Dist-PU (KM2)76.4322.78
PULDA (KM2)76.7615.22

W2: Investigating the Performance Gains from the Learned Retriever

We thank the reviewer for suggesting this meaningful comparison. First, we clarify that the “w/o FGCR” ablation in the original manuscript did not perform iterative self-training and therefore does not directly answer this question. To address it, we implemented the suggested direct teacher-only self-training variant. It initializes the context using the same GeoDNI results, then iteratively selects pseudo-labeled examples and updates the context based on class predictions from the frozen TabPFN, without training an FGCR retriever.

As shown in Table 2, under the 5-shot setting, this variant achieves an average AUC of 73.99% and an average Pos.-F1 of 54.98%. This shows that direct self-training with TabPFN feedback is a competitive simple alternative. DRIP-MPU further improves the average AUC and Pos.-F1 by 1.42 and 1.77 percentage points, respectively. The learned retriever in the full method thus provides additional average gains on both metrics.

Table 2. Performance comparison between DRIP-MPU and the teacher-only self-training variant in the 5-shot setting.

MethodAUC (%)Pos.-F1 (%)
Teacher-only self-training73.99±1.5573.99 \pm 1.5554.98±1.1754.98 \pm 1.17
DRIP-MPU75.41±1.10\mathbf{75.41 \pm 1.10}56.75±0.90\mathbf{56.75 \pm 0.90}

Thanks for running the additional experiments. The results on the ten multiclass datasets address most of my concern about using a single hidden negative class. I’d still be interested in how the method performs when a hidden class closely resembles a labeled class. The teacher-only comparison answers my second question. I keep my original positive rating.

Reviewer_MZrcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses few-shot tabular multi-positive-unlabeled (MPU) classification. A frozen TabPFN plays two roles: a teacher that gives out-of-fold probabilities, and the final in-context classifier. The teacher's outputs train a small MLP "retriever" (FGCR) that selects and pseudo-labels unlabeled rows for the context, starting from geometrically chosen diverse pseudo-negatives (GeoDNI). On 25 OpenML datasets, it reports 75.4% AUC and 56.7% Pos.-F1 at 5 shots, well above the PU/MPU baselines.

Critical strengths
  1. The problem is well motivated, and framing it as context construction for a tabular foundation model is a sensible idea.
  2. The experiments are broad, including 25 datasets, 10 seeds, and 3 shot budgets.
  3. The frozen-TabPFN control (Table 5a) and the component ablations address the obvious "it's just TabPFN" objection.
  4. The paper reports results with oracle priors, which is honest.
Critical weaknesses
  1. Missing baselines for the core claim.

    a. The w/o-FGCR variant (GeoDNI + TabPFN, 52.5 F1) already beats every baseline+TabPFN (best 47.6). So much of the gain comes from the geometric heuristic, not from learned retrieval.

    b. There is no comparison to simple self-training with TabPFN directly (no student MLP), or to kNN/prototype/outlier-based pseudo-negatives plus TabPFN.

    c. FGCR is an MLP on the same features, so the case that "distillation" adds value beyond smoothing is not made.

    d. LoCalPFN and TuneTables are discussed but never adapted as baselines.

  2. Hyperparameter selection and leakage.

    a. About 15 hyperparameters were tuned. The sensitivity plots (Fig. 3) use test-set metrics, and the defaults sit at the peaks.

    b. No validation protocol is described, and no evidence is given that baselines received comparable tuning.

  3. Hidden prior knowledge.

    The context budgets ξ are fixed fractions of |U| (positives 5%, negatives 10→20%). This implicitly encodes class proportions, while baselines must estimate priors with KM2. Fig. 3j shows AUC degrading with ξ, which undercuts the "robust" claim.3.

  4. Metric and statistics concerns.

    a. Many baselines get F1 ≈ 0–3% (BBE-PU, ScalePU, nnPU, AREA), which suggests thresholding or calibration failure rather than weak ranking. The 46.1% relative gain is therefore inflated.

    b. The reported SDs are over seed-wise dataset averages, which hides cross-dataset variance. There are no significance tests, and several test sets are tiny (sonar has 208 instances in total).

    c. DRIP-MPU's AUC is near or below chance on several datasets (climate-model 45.2, dresses-sales, GesturePhase).

    d. With true priors, PUET/Dist-PU beat it badly on climate-model (95.4 vs 51.1 F1) and are competitive on segment and sonar. This is not discussed.

  5. Inconsistencies. Section 5.1 says 10 baselines, but the appendix tables include 17 methods (VPU, PAN, HolisticPU, P3Mix-C/E and others) that are absent from Table 1. Table 2 shows only the four baselines that improved with TabPFN.

  6. Realism and cost.

    a. The negative class is a single held-out class, unlike a heterogeneous "everything else" negative.

    b. Runtime is not reported: 30 rounds × 3-fold TabPFN-2.5 on pools up to ~8k rows.

    c. Possible overlap between OpenML datasets and TabPFN-2.5's real-data training is not addressed.

  7. No analysis of the self-reinforcing loop.

    There is no pseudo-label purity per round, negative recall, or convergence and confirmation-bias analysis. Contradicting the paper's own mechanism claim, the OOF ablation yields only a modest drop (−1.8).

Questions for authors
  1. How were hyper-parameters chosen without test labels, and how do results change under a validation-based protocol?

  2. Performance when the true negative fraction differs strongly from the assumed ξ budgets?

  3. Results for teacher-only self-training (no FGCR), and for GeoDNI + TabPFN with oracle negatives as an upper bound?

  4. Wall-clock comparison to baselines?

5 comments
AuthorsOpenReview ↗

Rebuttal Answer [part1]

Thank you for pointing out the weaknesses of our paper. We first address your four questions, then respond to the seven weaknesses you identified.

Q1: Hyperparameter Settings

We apologize that our previous description gave the impression that we selected the hyperparameters based on test-set performance. The experiments used a predefined parameter configuration that was applied consistently across all datasets and random seeds. No parameters were selected based on test-set performance. This configuration is intended to provide a unified protocol for applying DRIP-MPU to different datasets, and its parameter values are documented in Appendix B of the manuscript.

Appendix E of the manuscript presents a sensitivity analysis of this predefined configuration. The analysis evaluates robustness around the fixed settings; it was not used to select an optimal configuration for each test dataset. Even the best-performing hyperparameter choices in this analysis yield only modest changes relative to the predefined settings. We therefore believe that precise tuning on a validation set is not necessary to achieve strong performance.

We additionally evaluated a stratified 60/20/20 train/val/test protocol on all 25 datasets with ten seeds in the 5-shot setting. Keeping all other hyperparameters fixed, we selected the checkpoint with the highest validation AUC among epochs 1–600, without using test labels. Compared with fixed stopping at epoch 600 on identical training data, mean test AUC increased from 75.00% to 76.09%, and Pos.-F1 from 54.93% to 55.57%.

Q2: Large Discrepancies Between the True Negative Fraction and ξ\xi

Thank you for this question. The negative-context budget ξCneg\xi_{C_{\mathrm{neg}}} is a manually specified retrieval budget, rather than an estimate of the negative-class prior. Our aim is to construct an informative pseudo-labeled context, whose class composition need not reproduce the population proportions. GeoDNI initializes a small, geometrically plausible and diverse negative set, and FGCR subsequently updates and supplements the context.

The existing benchmark includes substantial mismatches between the retrieval budget and the true negative fraction. The results on the four datasets in Table 2 below, all of which have markedly imbalanced positive-to-negative class ratios, show that strong performance does not require the retrieval budget to match the true negative fraction.

Table 1. True negative fractions and preset budgets.

DatasetTrue (%)Budget (%)
wdbc63.3320.00
analcatdata_author38.5115.00
segment85.9520.00
mfeat-karhunen10.2915.00

Table 2. AUC and Pos.-F1 (%) in the 5-shot setting. Bold and underlined values indicate the best and second-best means for each dataset and metric, respectively.

Methodwdbc AUCwdbc Pos.-F1analcatdata_author AUCanalcatdata_author Pos.-F1segment AUCsegment Pos.-F1mfeat-karhunen AUCmfeat-karhunen Pos.-F1
DRIP-MPU (Ours)99.0±1.1\mathbf{99.0 \pm 1.1}91.2±3.4\mathbf{91.2 \pm 3.4}99.1±1.6\mathbf{99.1 \pm 1.6}95.3±3.9\mathbf{95.3 \pm 3.9}97.4±3.8‾\underline{97.4 \pm 3.8}60.4±25.4‾\underline{60.4 \pm 25.4}95.3±3.1\mathbf{95.3 \pm 3.1}82.4±3.8\mathbf{82.4 \pm 3.8}
BBE-PU93.0±3.3‾\underline{93.0 \pm 3.3}9.8±14.19.8 \pm 14.192.6±2.492.6 \pm 2.40.0±0.00.0 \pm 0.092.4±4.992.4 \pm 4.90.0±0.00.0 \pm 0.090.5±0.790.5 \pm 0.70.0±0.00.0 \pm 0.0
LaGAM-PU88.2±7.788.2 \pm 7.761.3±11.9‾\underline{61.3 \pm 11.9}95.2±1.695.2 \pm 1.667.9±3.0‾\underline{67.9 \pm 3.0}85.4±8.685.4 \pm 8.650.4±13.750.4 \pm 13.791.9±1.7‾\underline{91.9 \pm 1.7}68.4±3.7‾\underline{68.4 \pm 3.7}
CGenPU80.6±20.880.6 \pm 20.859.7±18.559.7 \pm 18.558.5±9.758.5 \pm 9.725.5±7.225.5 \pm 7.276.3±17.476.3 \pm 17.440.3±16.640.3 \pm 16.661.9±5.261.9 \pm 5.211.4±4.311.4 \pm 4.3
PUET (KM2)87.9±4.887.9 \pm 4.88.3±14.58.3 \pm 14.595.1±2.395.1 \pm 2.327.5±12.427.5 \pm 12.494.6±5.894.6 \pm 5.848.6±15.648.6 \pm 15.689.8±0.889.8 \pm 0.87.8±3.07.8 \pm 3.0
ScalePU (KM2)86.9±7.586.9 \pm 7.50.0±0.00.0 \pm 0.079.3±2.679.3 \pm 2.61.8±5.61.8 \pm 5.692.2±4.292.2 \pm 4.27.1±5.77.1 \pm 5.775.9±2.375.9 \pm 2.30.0±0.00.0 \pm 0.0
nnPU (KM2)71.6±13.771.6 \pm 13.70.5±1.50.5 \pm 1.585.9±5.685.9 \pm 5.63.0±9.63.0 \pm 9.673.2±2.573.2 \pm 2.50.0±0.00.0 \pm 0.082.4±1.782.4 \pm 1.70.0±0.00.0 \pm 0.0
Dist-PU (KM2)88.3±11.388.3 \pm 11.310.6±20.010.6 \pm 20.096.5±3.8‾\underline{96.5 \pm 3.8}55.8±19.955.8 \pm 19.993.9±3.793.9 \pm 3.743.6±32.043.6 \pm 32.090.8±1.090.8 \pm 1.027.5±6.327.5 \pm 6.3
PULDA (KM2)78.9±8.078.9 \pm 8.02.3±3.22.3 \pm 3.286.5±2.986.5 \pm 2.93.4±5.53.4 \pm 5.598.1±1.0\mathbf{98.1 \pm 1.0}76.9±12.6\mathbf{76.9 \pm 12.6}89.7±1.489.7 \pm 1.42.6±1.02.6 \pm 1.0
UREA (KM2)62.0±4.662.0 \pm 4.637.2±12.237.2 \pm 12.255.8±2.555.8 \pm 2.514.1±5.914.1 \pm 5.984.3±8.484.3 \pm 8.451.2±15.951.2 \pm 15.952.7±1.452.7 \pm 1.45.1±2.05.1 \pm 2.0
AREA (KM2)71.5±12.671.5 \pm 12.63.3±2.23.3 \pm 2.257.1±7.557.1 \pm 7.52.7±2.92.7 \pm 2.990.1±7.190.1 \pm 7.17.2±4.47.2 \pm 4.459.5±2.459.5 \pm 2.42.5±1.02.5 \pm 1.0
AuthorsOpenReview ↗

Rebuttal Answer [part2]

Q3: Teacher-Only Self-Training and an Oracle-Negative Upper Bound

We added a teacher-only self-training variant that removes FGCR and its training process. It retains the same GeoDNI initialization, out-of-fold (OOF) prediction mechanism, context budgets, number of refinement rounds, and final prediction configuration. In each round, it directly updates the pseudo-labels and context using TabPFN's predicted probabilities. This comparison isolates the additional contribution of FGCR over direct self-training.

We evaluated this variant under the most challenging 5-shot setting. As shown in Table 3 below, it achieves an average AUC of 73.99% and an average Pos.-F1 of 54.98%, indicating that direct self-training with TabPFN feedback is a competitive simple alternative. DRIP-MPU further improves the average AUC and Pos.-F1 by 1.42 and 1.77 percentage points, respectively. Thus, the learned retriever in the full method provides additional average gains on both metrics.

Table 3. Performance comparison between DRIP-MPU and the teacher-only self-training variant in the 5-shot setting.

MethodAUC (%)Pos.-F1 (%)
Teacher-only self-training73.99±1.5573.99 \pm 1.5554.98±1.1754.98 \pm 1.17
DRIP-MPU75.41±1.10\mathbf{75.41 \pm 1.10}56.75±0.90\mathbf{56.75 \pm 0.90}

To address your second question, we replaced the pseudo-negative examples selected by GeoDNI with true negative examples sampled uniformly at random under the same budget, while retaining positive-example retrieval. This provides an oracle comparison with access to additional label information. As shown in Table 4 below, Oracle-DRIP improves average AUC by 8.28.2 percentage points over DRIP-MPU, indicating that true negative labels can further improve ranking performance.

However, Pos.-F1 decreases. True negative labels ensure label correctness, but do not guarantee that the resulting context is better suited to few-shot positive-class recognition. Whereas DRIP-MPU progressively selects negative examples, Oracle-DRIP uses a fixed set of randomly sampled true negatives from initialization onward. This changes the distribution of negative examples and the early class balance. With limited positive support, these changes lead the model to predict the positive class more conservatively, reducing false positives but increasing false negatives and ultimately lowering Pos.-F1.

Table 4. Comparison of DRIP-MPU and Oracle-DRIP in the five-shot setting (%).

MethodAUCPos.-F1
DRIP-MPU75.41±1.1075.41 \pm 1.1056.75±0.90\mathbf{56.75 \pm 0.90}
Oracle-DRIP83.63±0.89\mathbf{83.63 \pm 0.89}51.34±1.4251.34 \pm 1.42

Q4: Runtime Comparison with Baseline Methods

Thank you for pointing out that the original manuscript did not report runtime. We have added the average runtime per run under different shot settings, as shown in the table. DRIP-MPU takes 61.08 seconds on average, which is close to CGenPU but longer than most other baselines. The additional time is mainly due to multiple rounds of retriever training and context updates. Given the substantial improvements in AUC and positive-class F1, we consider this runtime acceptable.

Table 5. Average recorded runtime (seconds) per dataset and seed. The final column averages the three shot settings.

Method5-shot10-shot15-shotAverage
DRIP-MPU (Ours)60.2063.6059.4461.08
BBE-PU7.257.297.197.24
LaGAM-PU41.5737.7645.8641.73
CGenPU59.6758.4559.0759.06
PUET (KM2)1.211.782.101.70
ScalePU (KM2)4.684.794.574.68
nnPU (KM2)3.253.153.143.18
Dist-PU (KM2)10.5410.6510.7510.65
PULDA (KM2)10.4610.6310.6310.57
UREA (KM2)11.5311.1611.7711.49
AREA (KM2)11.1510.6311.0510.94
AuthorsOpenReview ↗

Rebuttal Answer [part3]

W1

a: Performance Gains from FGCR

All ablated variants in Table 3 outperform the strongest baseline+TabPFN (47.60% Pos.-F1), even without GeoDNI; their Pos.-F1 ranges from 50.39% to 55.59%. Relative to full DRIP-MPU (56.75%), removing GeoDNI, FGCR, Ldistilled\mathcal{L}_{\mathrm{distilled}}, and Lgeo\mathcal{L}_{\mathrm{geo}} reduces Pos.-F1 by 6.36, 4.21, 2.05, and 1.16 percentage points, respectively. In particular, adding FGCR raises Pos.-F1 from 52.54% to 56.75%. Thus, geometric initialization, learned retrieval, and both supervision signals contribute; neither GeoDNI nor FGCR alone accounts for the full gain.

b: Self-Training Variants and Alternative Pseudo-Negative Construction

Our response to Q3 covers the self-training variants. We also replaced GeoDNI with KNN-distance, positive-prototype-distance, and Isolation Forest methods, keeping all other components and hyperparameters fixed. In Table 6, DRIP-MPU exceeds the strongest alternative, the prototype method, by 5.06 AUC and 12.31 Pos.-F1 percentage points on average, supporting GeoDNI's effectiveness.

Table 6. Five-shot results on 25 datasets for three alternatives to GeoDNI.

MethodAUC (%)Pos.-F1 (%)
KNN negatives + FGCR69.71±1.7069.71 \pm 1.7043.95±1.1243.95 \pm 1.12
Prototype negatives + FGCR70.35±1.4670.35 \pm 1.4644.44±1.2744.44 \pm 1.27
Isolation Forest negatives + FGCR65.14±1.5665.14 \pm 1.5640.80±2.5140.80 \pm 2.51
DRIP-MPU75.41±1.10\mathbf{75.41 \pm 1.10}56.75±0.90\mathbf{56.75 \pm 0.90}

c: Teacher Distribution Matching versus Smoothing

We replaced teacher distribution matching with a local smoothing loss on a fixed 10-nearest-neighbor graph in the input feature space of the unlabeled training pool, using normalized Gaussian distance weights. Across 25 datasets, ten seeds, and the 5-shot setting, DRIP-MPU exceeds this control by 2.70 AUC and 3.87 Pos.-F1 percentage points on average (Table 7). This supports an additional benefit from teacher distribution matching relative to local smoothing.

Table 7. Average results (%) across 25 datasets in the 5-shot setting.

MethodAUCPos.-F1
Smoothing replacement72.7152.88
DRIP-MPU75.4156.75

d: LoCalPFN and TuneTables

We discuss LoCalPFN and TuneTables because they relate to retrieval and context optimization. Their original settings require labeled support data, whereas ours has few labeled positives and no labeled negatives. Adapting them with constructed pseudo-negatives would introduce choices about pseudo-labels, noise handling, and tuning. The results would therefore be difficult to interpret as representative of the original methods or as a fair comparison.

W2

a: Hyperparameter Settings

We agree that the default settings appearing near peaks in some Appendix E curves could suggest test-based selection. They were fixed before test-set evaluation as a unified configuration across datasets. Figure 3 is a post hoc sensitivity analysis and was not used to choose the defaults. Several defaults, including α\alpha and ξc\xi_c, do not lie at their curve peaks (Fig. 3(h), (j), and (k)). Varying one parameter at a time changes performance only modestly.

b: Baseline Hyperparameters

For seven baselines, we used PU-Bench [1] implementations and their method-specific defaults. We implemented ScalePU, UREA, and AREA because we could not find public code, following the hyperparameters in their original papers.

[1] Chen, Q., Zhang, H., et al. “PU-Bench: A Unified Benchmark for Rigorous and Reproducible PU Learning.” ICLR 2026.

W3: Use of the True Class Prior

ξc\xi_c limits the number of samples per class in TabPFN's context; it does not estimate class proportions in the unlabeled pool. We used neither true class proportions nor test labels to set it. Its role differs from the KM2 prior used for baseline risk estimation. As shown in our response to Q2, DRIP-MPU also performs well when the true negative fraction differs substantially from the budget set by ξc\xi_c.

W4

a: Baseline Decision Rules and the Pos.-F1 Gain

We used PU-Bench's default decision rules for seven baselines. For our implementations, ScalePU thresholds sigmoid scores at 0.5, while UREA and AREA take the argmax of their original class scores. Class-specific PU classifiers in multi-positive tasks follow the corresponding rules, without extra threshold calibration.

Near-zero Pos.-F1 may reflect scarce supervision, prior-estimation errors, or calibration rather than poor ranking; hence we report both AUC and Pos.-F1. The reported 46.1% relative Pos.-F1 gain compares DRIP-MPU with LaGAM-PU, the strongest baseline on this metric, rather than a near-zero-F1 baseline. Pos.-F1 rises from 38.8% to 56.7%, a 17.9-point absolute gain. Excluding near-zero-F1 baselines does not change this comparison.

Reviewer_2BucOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes DRIP-MPU for few-shot tabular classification with few labeled positive examples and no labeled negatives. Instead of directly learning a classifier, it iteratively trains a retriever using a frozen TabPFN as teacher; the retriever pseudo-labels and selects unlabeled examples to construct better context for TabPFN, which ultimately performs classification. Its key distinction is learning the context rather than the final classifier, though the iterative feedback process has no convergence guarantee.

Critical strengths
  • Instead of training the final classifier, the method learns to construct pseudo-labeled context for a frozen TabPFN, using TabPFN both as teacher and final classifier. This clearly differentiates it from standard PU/MPU approaches.

  • It evaluates on 25 datasets, across 5-, 10-, and 15-shot settings, and reports the best average AUC and positive-class F1 among the evaluated baselines.

Critical weaknesses
  • The method iteratively uses frozen TabPFN predictions to update the retriever, whose pseudo-labeled context is then fed back to TabPFN for the next round. However, the paper does not establish that this feedback process converges or avoids reinforcing erroneous pseudo-labels. Although sensitivity to the number of refinement rounds is reported, this does not demonstrate convergence, raising concerns about stability and generalization to different datasets.

  • The proposed method relies on several heuristically designed components, including geometry-based scoring, confidence weighting, context selection, and refinement rules, with many associated hyperparameters.

  • The method section first introduces multiple scoring functions, weighting terms, losses, and hyperparameters, making the approach unnecessarily difficult to follow. The central idea could be explained more intuitively before the detailed formulation: TabPFN teaches the retriever, the retriever constructs a better pseudo-labeled context, and the updated context is fed back to TabPFN iteratively.

Questions for authors

.

1 comment
AuthorsOpenReview ↗

Rebuttal Answer

Thank you for carefully reading our paper and providing valuable comments.

W1: Additional Evidence of Convergence in the Feedback Process

Thank you for pointing out the lack of convergence evidence for the feedback process in our paper. To examine the convergence behavior and stability of the feedback process, we tracked AUC, Pos.-F1, pseudo-label accuracy, and the joint FGCR loss throughout training in the 5-shot setting. AUC, Pos.-F1, and pseudo-label accuracy generally improve during early training and then remain broadly stable. The mean joint loss decreases sharply at first and more gradually thereafter, with small fluctuations in later epochs. These trends provide empirical evidence of stable training and apparent convergence.

The training curves are available at the anonymous link and have also been included in Appendix F (Figure 4) of the revised manuscript. The figure displays the first 200 FGCR training epochs, by which point the metrics have largely stabilized.

W2: The Role of Each Component and Hyperparameter Sensitivity

Thank you for raising this point. Each core component is designed to address a specific challenge. Because no labeled negative examples are available initially, we use a geometry-based score to initialize the negative-class context. Because unreliable pseudo-labels may affect the teacher's predictions, we introduce confidence optimization. To further refine the context and control its size and quality, we design strict rules for context selection and updating.

The ablation results in Table 3 validate the contributions of these components, while the one-at-a-time hyperparameter sensitivity results in Figure 3 of Appendix E show that their effectiveness does not depend on precise hyperparameter tuning.

W3: Improving the Presentation of the Method

Thank you for your suggestion on the presentation of the method. The overview in Section 4.1 and Figure 2 describe the core idea and the full feedback process. Following your suggestion, we have expanded the intuitive explanation in Section 4.1 of the revised manuscript to make the process and the connections between its components clearer.

DRIP-MPU first uses geometric information to construct an initial pseudo-labeled context, enabling TabPFN to make context-conditioned predictions for unlabeled samples. TabPFN then acts as a teacher: its out-of-fold (OOF) predictions provide supervision for training the retriever. The retriever uses this signal to construct a more informative pseudo-labeled context from the unlabeled data for the next round. Feeding the updated context back into TabPFN creates an iterative feedback process. After the final round, TabPFN serves as the in-context classifier and uses the resulting context to predict the labels of unlabeled samples.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading and constructive feedback. We have addressed their questions, conducted additional experiments, and revised the manuscript accordingly. The main responses are organized around three topics: the feedback process and evaluation protocol, learned retrieval and context construction, and the scope and runtime of DRIP-MPU.

Feedback process and evaluation protocol

  • We revised Section 4.1 to explain the roles of TabPFN, FGCR, and context updates in the feedback process more clearly. (Reviewer 2Buc, Q3)

  • We added curves for AUC, positive-class F1, pseudo-label accuracy, and loss. The metrics improve early and then remain broadly stable, providing evidence of stable training. (Reviewer 2Buc, Q1)

  • We clarified that the main experiments used one predefined hyperparameter configuration across datasets and seeds. A validation-based checkpoint experiment on all 25 datasets and the sensitivity analysis show strong performance under a validation-based protocol and limited sensitivity to the examined settings. (Reviewers 2Buc, Q2; MZrc, Q1)

Learned retrieval and context construction

  • We compared DRIP-MPU with teacher-only iterative self-training under the same initialization, out-of-fold predictions, context budgets, and refinement rounds. DRIP-MPU achieves 75.41% AUC and 56.75% positive-class F1, versus 73.99% and 54.98% for self-training, showing FGCR's additional contribution. (Reviewers MZrc, Q3; AyQS, Q2)

  • We clarified how geometry-based initialization, confidence optimization, and context selection and updating address distinct challenges. The component ablations and hyperparameter sensitivity results support their contributions and robustness. (Reviewer 2Buc, Q2)

  • An oracle-negative variant achieves 83.63% AUC and 51.34% positive-class F1, versus 75.41% and 56.75% for DRIP-MPU. Correct negative labels improve ranking, while context composition also affects positive-class recognition. (Reviewer MZrc, Q3)

  • We clarified that the negative-context budget controls the number of retrieved examples rather than estimating the true negative-class proportion. DRIP-MPU maintains strong performance on four highlighted datasets where the two differ substantially. (Reviewer MZrc, Q2)

Scope and runtime

  • With multiple hidden negative classes on ten multiclass datasets, DRIP-MPU achieves 81.25% AUC and 52.11% positive-class F1, outperforming the four compared baselines. (Reviewer AyQS, Q1)

  • DRIP-MPU averages 61.08 seconds per run, close to CGenPU and longer than most other baselines. Given its AUC and positive-class F1 gains, we consider this runtime acceptable. (Reviewer MZrc, Q4)

We hope these revisions and additional results address the reviewers' central concerns.

Revision locations in the manuscript

  • Section 4.1 expands the intuitive overview of the method and feedback process.
  • Section 5.1 clarifies dataset selection, including the exclusion of datasets used in TabPFN pretraining.
  • Section 5.3 now points to the complete frozen-TabPFN comparison with all ten main baselines in Appendix C.
  • New Appendix F presents feedback-process training dynamics.
  • Appendix G clarifies the main baselines, supplementary methods, and oracle variants in the complete per-dataset results.
  • New Appendix H reports cross-dataset variability.
  • New Appendix I compares DRIP-MPU with teacher-only self-training.
  • New Appendix J evaluates multiple hidden negative classes.
  • New Appendix K reports runtime comparisons.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.