acceptodds
Under review as a conference paper at ICLR 2027

TiltMIL: Learning Finite Statistics from Functional Bag Representations for Multiple Instance Learning

Abstract

Multiple instance learning (MIL) predicts bag-level labels from multiple instances using only bag-level supervision and is widely applied to whole-slide image analysis in computational pathology. However, existing MIL methods typically compress an instance set directly into a finite-dimensional representation whose retained information form is predetermined by the representation mechanism, although different tasks may require different forms of compositional information. We propose TiltMIL, a Bag composition modeling framework based on finite cumulant-generating function (CGF) readouts. TiltMIL represents a bag as an empirical probability measure in a task-specific feature space, uses the CGF as a functional distributional representation, and learns task-relevant tilt locations to extract finite CGF function values and cumulant statistics for prediction. A lightweight hierarchical fusion mechanism combines these readouts, while empirical perturbation consistency improves stability to finite composition changes. Experiments on three WSI-MIL benchmarks with two feature encoders demonstrate consistent performance improvements across multiple pathology classification tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multiple Instance Learning, Set Representation Learning, Computational Pathology, Whole-Slide Image Classification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 5 reviews, above 85% of the others.
Reviewer_oAAVOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TiltMIL, a multiple instance learning method that uses the cumulant-generating function (CGF) to capture different statistical properties of a bag, including higher-order information. The method is evaluated on three pathology datasets, with particularly strong results on PANDA. Ablation studies support the contribution of second-order statistics and consistency regularisation.

Critical strengths
  1. Interesting approach to bag representation.

Using learned CGF readouts to capture different aspects of a bag's distribution is an interesting idea. The inclusion of second-order statistics is particularly relevant to pathology, where tissue composition can be important.

  1. Strong experimental results.

The evaluation covers ten baselines and two feature extractors. Results on PANDA are particularly convincing, with TiltMIL achieving 80.73% accuracy compared with 76.31% for the strongest baseline. The ablations support the importance of second-order statistics and consistency regularisation. The robustness experiments are also useful.

Critical weaknesses
  1. Unclear advantage over simpler pooling methods.

My main concern is that the CGF readouts are closely related to attention-weighted means and variances. It is therefore unclear whether the CGF formulation offers advantages over a simpler multi-query attention model with statistical pooling. A matched comparison would help establish the novelty of the approach.

  1. Evaluation details and consistency of improvements.

The PANDA results are strong, but improvements on TCGA-NSCLC and TCGA-RCC are much smaller. It would be useful to establish whether these differences are consistent across folds. The authors should also clarify whether cross-validation splits are patient-independent, particularly for TCGA datasets containing multiple slides per patient.

Questions for authors

Could the authors compare TiltMIL with a multi-query attention model using weighted mean and variance pooling under similar training conditions?

Are the cross-validation splits patient-independent, and are the same partitions used for all methods?

Could the authors provide fold-level comparisons for the TCGA datasets to assess whether the smaller improvements are consistent?

2 comments
AuthorsOpenReview ↗

Response to Reviewer oAAV (1/2)

We thank the reviewer for the effort in reviewing our work. Our point-to-point responses are as follows.

W1 & Q1: On Comparisons with Multi-Query Statistical Pooling


It is therefore unclear whether the CGF formulation offers advantages over a simpler multi-query attention model with statistical pooling. A matched comparison would help establish the novelty of the approach.

Could the authors compare TiltMIL with a multi-query attention model using weighted mean and variance pooling under similar training conditions?

We thank the reviewer for the suggestion. We considered two multi-query attention baselines: Pooling by Multihead Attention (PMA [1]), which uses learnable seed queries to aggregate instance features, and ACMIL-MHA [2], our multi-head attention variant of ACMIL's multi-branch aggregation. For both methods, we extended the original attention-weighted feature aggregation with weighted variance pooling (Var), using the corresponding instance attention weights. Under the same dataset and training protocol, we also examined capacity-matched variants (CM) and the combination of variance pooling and EPC regularization (Var+EPC).

On PANDA–UNI, adding weighted variance improves PMA's Macro-F1 from 70.17% to 71.44%, while further incorporating EPC increases it to 73.13%. Similarly, ACMIL-MHA improves from 67.75% to 69.04% with variance pooling and reaches 71.12% with Var+EPC. Nevertheless, TiltMIL achieves 77.35%, outperforming both extended attention-based baselines. On TCGA-NSCLC–UNI, the differences are smaller. TiltMIL achieves an Accuracy of 93.38% and a Macro-F1 of 93.36%, compared with 92.71% and 92.69% for ACMIL-MHA+Var+EPC, although its AUC is slightly lower (96.94% vs. 97.17%). Furthermore, the original Table 2 shows that removing CGF function values or variance from TiltMIL reduces predictive performance.

These results suggest that augmenting attention-based aggregation with weighted variance and EPC does not fully reproduce the performance of TiltMIL's combined finite statistical readout design, particularly on PANDA.

MethodPANDA AccPANDA AUCPANDA F1-scoreNSCLC AccNSCLC AUCNSCLC F1-score
PMA [1]75.01 ± 1.8693.88 ± 0.3870.17 ± 2.0090.31 ± 1.9696.12 ± 1.9990.26 ± 1.95
PMA+Var76.10 ± 1.8293.98 ± 0.5571.44 ± 2.3991.55 ± 2.4796.34 ± 2.6591.51 ± 2.47
PMA+CM74.88 ± 1.5993.95 ± 0.3870.60 ± 1.6990.88 ± 2.5196.37 ± 2.4990.88 ± 2.51
PMA+Var+EPC76.31 ± 0.6194.00 ± 0.1873.13 ± 0.7290.84 ± 2.7496.35 ± 2.1290.80 ± 2.76
ACMIL-MHA [2]72.53 ± 2.1192.90 ± 0.6067.75 ± 3.5591.37 ± 3.7996.93 ± 2.4391.35 ± 3.80
ACMIL-MHA+Var73.68 ± 2.0993.19 ± 0.5969.04 ± 2.6192.10 ± 3.2597.03 ± 1.8492.07 ± 3.31
ACMIL-MHA+CM73.07 ± 2.3293.06 ± 0.5468.43 ± 2.9191.64 ± 3.1297.02 ± 1.9191.62 ± 3.14
ACMIL-MHA+Var+EPC74.43 ± 1.2093.88 ± 0.4471.12 ± 1.5792.71 ± 2.4097.17 ± 1.5992.69 ± 2.45
TiltMIL80.73 ± 0.4696.04 ± 0.1777.35 ± 0.6593.38 ± 1.9796.94 ± 1.8793.36 ± 1.97

[1] Set Transformer: A Framework for Attention-Based Permutation-Invariant Neural Networks

[2] Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification

AuthorsOpenReview ↗

Response to Reviewer oAAV (2/2)

W2 & Q2 & Q3: Evaluation details and consistency of improvements.

Are the cross-validation splits patient-independent, and are the same partitions used for all methods?

For the TCGA datasets, we strictly enforce patient-independent cross-validation, ensuring that all slides from the same patient remain within the same training, validation, or test partition, thereby preventing patient-level data leakage. For PANDA, patient identifiers are not provided in the original dataset, so we adopt a standard random five-fold split at the slide level. All methods are evaluated using the same data partitions within each experimental setting to ensure fair comparisons.

Could the authors provide fold-level comparisons for the TCGA datasets to assess whether the smaller improvements are consistent?

We thank the reviewer for the suggestion. We conducted paired fold-level comparisons on TCGA-NSCLC and TCGA-RCC using identical patient-independent five-fold partitions. Under ResNet-50 features, TiltMIL improves NSCLC AUC in all five folds, while the other metrics exhibit moderate fold-level variability. Under UNI features, the smaller gains are less consistent: NSCLC Accuracy and Macro-F1 improve in only two of five folds, while RCC shows mixed improvements across metrics. These analyses provide a more detailed assessment of fold-level performance and complement the aggregate results reported in Table 1.

Fold-Level Performance Comparison on TCGA Datasets (ResNet-50).

Dataset / MetricFold 1 ΔFold 2 ΔFold 3 ΔFold 4 ΔFold 5 ΔMean ΔImproved Folds
NSCLC Acc-0.91+4.55+0.47-2.46+1.49+0.633
NSCLC AUC+0.78+2.03+0.32+1.64+0.43+1.045
NSCLC Macro-F1-0.84+4.56+0.52-2.37+1.48+0.673
RCC Acc+0.54+0.56-3.80-5.10+4.40-0.683
RCC AUC+0.07+1.97-1.47-0.55+0.95+0.193
RCC Macro-F1+1.48+0.60-4.09-3.81+6.56+0.153

Fold-Level Performance Comparison on TCGA Datasets (UNI).

Dataset / MetricFold 1 ΔFold 2 ΔFold 3 ΔFold 4 ΔFold 5 ΔMean ΔImproved Folds
NSCLC Acc-1.37+2.02-0.47-0.49+0.50+0.042
NSCLC AUC-0.42-1.11-1.07-1.41-0.62-0.930
NSCLC Macro-F1-1.37+1.96-0.48-0.48+0.50+0.022
RCC Acc0.00-1.110.00+2.04+1.10+0.412
RCC AUC-0.07-0.50-0.28+0.10+0.01-0.152
RCC Macro-F1+0.01-0.42+0.56+2.96-0.13+0.593

Δ = TiltMIL − best baseline. The fixed reference baselines are ACMIL for NSCLC, 2DMamba for RCC–ResNet-50, and PTCMIL for RCC–UNI.

Reviewer_43z8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is about multiple instance learning (MIL) in which there is a unique label for a whole slide and a bag of patches (parts of the slide) that subtend the same label. In different pathologies, different aspects can be used for classification: tumor classification relies on the present of a localized lesion, whereas cancer grading depends on the proportions of different tissues. So, its is important not to choose the representation of the bag before classification. Here, contrary to classical methods that summarize instances into a chosen finite representation, they represent the bag of instances/patches as a distribution, described by the cumulant-generating function (CGF), from which they learn adaptative task-relevant information and location for the final prediction. They evaluate their methods on whole-slide image benchmarks (PANDA, TCGA-NSCLC and TCGA-RCC). They show substantial improvements for PANDA, that is prostate cancer grading.

Critical strengths
  1. A clear motivated separation between representation and supervision in the MIL framework They clearly state the principles of TiltMIL that can be schematized as follows: they construct a richer representation of the bag and then decide which information should be used for prediction.
  • At self-supervised level: Instance/patch → feature space → empirical probability measure over instances → represented by the cumulant-generating function → extract some explicit statistics
  • At supervision level: learn tilt locations (tilt!=patch, it corresponds to a ponderation over patches) subtending classification → fusion with base statistics (without ponderation) → classification They clearly motivate their approach by first proving mathematically and showing that, in some cases, adding a patch can decrease classification performance, thus it is necessary to give more strength to some instances.
  1. A sound validation on WSI datasets The authors don’t stay at the theoretical level, and prove that their approach can be applied and improve MIL learning on classical benchmarks, particularly on PANDA.
Critical weaknesses

Main weaknesses that I see are on the reporting of the experimental results and on the ablation study:

  1. Needs better significativity testing Even if it is often the case in the MIL literature, reporting standard deviations in cross-validation with only one seed doesn’t look enough. I would at least have expected a repeat of the experiments with another seed, and an acknowledgment that other significativity testings could have been done.

  2. Could EPC alone explain the results? There are a lot of changes between the tested benchmarks and TiltMIL. The ablation has only been done inside the TiltMIL framework. We are left wondering if EPC alone could improve in a similar way the results of classical benchmarks.

Questions for authors

Questions that could change my assessment

  • Have you added EPC on classical beanchmarks, to be sure that EPC alone is not enough to improve the results?
  • Have you tested Table 1 with another seed to be sure that the results hold? Even if the results don’t hold for TCGA-NSCLC and TCGA-RCC, where the results can already be close to the “plateau”, I think it is necessary to have at least the variability due to the seeds. More globally, it would be useful to assess and discuss the significativity of the results (at least, how you could assess in future works, and available datasets, the variance due to the data). It would be good to have another table 1 with another seed to be sure of the results.

Clarifications

  • Why have you chosen TCGA-NSCLC and TCGA-RCC, where the results are already close to “plateau”?
  • minor comment: please remove “seed 2026”, the exact seed number being not useful if you don’t provide the corresponding github code
4 comments
AuthorsOpenReview ↗

Response to Reviewer 43z8 (1/2)

We sincerely thank the reviewer for the time and effort in evaluating our work. Our point-by-point responses are as follows.

W1 & Q2: Needs better significativity testing


Have you tested Table 1 with another seed to be sure that the results hold? ... It would be good to have another table 1 with another seed to be sure of the results.

We additionally repeated five-fold cross-validation on the original three datasets using seeds 2025 and 2027. As shown in Table 10, TiltMIL achieves the best results in 15 of the 18 evaluations and ranks among the top three in 16. Notably, for the NSCLC–UNI AUC highlighted by the reviewer, TiltMIL achieves 96.94% under the original seed 2026, while the results across the three seeds are 97.77%, 96.94%, and 97.33%, yielding an average of 97.35%. Consequently, the gap to the strongest baseline narrows from 0.93 to 0.38 percentage points. Meanwhile, the three-seed average Accuracy and Macro-F1 under this setting remain the best among the compared methods. These results indicate that some performance differences observed under the original single-seed setting are influenced by randomness, while further supporting TiltMIL's overall competitiveness across multiple datasets and feature settings. We further assessed performance variability using patient-level bootstrap and seed-level 95% confidence intervals. On PANDA–UNI, the seed-level intervals for Accuracy (+4.38 pp, [3.60, 5.16]), AUC (+1.68 pp, [1.54, 1.82]), and Macro-F1 (+5.03 pp, [3.95, 6.11]) all exclude zero, while those on TCGA-NSCLC and RCC generally include zero, indicating greater uncertainty in the smaller performance differences.

W2 & Q1: On the Independent Contribution of EPC


We are left wondering if EPC alone could improve in a similar way the results of classical benchmarks.

Have you added EPC on classical beanchmarks, to be sure that EPC alone is not enough to improve the results?

We thank the reviewer for the suggestion. To evaluate the independent contribution of EPC, we further incorporated EPC into ABMIL, CLAM-SB, TransMIL, and PTCMIL, conducting controlled experiments while preserving the original architectures of these baselines. The results show that the effects of EPC are method- and task-dependent. On PANDA, EPC improves the Accuracy of ABMIL and TransMIL by 1.53 and 2.47 percentage points, respectively, but does not improve the Accuracy of CLAM-SB. On TCGA-NSCLC, the performance changes are generally modest, with slight decreases in some metrics. These results indicate that EPC can improve the predictive performance of certain established MIL methods, but its contribution alone is insufficient to explain TiltMIL's overall performance gains on PANDA. We will include these controlled results in the revised manuscript to further clarify the practical contribution of EPC regularization.

MethodPANDA AccPANDA AUCPANDA F1-scoreNSCLC AccNSCLC AUCNSCLC F1-score
ABMIL74.62±0.5093.77±0.2269.96±1.1292.52±2.0597.59±1.0692.50±2.04
ABMIL+EPC76.15±0.3594.73±0.9171.46±0.5293.11±2.8297.49±1.5293.09±2.82
Δ+1.53+0.96+1.50+0.59-0.11+0.59
CLAM-SB75.69±0.7694.31±0.3071.53±1.4892.13±1.9797.03±1.9692.13±1.97
CLAM-SB+EPC75.57±0.3195.22±1.2171.46±0.9892.53±2.5797.40±1.6392.52±2.57
Δ-0.12+0.91-0.07+0.39+0.37+0.39
TransMIL66.05±2.0591.08±0.7559.98±1.8092.32±2.8497.32±1.2692.29±2.84
TransMIL+EPC68.52±1.5492.30±0.5562.37±1.7292.28±1.2997.20±2.2092.24±1.31
Δ+2.47+1.22+2.39-0.04-0.12-0.06
PTCMIL71.82±1.0592.97±0.3665.43±1.7693.25±2.3297.55±1.8793.25±2.32
PTCMIL+EPC73.14±1.2192.84±0.6666.79±1.3093.36±1.8897.46±1.1793.28±1.93
Δ+1.32-0.13+1.36+0.11-0.09+0.03
TiltMIL (Ours)80.73±0.4696.04±0.1777.35±0.6593.38±1.9796.94±1.8793.36±1.97
AuthorsOpenReview ↗

Response to Reviewer 43z8 (2/2)

Clarifications


C1:Why have you chosen TCGA-NSCLC and TCGA-RCC, where the results are already close to “plateau”?

We selected TCGA-NSCLC and TCGA-RCC primarily to further evaluate the applicability of TiltMIL beyond the six-class prostate cancer grading task on PANDA, across different organ origins (lung cancer subtypes in NSCLC and renal cancer subtypes in RCC) and classification settings (binary classification for NSCLC and three-class classification for RCC).

To further evaluate the applicability of TiltMIL across different cancer types and prediction tasks, we conducted additional patient survival prediction experiments on three TCGA cohorts, STAD (stomach adenocarcinoma), BLCA (bladder urothelial carcinoma), and BRCA (breast invasive carcinoma), and one CPTAC cohort, PDA (pancreatic ductal adenocarcinoma), using the concordance index (C-index) for evaluation. As shown in Table 11, TiltMIL achieves the highest mean C-index on three of the four cohorts: 0.637 on STAD, 0.674 on BRCA, and 0.627 on PDA. On BLCA, TiltMIL achieves 0.631, close to the best result of 0.636. These experiments provide further empirical support for TiltMIL's applicability across cancer types and prediction tasks.

minor comment: please remove “seed 2026”, the exact seed number being not useful if you don’t provide the corresponding github code

We have made the corresponding revisions in the main text.

Reviewer_43z8OpenReview ↗

Contribution of EPC in the Supplementary

I thank the authors for their rebuttal. COuld you please add the results about the contribution of EPC in the Supplementary? I believe that these results are useful. I have no more questions. I keep my positive rating.

Thank you for the helpful suggestion. We have added the results on the contribution of EPC in Appendix H. We appreciate your constructive feedback.

Reviewer_61hqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TiltMIL, a task-adaptive finite statistical representation framework for multiple instance learning Specifically, it first represents each bag as an empirical probability measure in a learned feature space and constructs a functional distributional representation using the Cumulant-Generating Function (CGF). By learning multiple exponential tilting locations, TiltMIL adaptively extracts finite statistical information, including CGF values, first-order cumulants, and second-order cumulants, which are subsequently integrated through Hierarchical Statistical Fusion. Furthermore, the authors introduce Empirical Perturbation Consistency (EPC) to improve robustness against instance-level perturbations and incomplete bag observations. Experiments are conducted on three types of synthetic MIL tasks and three real-world pathology datasets. The results demonstrate that TiltMIL achieves competitive performance across multiple tasks.

Critical strengths
  1. The paper addresses a well-motivated research problem by recognizing that different multiple instance learning tasks require different types of information about bag composition. Through three controlled MIL tasks, i.e., Presence, Co-occurrence, and Competitive, the authors investigate how instance presence, co-occurrence of different evidence types, and relative composition proportions affect bag-level predictions. These experiments reveal the complex responses of conventional normalized attention pooling to changes in instance composition.
  2. The paper conducts experiments on three pathology datasets using two feature extractors, ResNet-50 and UNI, and achieves competitive results across multiple experimental settings.
Critical weaknesses
  1. The paper achieves task-adaptive finite statistical representations by learning different exponential tilting locations. However, this adaptive mechanism primarily relies on learning a set of trainable tilting parameters through bag-level supervision, which is conceptually similar to existing learnable query-based feature aggregation mechanisms. Although the authors provide a clear statistical interpretation of these parameters, they do not sufficiently explain what unique advantages this parameterization offers in terms of the underlying learning mechanism.
  2. Although TiltMIL achieves relatively substantial improvements on the PANDA dataset, its advantages on TCGA-NSCLC and TCGA-RCC are comparatively limited. In particular, under the UNI feature setting, the Accuracy and Macro-F1 on TCGA-NSCLC improve by only approximately 0.04 and 0.02 percentage points, respectively, over the strongest baseline, while the Accuracy improvement on TCGA-RCC is only about 0.40 percentage points. These results suggest that the effectiveness of the proposed method may be highly dependent on the specific task and its bag composition characteristics.
Questions for authors
  1. The paper motivates its method by leveraging the property that the complete Cumulant-Generating Function uniquely characterizes an empirical distribution. However, the actual model retains only a finite number of learned CGF evaluations and low-order cumulants. Under what conditions can this finite representation distinguish different empirical bag distributions, and how do the number and locations of the exponential tilting points affect representational sufficiency or approximation error?
  2. TiltMIL extracts different bag-level statistical information by learning multiple exponential tilting locations, but the paper does not sufficiently investigate whether these locations actually capture complementary perspectives of the underlying bag distributions. Could different tilting vectors converge to similar directions, resulting in highly redundant statistical readouts?
  3. The paper achieves substantial improvements on the PANDA dataset, but the gains on TCGA-NSCLC and TCGA-RCC are relatively modest. Provide a more detailed class-wise analysis of the classification results to explain the differences in performance improvements across these datasets.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 61hq (1/3)

Thank you for your helpful comments and suggestions. Here are our point-by-point responses.

W1: On the Advantages of TiltMIL's Parameterization

Although the authors provide a clear statistical interpretation of these parameters, they do not sufficiently explain what unique advantages this parameterization offers in terms of the underlying learning mechanism.

We thank the reviewer for the suggestion. We believe that the main advantage of TiltMIL's parameterization lies in combining finite readouts of different statistical types with task-adaptive observation locations, allowing the model to retain different forms of bag composition information within a finite representation. Specifically, TiltMIL constructs different types of statistical readouts from the CGF and its induced tilted distributions, including CGF function values, first-order statistics, and statistics retained up to second order. Among these, the first-order tilted mean is indeed related to learnable query-based aggregation in its computational form. Compared with multi-query aggregation methods that typically employ weighted feature means, TiltMIL explicitly organizes different statistical readouts and learns their observation locations separately, allowing task supervision to adapt the corresponding observation views for different types of statistical information.

On the role of heterogeneous statistical readouts. Table 2 has already shown that removing either the CGF function-value readout or variance affects predictive performance. To further validate this, we conducted additional targeted experiments. Specifically, we used PMA as a learnable query-based aggregation baseline and included a multi-head attention (MHA) variant of ACMIL as a reference. We examined the effects of query capacity, weighted variance extensions, and EPC regularization. The results show that introducing weighted variance and EPC can improve the predictive performance of some baselines, but these extensions do not fully reproduce TiltMIL's performance on PANDA. Specifically, PMA+Var+EPC achieves a Macro-F1 of 73.13%, whereas TiltMIL reaches 77.35%. On TCGA-NSCLC, the differences among methods are relatively limited, while the full model remains competitive.

MethodPANDA AccPANDA AUCPANDA F1-scoreNSCLC AccNSCLC AUCNSCLC F1-score
PMA [1]75.01±1.8693.88±0.3870.17±2.0090.31±1.9696.12±1.9990.26±1.95
PMA+Var76.10±1.8293.98±0.5571.44±2.3991.55±2.4796.34±2.6591.51±2.47
PMA+CM74.88±1.5993.95±0.3870.60±1.6990.88±2.5196.37±2.4990.88±2.51
PMA+Var+EPC76.31±0.6194.00±0.1873.13±0.7290.84±2.7496.35±2.1290.80±2.76
ACMIL-MHA [2]72.53±2.1192.90±0.6067.75±3.5591.37±3.7996.93±2.4391.35±3.80
ACMIL-MHA+Var73.68±2.0993.19±0.5969.04±2.6192.10±3.2597.03±1.8492.07±3.31
ACMIL-MHA+CM73.07±2.3293.06±0.5468.43±2.9191.64±3.1297.02±1.9191.62±3.14
ACMIL-MHA+Var+EPC74.43±1.2093.88±0.4471.12±1.5792.71±2.4097.17±1.5992.69±2.45
TiltMIL80.73±0.4696.04±0.1777.35±0.6593.38±1.9796.94±1.8793.36±1.97

CM denotes capacity matching. For PMA, we match both the number of trainable parameters and learnable query heads to TiltMIL. For ACMIL-MHA, due to its distinct architecture, we match only the number of trainable parameters.

On the roles of statistical readouts and observation-location learning. We further conducted a two-factor ablation study within the same TiltMIL architecture, independently controlling the statistical readout types and whether the observation-location parameters are learnable. On PANDA–UNI, when observation locations are fixed, introducing heterogeneous statistical readouts improves Accuracy from 74.28% to 77.14%. With learnable observation locations, Accuracy increases from 77.47% to 80.73%. Similar trends in Accuracy are observed on TCGA-NSCLC–UNI. These results support the individual contributions of retaining heterogeneous statistical information and learning task-adaptive observation locations.

PANDA–UNI: Statistical Readouts and Observation-Location Learning

Observation Locations1st-order Acc1st-order AUCHeterogeneous AccHeterogeneous AUC
Random Fixed t74.28 ± 1.2194.79 ± 0.4377.14 ± 0.7995.57 ± 0.25
Learnable t77.47 ± 0.8895.44 ± 0.2980.73 ± 0.4696.04 ± 0.17

TCGA-NSCLC–UNI: Statistical Readouts and Observation-Location Learning

Observation Locations1st-order Acc1st-order AUCHeterogeneous AccHeterogeneous AUC
Random Fixed t90.61 ± 2.6296.15 ± 1.8692.81 ± 1.8296.95 ± 1.78
Learnable t92.41 ± 2.1197.07 ± 1.8593.38 ± 1.9796.94 ± 1.87

[1] Set Transformer: A Framework for Attention-Based Permutation-Invariant Neural Networks

[2] Attention-Challenging Multiple Instance Learning for Whole Slide Image Classification

AuthorsOpenReview ↗

Response to Reviewer 61hq (2/3)

W2 & Q3: On performance differences across datasets


Although TiltMIL achieves relatively substantial improvements on the PANDA dataset ... These results suggest that the effectiveness of the proposed method may be highly dependent on the specific task and its bag composition characteristics.

First, the three datasets involve different classification settings: PANDA consists of six ordered ISUP grades, while TCGA-NSCLC and TCGA-RCC are binary and three-class classification tasks, respectively. We acknowledge that, although TiltMIL achieves relatively modest improvements on some TCGA–UNI metrics, its overall performance is not limited to PANDA. Specifically, across all settings, TiltMIL achieves the best results in 15 out of 18 evaluations and the second-best result in one. In comparison, CLAM-MB achieves 6 top-two results. These results indicate that, although the relative performance of different methods varies across task settings, TiltMIL maintains strong overall competitiveness across tasks.

Provide a more detailed class-wise analysis of the classification results to explain the differences in performance improvements across these datasets.

Following the reviewer's suggestion, we further examine the classification results across the six ISUP grades in PANDA, the two lung cancer subtypes in TCGA-NSCLC, and the three renal cancer subtypes in TCGA-RCC. We use confusion matrices to analyze the sources of performance differences across datasets. As shown in Fig. 9, on PANDA, TiltMIL improves classification performance across multiple ISUP grades under both feature settings, particularly for intermediate grades (ISUP 1–4), reducing confusion between certain adjacent grades. On TCGA-NSCLC, TiltMIL consistently improves the recall of LUAD under both feature settings, while the recall of LUSC decreases slightly. These results suggest that TiltMIL's gains on PANDA arise from improvements across multiple classes, whereas its gains on TCGA-NSCLC and TCGA-RCC are more localized to specific classes and accompanied by trade-offs between classes, helping explain the differences in overall performance improvements across datasets.

Q1: On the Sufficiency of Finite CGF Representations and the Selection of Observation Locations

We thank the reviewer for highlighting the relationship between the complete CGF and finite statistical readouts.

What conditions can this finite representation distinguish different empirical bag distributions?

We provide a conditional identifiability analysis in Appendix C. For smoothly parameterized distribution families, a full-column-rank Jacobian of the finite readout mapping guarantees local distinguishability. For distributions with a known common finite support, the probability masses can be uniquely recovered when the exponential observation matrix, augmented with the probability normalization constraint, has full column rank. These results indicate that the distinguishability of finite readouts depends on the distribution family and the configuration of observation locations. For bag-level classification, TiltMIL further selects finite statistical observations through task supervision to retain composition information relevant to downstream decisions, without requiring complete reconstruction of arbitrary empirical distributions.

how do the number and locations of the exponential tilting points affect representational sufficiency or approximation error?

Figure 5 analyzes the effects of different observation budgets on predictive performance. We further evaluate the recoverability of bag composition histograms on a controlled dataset with known composition proportions, using a linear probe on frozen representations. As we can see, increasing the number of observation locations improves composition recovery, while learned locations achieve higher information-retention efficiency. For example, only 12 Co-occurrence-supervised locations achieve an R2R^2 of 0.169, comparable to the result of 0.170 obtained with 48 random locations.

Location SourceRHistogram R2R^2 ↑MAE ↓JS ↓
Random30.012±0.0030.0404±0.00010.0487±0.0003
Random120.035±0.0130.0395±0.00030.0468±0.0008
Random480.170±0.0140.0357±0.00040.0383±0.0010
Co-occurrence learned30.069±0.0190.0380±0.00030.0426±0.0007
Co-occurrence learned120.169±0.0220.0343±0.00040.0333±0.0012
Co-occurrence learned480.290±0.0080.0307±0.00020.0258±0.0008
Composition learned30.059±0.0220.0386±0.00080.0440±0.0019
Composition learned120.180±0.0150.0345±0.00020.0346±0.0006
Composition learned480.406±0.0130.0288±0.00030.0244±0.0006
AuthorsOpenReview ↗

Response to Reviewer 61hq (3/3)

(Continued from the previous comment)

We further explore that the composition recovery R2R^2 for Co-occurrence-supervised locations increases from 0.238 when using only CGF function values to 0.266 after incorporating first-order statistics, and further to 0.290 after including second-order statistics. Meanwhile, Composition-supervised locations achieve higher composition recoverability, whereas the original Table 3 shows that Co-occurrence-supervised locations slightly outperform them across all three classification tasks. These results indicate that the recoverability of overall composition information does not necessarily align with its discriminative utility for a specific task, further supporting TiltMIL's motivation to learn task-relevant observation locations under a finite statistical budget.

Location SourceReadoutHistogram R2R^2 ↑MAE ↓JS ↓
RandomK0.098±0.0150.0377±0.00050.0428±0.0013
RandomK+κ₁0.142±0.0140.0360±0.00040.0389±0.0009
RandomK+κ₁+κ₂0.170±0.0140.0357±0.00040.0383±0.0010
Co-occurrenceK0.238±0.0110.0321±0.00040.0285±0.0011
Co-occurrenceK+κ₁0.266±0.0090.0309±0.00030.0260±0.0007
Co-occurrenceK+κ₁+κ₂0.290±0.0080.0307±0.00020.0258±0.0008
CompositionK0.311±0.0180.0313±0.00050.0288±0.0011
CompositionK+κ₁0.390±0.0140.0290±0.00030.0248±0.0007
CompositionK+κ₁+κ₂0.406±0.0130.0288±0.00030.0244±0.0006

Q2: Are different exponential tilting locations complementary?


Could different tilting vectors converge to similar directions, resulting in highly redundant statistical readouts?

We thank the reviewer for the suggestion. We agree that learning multiple exponential tilting locations does not necessarily guarantee complementary statistical observations, and redundancy may exist among different locations. To investigate this, we conducted a redundancy analysis on the full TiltMIL model trained on the controlled Co-occurrence task. For the CGF function-value (ΦKΦ_K), first-order statistical (Φκ(1)Φ_κ⁽¹⁾), and second-order statistical (Φκ(2)Φ_κ⁽²⁾) readout branches, we examined the directional similarity of tilting locations, the Jensen–Shannon Divergence (JS) between the resulting tilted instance weights, and the CKA similarity of the actual statistical readouts.

As shown in Fig. 8, different locations within the ΦKΦ_K and Φκ(1)Φ_κ⁽¹⁾ branches produce substantially different instance-weight distributions, with median JS values of approximately 0.59 and 0.55, respectively. Their statistical readouts also generally exhibit low similarity, suggesting that these locations provide differentiated statistical observations. In contrast, the Φκ(2)Φ_κ⁽²⁾ branch shows a median JS of only approximately 0.01 and a readout CKA of approximately 0.95, indicating substantial redundancy within this branch. These results show that different statistical readout branches exhibit varying degrees of diversity in their observations: some retain more differentiated statistical information, while others exhibit overlapping statistical responses. Further details are provided in Appendix D.

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

This paper proposes TiltMIL, a multiple instance learning framework that represents each bag through a cumulant-generating function (CGF) and learns task-adaptive finite statistical readouts. Experiments on three whole-slide image classification benchmarks demonstrate competitive performance across different pathology tasks and feature encoders

Critical strengths

1.The paper provides an interesting and well-motivated perspective on MIL by introducing task-adaptive statistical representations rather than relying solely on conventional instance aggregation.

2.The proposed framework is reasonably well designed, and the combination of CGF-based readouts and learnable tilt locations offers a flexible way to capture bag-level composition.

Critical weaknesses

1.Insufficient computational efficiency analysis. The proposed framework introduces additional CGF-based readouts, learnable tilt locations, and perturbation-consistency regularization. However, the paper lacks a quantitative comparison of the additional training and inference costs. The authors should report training time, inference latency overhead compared with representative MIL baselines under identical experimental settings. Such results are important for evaluating the practicality of the proposed method on large-scale whole-slide images.

2.Limited experimental generalizability. Although the proposed method demonstrates promising results on three pathology datasets, further evaluation on more diverse datasets would strengthen the evidence for its effectiveness and generalizability. I would be very willing to increase my score if the authors could provide additional results demonstrating consistent improvements across more datasets.

Questions for authors

1.Insufficient computational efficiency analysis.

2.Limited experimental generalizability.

I would be very willing to increase my score if the authors could provide additional results demonstrating consistent improvements across more datasets.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MkPd

We thank the reviewer for the careful reading of our manuscript. Our point-by-point replies are provided below.

W1 & Q1: Insufficient computational efficiency analysis


The authors should report training time, inference latency overhead compared with baselines.

We thank the reviewer for the suggestion. We compared TiltMIL with ten representative MIL methods under identical settings (FP32, 4,096 instances per bag, 1,024-dimensional features, and binary classification), reporting parameters, GFLOPs, and per-bag training and inference times.

As shown below, TiltMIL requires only 0.496M parameters and 3.478 GFLOPs. Its inference and training times are 2.99 and 28.73 ms/bag, respectively. Although computing CGF values and weighted statistics introduces runtime overhead compared with lightweight pooling methods, TiltMIL remains substantially faster than 2DMamba (17.26/114.66 ms) and PTCMIL (26.65/164.67 ms) in inference/training. These results demonstrate competitive computational efficiency with a relatively small model size.

MethodParams(M)GFLOPsInference(ms/Bag)Training(ms/Bag)
ABMIL0.5924.8370.623.44
CLAM-SB0.7916.4490.907.39
CLAM-MB0.7916.4551.288.98
DSMIL0.1541.2441.256.76
TransMIL2.67254.46410.0937.91
DTFD-MIL0.7896.4575.8725.51
ACMIL6.31827.9694.4942.71
WiKG2.10331.1916.8618.92
2DMamba7.37562.18817.26114.66
PTCMIL1.57032.54326.65164.67
TiltMIL(Ours)0.4963.4782.9928.73

W2 & Q2: Limited experimental generalizability


Further evaluation on more diverse datasets would strengthen the evidence for its effectiveness and generalizability.

We thank the reviewer for the suggestion. To further evaluate the applicability of TiltMIL across different cancer types and prediction tasks, we conducted additional patient survival prediction experiments on three TCGA cohorts, STAD (stomach adenocarcinoma), BLCA (bladder urothelial carcinoma), and BRCA (breast invasive carcinoma), and one CPTAC cohort, PDA (pancreatic ductal adenocarcinoma), using the concordance index (C-index) for evaluation. TiltMIL achieves the highest mean C-index on three of the four cohorts: 0.637 on STAD, 0.674 on BRCA, and 0.627 on PDA. On BLCA, TiltMIL achieves 0.631, close to the best result of 0.636. These experiments extend our original classification evaluations to survival prediction tasks across different cancer types, providing further empirical support for TiltMIL's applicability across cancer types and prediction tasks.

MethodSTADBLCABRCAPDA
ABMIL0.604 ± 0.0700.629 ± 0.0470.652 ± 0.0550.617 ± 0.111
CLAM-SB0.624 ± 0.0840.618 ± 0.0350.656 ± 0.0540.596 ± 0.110
CLAM-MB0.626 ± 0.0810.636 ± 0.0320.650 ± 0.0600.569 ± 0.099
DSMIL0.626 ± 0.0720.634 ± 0.0320.618 ± 0.0820.559 ± 0.111
TransMIL0.605 ± 0.0850.602 ± 0.0540.649 ± 0.0760.584 ± 0.126
DTFD0.621 ± 0.0740.624 ± 0.0340.665 ± 0.0420.586 ± 0.112
ACMIL0.626 ± 0.0740.632 ± 0.0450.658 ± 0.0490.606 ± 0.082
WiKG0.603 ± 0.0640.595 ± 0.0410.670 ± 0.0620.573 ± 0.053
2DMamba0.632 ± 0.0760.626 ± 0.0460.664 ± 0.0810.600 ± 0.071
PTCMIL0.612 ± 0.0760.614 ± 0.0530.652 ± 0.0620.567 ± 0.099
TiltMIL0.637 ± 0.0650.631 ± 0.0140.674 ± 0.0810.627 ± 0.073

We also conducted additional experiments on SIVAL [1], a non-WSI MIL dataset with 25 classes. TiltMIL achieves the highest performance across all three metrics.

MethodAccAUCMacro-F1
ABMIL59.20±4.6091.98±2.2659.26±4.50
CLAM-SB61.93±3.9593.41±0.9761.75±3.85
CLAM-MB68.33±0.4294.82±1.1168.24±0.53
DSMIL53.47±4.4490.67±1.4754.00±4.11
TransMIL56.87±4.8290.10±2.4255.82±4.80
DTFD60.67±3.5692.36±1.0560.47±3.43
ACMIL75.47±3.0697.36±0.3575.99±3.02
WiKG69.27±1.6395.04±0.6968.67±1.84
2DMambaN/AN/AN/A
PTCMIL72.13±4.5695.71±1.3370.69±4.63
TiltMIL (Ours)76.47±2.7297.60±0.4876.18±2.79

N/A: 2DMamba requires spatial coordinates, which are unavailable in SIVAL’s region-feature representation.

[1] Localized content based image retrieval.

Thank you very much for your efforts, and thanks to the authors for providing additional experiments. I have decided to maintain my score.

We sincerely thank the reviewer for taking the time to consider our additional experiments and responses. We greatly appreciate your constructive feedback and thoughtful consideration throughout the discussion.

Reviewer_Q6YBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes TiltMIL, a multiple instance learning framework that models bag composition through a cumulant-generating function (CGF). Instead of directly aggregating instance features into a fixed bag representation, the method learns task-dependent tilt locations from which it extracts CGF values and statistical quantities, including first-order means and second-order variances. These statistics are combined through a hierarchical fusion module, with an additional perturbation consistency objective to improve robustness. Experiments on PANDA, TCGA-NSCLC, and TCGA-RCC demonstrate competitive performance using both ResNet-50 and UNI features.

Critical strengths

(1) The paper offers an interesting perspective on MIL aggregation. Different pathology tasks may depend on the presence, proportion, or heterogeneity of specific tissue patterns, and learning task-adaptive distributional statistics is a reasonable way to address these differences. The methodological motivation is generally clear.

(2) The experimental results are competitive. The authors compare TiltMIL with ten MIL baselines across three pathology datasets and two feature encoders. The improvements on PANDA are particularly encouraging. The inclusion of controlled composition experiments, statistical readout ablations, and bag perturbation analysis also strengthens the evaluation.

Critical weaknesses

(1) The distinction from existing distribution-aware MIL methods needs to be better established. Although DeepDisMISL and Distribution-based MIL Pooling are discussed, they are not included in the main experimental comparisons. More importantly, learnable exponential tilting is closely related to multi-query attention, while first- and second-order statistics can also be obtained through other weighted aggregation mechanisms. It is therefore unclear whether the performance gains come specifically from the CGF formulation or simply from using multiple learnable queries and additional statistical features.

(2) Some experiments do not fully support the central methodological claims. In Table 2, adding first-order statistics alone provides limited improvement, while second-order statistics help substantially on PANDA but are less consistent on TCGA-NSCLC. The paper emphasizes the expressive properties of the full CGF, but the actual model only evaluates a finite number of tilt locations and a restricted set of statistics. The connection between these theoretical properties and the implemented model could be made more convincing.

(3) The improvements are not uniformly strong across all settings. With UNI features, TiltMIL does not achieve the best AUC on TCGA-NSCLC or TCGA-RCC, and some gains are relatively small compared with the reported cross-validation variability. Additional validation across independent cohorts or random seeds would strengthen the generalization claims.

Questions for authors

The most important missing experiment is a controlled comparison against multi-query attention and existing distribution-aware MIL approaches, particularly DeepDisMISL, with matched model capacity, query numbers, and regularization.

It would also be useful to demonstrate that the learned second-order statistics capture meaningful aspects of tissue composition rather than merely improving classification accuracy. For example, controlled experiments with known tissue proportions or lesion distributions could clarify what information each statistical branch actually learns.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Q6YB (1/3)

We thank the reviewer for the effort in reviewing our work. Our point-by-point responses are as follows.

W1 & Q1: On Comparisons with Distribution-Aware and Multi-Query Methods


The distinction from existing distribution-aware MIL methods needs to be better established. ... It is therefore unclear whether the performance gains come specifically from the CGF formulation or simply from using multiple learnable queries and additional statistical features.

We thank the reviewer for highlighting the connections between TiltMIL and existing distribution-aware and multi-query aggregation methods. Exponential tilting indeed shares a similar instance-reweighting operation with multi-query attention. The distinguishing design of TiltMIL is to treat learnable tilting locations as finite observation points of the CGF and jointly extract CGF function values and statistics of the induced tilted distributions, thereby constructing a task-adaptive representation with explicit statistical structure.

The most important missing experiment is a controlled comparison against multi-query attention and existing distribution-aware MIL approaches

To further examine whether the improvements can be explained by query capacity, additional statistical features, or regularization, we conducted controlled experiments using Pooling by Multihead Attention (PMA [1]), considering capacity matching (CM), weighted variance extensions (Var), and EPC regularization. Adding variance and EPC improves PMA's Macro-F1 from 70.17% to 73.13%, whereas TiltMIL achieves 77.35%. These results indicate that augmenting query-based aggregation with weighted variance and regularization does not fully reproduce TiltMIL's performance in this setting.

MethodPANDA AccPANDA AUCPANDA F1-scoreNSCLC AccNSCLC AUCNSCLC F1-score
PMA [1]75.01±1.8693.88±0.3870.17±2.0090.31±1.9696.12±1.9990.26±1.95
PMA+CM74.88±1.5993.95±0.3870.60±1.6990.88±2.5196.37±2.4990.88±2.51
PMA+Var76.10±1.8293.98±0.5571.44±2.3991.55±2.4796.34±2.6591.51±2.47
PMA+Var+EPC76.31±0.6194.00±0.1873.13±0.7290.84±2.7496.35±2.1290.80±2.76
DeepDisMISL [2]58.84±1.1185.75±0.6452.12±2.0787.50±3.5095.44±1.6187.43±3.60
DeepDisMISL+EPC58.42±1.5385.77±0.6551.80±1.7787.97±2.1194.76±1.4987.94±2.09
TiltMIL80.73±0.4696.04±0.1777.35±0.6593.38±1.9796.94±1.8793.36±1.97

We additionally evaluated TiltMIL against DeepDisMISL on survival prediction tasks, as DeepDisMISL was originally developed for prognostic modeling. TiltMIL achieves higher mean C-indices on TCGA-STAD and TCGA-BRCA, reaching 0.637 and 0.674, respectively. Furthermore, the original Table 2 shows that removing the CGF function-value readout decreases Accuracy by 2.19 and 4.07 percentage points on PANDA and TCGA-NSCLC, respectively. Together, these findings provide empirical support for the value of TiltMIL's finite statistical readout design, beyond simply increasing query capacity or adding regularization.

MethodSTAD(C-index↑)BRCA(C-index↑)
DeepDisMISL0.630±0.0640.646±0.020
DeepDisMISL+EPC0.626±0.0460.650±0.015
TiltMIL0.637±0.0650.674±0.081

[1] Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., and Teh, Y. W. Set Transformer: A Framework for Attention-Based Permutation-Invariant Neural Networks

[2] Li, X., Jonnagaddala, J., Cen, M., Zhang, H., and Xu, S. Colorectal Cancer Survival Prediction Using Deep Distribution Based Multiple-Instance Learning

W2: On the Contribution of Finite and Heterogeneous Statistical Readouts


Second-order statistics help substantially on PANDA but are less consistent on TCGA-NSCLC.

We agree that the contributions of different statistical readouts are task-dependent, which is consistent with our motivation that different bag-level decisions may rely on different aspects of tissue composition. Specifically, PANDA involves prostate cancer grading, where ISUP grades are related to the presence and relative proportions of different Gleason patterns, whereas TCGA-NSCLC focuses on distinguishing histological subtypes (LUAD and LUSC), which relies more directly on characteristic tissue morphology. These differences in task requirements may explain why second-order statistics provide more substantial gains on PANDA but relatively limited improvements on TCGA-NSCLC.

AuthorsOpenReview ↗

Response to Reviewer Q6YB (2/3)

The paper emphasizes the expressive properties of the full CGF, but the actual model only evaluates a finite number of tilt locations and a restricted set of statistics. The connection between these theoretical properties and the implemented model could be made more convincing.

We thank the reviewer for this suggestion. To further clarify the connection between the theoretical properties of the complete CGF and the implemented finite statistical readouts, we provide a conditional distinguishability analysis in Appendix C, establishing sufficient conditions for local distinguishability in smoothly parameterized distribution families and exact recovery of probability masses on a known common finite support. We further conducted controlled bag-composition recovery experiments to examine the effects of both observation locations and statistical readout types. As shown in Table 12, 12 Co-occurrence-supervised locations achieve a histogram recovery R² of 0.169, comparable to 0.170 obtained with 48 random locations. Furthermore, at R=48, incorporating first- and second-order statistics progressively improves R² from 0.238 to 0.266 and 0.290, respectively (Table 13). These results provide additional theoretical and empirical evidence for the representational capabilities of finite CGF-based readouts and the importance of observation-location selection.

W3 The improvements are not uniformly strong across all settings.


Additional validation across independent cohorts or random seeds would strengthen the generalization claims.

To further evaluate the applicability of TiltMIL across cancer types and tasks, we conducted survival prediction experiments on three additional TCGA cohorts, STAD, BLCA, and BRCA and one CPTAC cohort, PDA, using the concordance index (C-index). TiltMIL achieves the highest mean C-index on three of the four cohorts: 0.637 on STAD, 0.674 on BRCA, and 0.627 on PDA. On BLCA, TiltMIL achieves 0.631, close to the best result of 0.636. These experiments extend the evaluation beyond the original classification benchmarks and provide additional empirical evidence of applicability across cancer cohorts and prediction tasks.

MethodSTADBLCABRCAPDA
ABMIL0.604 ± 0.0700.629 ± 0.0470.652 ± 0.0550.617 ± 0.111
CLAM-SB0.624 ± 0.0840.618 ± 0.0350.656 ± 0.0540.596 ± 0.110
CLAM-MB0.626 ± 0.0810.636 ± 0.0320.650 ± 0.0600.569 ± 0.099
DSMIL0.626 ± 0.0720.634 ± 0.0320.618 ± 0.0820.559 ± 0.111
TransMIL0.605 ± 0.0850.602 ± 0.0540.649 ± 0.0760.584 ± 0.126
DTFD0.621 ± 0.0740.624 ± 0.0340.665 ± 0.0420.586 ± 0.112
ACMIL0.626 ± 0.0740.632 ± 0.0450.658 ± 0.0490.606 ± 0.082
WiKG0.603 ± 0.0640.595 ± 0.0410.670 ± 0.0620.573 ± 0.053
2DMamba0.632 ± 0.0760.626 ± 0.0460.664 ± 0.0810.600 ± 0.071
PTCMIL0.612 ± 0.0760.614 ± 0.0530.652 ± 0.0620.567 ± 0.099
TiltMIL0.637 ± 0.0650.631 ± 0.0140.674 ± 0.0810.627 ± 0.073

We additionally repeated five-fold cross-validation on the original three datasets using seeds 2025 and 2027. To maintain consistency with other follow-up analyses, we retain the original results in Table 1 and report the multiple-seeds results in Table 10. TiltMIL achieves the best performance in 15 of the 18 evaluations and ranks among the top three in 16. For NSCLC–UNI AUC, the results across the three seeds are 97.77%, 96.94%, and 97.33%, yielding an average of 97.35%, compared with 96.94% under the original seed alone. The additional evaluations further support TiltMIL's overall competitiveness across datasets and feature settings.

AuthorsOpenReview ↗

Response to Reviewer Q6YB (3/3)

Q2: On the Tissue-Composition Information Captured by Second-Order Statistics.


It would also be useful to demonstrate that the learned second-order statistics capture meaningful aspects of tissue composition rather than merely improving classification accuracy. For example, controlled experiments with known tissue proportions or lesion distributions could clarify what information each statistical branch actually learns.

We thank the reviewer for suggesting a more direct analysis of the information captured by second-order statistical readouts. To investigate their responses to meaningful tissue-composition changes, we conducted an additional controlled experiment using Gleason region annotations from the Radboud subset of PANDA.

Specifically, we selected 47 whole-slide images containing high-confidence Gleason pattern 3 (G3), Gleason pattern 4 (G4), and non-cancerous tissue instances, and froze the trained PANDA–UNI TiltMIL model. Using fixed tissue-instance pools and a fixed total probability mass assigned to cancerous patches, we systematically varied the relative proportions of G3 and G4 and examined the responses of CGF function values, first-order tilted means, and second-order projected variances.

As shown in Fig. 10, CGF function values exhibit nonlinear responses to tissue mixtures (Panel E), first-order tilted means change with the relative G3/G4 composition (Panel F), and second-order projected variances capture additional feature dispersion arising from mixed tissue patterns (Panel G). When the total cancerous tissue mass is fixed at 60% and equally divided between G3 and G4, the mean variance of the mixed composition exceeds the average variance of the two single-pattern compositions by 0.2827.

To further identify the source of this increase, we decomposed the second-order variance into within-tissue and between-tissue components. As shown in Panel H, differences between G3 and G4 feature distributions account for approximately 97.6% of the average variance increase, while the remaining components contribute relatively little. These results demonstrate that the learned second-order readouts respond to controlled changes in tissue composition, rather than merely providing an unexplained improvement in classification accuracy. We provide further details and analyses in Appendix E.

Comments on the paper

2 comments
Area_Chair_48gEOpenReview ↗

Reminder – Please engage with author rebuttal

Dear Reviewers,

We are now past the halfway point of the discussion phase. The authors have posted their rebuttal, and I encourage all reviewers to engage actively in the discussion.

Please take the time to: i) read the authors' responses carefully; ii) check whether your concerns and questions have been adequately addressed; and iii) provide the authors with further feedback where needed.

Your active participation at this stage is important for reaching a well-informed final decision.

Thank you for your time and effort.

Best regards,

Your AC

AuthorsOpenReview ↗

Response Summary

We are grateful to all reviewers for their careful evaluation of our work and their constructive feedback. We would like to briefly summarize the main clarifications and additional experiments provided in our rebuttal:

  1. Comparison with multi-query and distribution-aware MIL methods: to examine whether the performance gains of TiltMIL stem from its CGF-based formulation rather than increased query capacity or additional statistical features (Reviewers Q6YB, 61hq, oAAV), we compared TiltMIL with PMA and ACMIL-MHA under capacity-matched, weighted-variance, and EPC-augmented settings, and included DeepDisMISL as a distribution-aware baseline (Appendix H.2). We further conducted a two-factor ablation to isolate the contributions of heterogeneous statistical readouts and learnable observation locations (Appendix H.3).

  2. Theoretical properties and interpretability of finite CGF representations: to address concerns regarding the distinguishability of finite CGF representations, the redundancy of learned tilting locations, and the information captured by second-order statistics (Reviewers Q6YB, 61hq), we provided conditional distinguishability analysis (Appendix C), investigated the redundancy of learned tilting locations and statistical readouts (Appendix D), and conducted controlled tissue-composition experiments using Gleason region annotations to examine the role of second-order statistics (Appendix E). Furthermore, we conducted controlled bag-composition recovery experiments to quantify how observation-location selection and heterogeneous statistical readouts affect the retention of composition information in finite representations (Appendix G).

  3. Generalizability and statistical reliability: in response to concerns regarding generalization across datasets and the consistency of performance improvements (Reviewers Q6YB, MkPd, 61hq, 43z8, oAAV), we extended our evaluation to additional survival prediction cohorts from TCGA and CPTAC, as well as the non-WSI SIVAL dataset (Appendix F, I). We also conducted additional multi-seed experiments, patient-level bootstrap analyses, seed-level confidence interval estimation, and fold-level and class-wise comparisons.

  4. Computational efficiency and the independent contribution of EPC: to address concerns regarding the computational overhead of TiltMIL and whether EPC alone could explain the performance improvements (Reviewers MkPd, 43z8), we compared model parameters, GFLOPs, training time, and inference latency against ten representative MIL methods. We also evaluated the independent contribution of EPC by incorporating it into several established MIL baselines. (Appendix H.1, H.4).

We have provided detailed discussions of other specific questions raised by the reviewers in our individual responses. We believe these revisions and clarifications address the key concerns raised by the reviewers and further strengthen the paper.

We sincerely thank the Area Chair and all reviewers for their time and efforts in reviewing our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.