acceptodds
Under review as a conference paper at ICLR 2027

EnergyMIL: Multiple Instance Learning as Strongly Convex Energy Minimization for Whole-Slide Image Classification

Abstract

Multiple instance learning (MIL) is the dominant framework for whole slide image (WSI) classification, where each slide becomes a bag of thousands of patch features supervised only by a slide label. Contextual MIL enriches these features through interactions among patches. Current designs specify the operations that compute context, while the target representation remains implicit. Making that target explicit gives contextual aggregation a principled endpoint. We design EnergyMIL to assign each WSI a unique contextual state defined over the entire slide. First, factorized anchor coupling connects tissue evidence across thousands of patches with storage linear in bag size. Then, a strongly convex contextual energy defines the state, and proximal refinement approaches it through a memory efficient learning path. Finally, complementary focal and global readouts produce the slide prediction. Across four datasets, EnergyMIL ranks first outright in eleven of the twelve dataset by metric comparisons and ties for first in the remaining one. Together, the formulation and results show how a well defined contextual state can provide a stable and scalable basis for integrating heterogeneous tissue evidence in WSI classification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MIL, WSI, Energy, Classification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3y6kOpenReview ↗
Rating
1(−2)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

EnergyMIL formulates WSI contextual modeling as a strongly convex energy minimization problem. It uses learned anchors for efficient global patch interaction, solves for a unique contextual state through iterative refinement, and combines focal and global evidence for slide classification.

Critical strengths
  1. Mathematical formulation is clear and internally consistent. The paper gives a well-defined inner optimization problem. Under its assumptions, the contextual state has a unique solution and proximal refinement converges linearly to that solution. The theory matches the actual model design rather than being added only as an explanation after the fact.

  2. Anchor-based global coupling is practical for WSI data. Direct patch-to-patch interaction can be expensive for thousands of patches. EnergyMIL routes communication through a small set of learned anchors, which allows global interaction while keeping memory roughly linear in the number of patches.

  3. Method is relatively lightweight. EnergyMIL has 0.535M parameters and 2.63G profiled FLOPs under the reported setup. This is much lower than several Transformer-based MIL methods while still allowing slide-wide interaction.

  4. The experimental comparison is reasonably broad within MIL. The paper compares against many recent MIL methods using the same CONCH features and the same five-fold splits. It also provides component ablations, parameter sensitivity, spatial visualizations, and paired slide-level statistical tests.

  5. The authors report uncertainty more carefully than many MIL papers. They do not rely only on average fold scores. They also perform paired bootstrap and permutation tests against MammothMIL and clearly state when statistical significance is limited.

Critical weaknesses
  1. Core mathematical idea is mainly transferred from implicit and equilibrium models. Defining a representation as a fixed point or optimization solution, guaranteeing uniqueness, and using iterative solvers are established ideas from DEQ, monotone operator networks, and energy-based implicit models. Therefore, the main novelty is the WSI-specific formulation and anchor coupling, not the general idea of an optimization-defined representation.

  2. This paper does not fully show that the equilibrium formulation itself is necessary. The strongest ablation effect comes from anchor coupling. In contrast, the gain from detached multi-step state refinement is much smaller. This means the experiments support the usefulness of global contextual interaction more strongly than they support the need for a strongly convex equilibrium state. A direct comparison with a matched finite-depth residual model using the same anchors, feature map, and readout would be important.

  3. Primary datasets are small. The four main datasets contain only 111, 192, 225, and 109 slides. The reported standard deviations mainly reflect variation across folds, not variation across independent training runs. This limits how strongly model stability can be concluded.

Questions for authors

Released code :1. The held-out fold is used for both model selection and final evaluation. The code evaluates the held-out fold after every epoch and selects the checkpoint with the highest held-out accuracy. This allows test-fold information to influence model selection and may lead to optimistic performance estimates. 2. Reported parameter count does not match the released implementation. The paper reports about 0.535M parameters, while the released default model contains about 0.470M parameters. This difference should be explained.

3 comments
AuthorsOpenReview ↗

We acknowledged the checkpoint selection limitation and will align the reported parameter count with the released code.

We thank the reviewer for the reassessment. We agree that selecting checkpoints by held out fold accuracy was a serious oversight in our experimental design. Our previous reply was inadequate, and we regret that. We accept that the submitted ranking lacks independent evaluation.

1. Checkpoint selection. All methods selected checkpoints by held out fold accuracy. This shared rule still allowed test information to influence model selection. We examined the two baselines with the highest submitted AUC on each dataset. Each cell shows submitted AUC followed by AUC at epoch 100, averaged over five folds.

DatasetEnergyMILHighest baselineSecond baseline
BML0.970 to 0.900MammothMIL 0.907 to 0.853CLAM-SB 0.892 to 0.863
HER20.892 to 0.885MammothMIL 0.876 to 0.837ILRA-MIL 0.860 to 0.824
AURORA0.915 to 0.899MammothMIL 0.913 to 0.871CLAM-SB 0.901 to 0.856
TCGA KICH0.634 to 0.596CLAM-SB 0.632 to 0.585DTFD-MIL 0.631 to 0.588

EnergyMIL retains higher epoch 100 AUC than these baselines, but this subset cannot establish an overall ranking. On TCGA KICH, MR-Block and RRT-MIL reach 0.617 and 0.607, above EnergyMIL's 0.596. ACC and macro F1 also change materially. Epoch 100 was not selected by independent validation. A corrected comparison requires validation within each outer training fold to select checkpoints for every method, followed by one evaluation on the outer test fold.

2. Parameter count. Direct enumeration with a 512 dimensional input and two class output identifies the discrepancy.

ItemParametersExplanation
Earlier implementation535,172Count used in the submission
Unused matrix65,536Registered with the skew component disabled
Unused attention scalar1Registered under the softmax readout
Released EnergyMIL469,635Parameters in the released model

The submitted 0.535M figure was incorrect. The correct size of the released model is 0.470M.

3. Mathematical contribution. We agree that equilibrium learning comes from prior implicit models. Our contribution is the WSI formulation and analysis.

Established principleEnergyMIL contribution
Optimization defined representationsA slide dependent contextual energy for variable size patch bags
Uniqueness and iterative inferenceA strongly convex target with a convergence guarantee under the stated conditions
Global contextual interactionFactorized learned anchor coupling across patches

These results specify the contextual state reached under fixed model parameters. They do not imply improved discrimination or generalization on their own.

4. Necessity of convergence. The one update control starts from zero and lacks cross patch interaction. We added fixed update controls with the same embedding, coupling, and readouts, plus a nearly parameter matched two layer residual anchor network.

Contextual modelBML ACCBML AUCBML F1TCGA KICH ACCTCGA KICH AUCTCGA KICH F1
One update0.9370.9580.9060.7980.5660.590
Five updates0.9380.9510.9070.8160.6180.654
Ten updates0.9280.9620.8880.8260.6180.665
Convergence based0.9110.9300.8490.8160.5830.638
Matched two layer anchors0.9020.9420.8400.7890.5770.560

On TCGA KICH, several updates improve all three metrics over one update. EnergyMIL exceeds the matched residual network in ACC and F1 on both datasets. Five or ten updates can match or exceed convergence based performance. The energy formulation supplies a unique contextual target, although fully reaching it is not necessary for the highest accuracy here. These controls share the original checkpoint rule and remain mechanism diagnostics.

5. Dataset size and stability.

DatasetSlidesEvaluation scope
BML111Three class pathology task
HER2192Biomarker classification
AURORA225 from 55 patientsSlide level folds
TCGA KICH109Renal cancer staging

The datasets cover distinct tasks. Two supplementary datasets add breadth. We agree that fold deviations from one training seed do not measure variation across independent runs. AURORA slide level folds do not establish generalization to unseen patients. We limit present claims to slide level observations. Future work will test EnergyMIL on larger WSI datasets with patient separated evaluation and independent validation. Those findings will guide further algorithm development.

  1. Test-set leakage is a serious methodological flaw that directly undermines the reliability of the paper's main experimental results. The authors should rigorously re-evaluate and correct the results reported in the submitted manuscript under a proper evaluation protocol. This is a serious issue that requires concrete corrective action, not empty promises.

  2. Authors have not provided a sufficiently substantiated and verifiable explanation for the discrepancy between the reported parameter count and the released implementation. Their current response remains unconvincing.

  3. Authors have not substantively addressed the Critical Weaknesses raised in my original review.

Given these unresolved concerns, I have decided to lower my rating.

AuthorsOpenReview ↗

We acknowledge the shortcomings in our assessment and consider this report a valuable lesson learned. Thank you.

Thank you for identifying this issue so clearly. It is a lesson we will carry into future method development, including code review, evaluation design, and validation on larger datasets. In light of your comments and the earlier reviews, we have revised and reorganized our response. We recognize that these revisions do not resolve the methodological flaw. They reflect our effort to examine it honestly and account for what went wrong. Thank you again for your time and candid feedback.

Reviewer_rB3BOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EnergyMIL formulates contextual aggregation for whole-slide image classification as a strongly convex energy-minimization problem. Learned anchors provide factorized cross-patch interactions, while proximal refinement approaches a unique contextual state. Training uses detached iterations followed by one differentiable update, and attention and mean-state readouts produce slide predictions. The paper establishes properties of the inner optimization problem and evaluates the method on four primary datasets and two supplementary datasets using frozen CONCH features. Results are promising on some tasks, but the evidence does not clearly establish the necessity or practical advantage of the optimization-based formulation over simpler aggregation.

Critical strengths

A coherent formulation with explicit mathematical guarantees. The factorized coupling and constrained feature map provide a well-defined contextual energy with a unique minimizer and convergence guarantees under the stated conditions. Avoiding a dense patch-interaction matrix and retaining only the final differentiable update are sensible design choices for large WSI bags.

Promising results supported by broad baseline comparisons. The paper compares numerous MIL methods under shared features and folds, and the BML improvements are substantial. Component ablations and paired statistical analyses provide useful evidence. Reporting less favorable supplementary results also helps delineate the method’s current scope.

Critical weaknesses

The added optimization machinery is not sufficiently justified by the task or experiments. The method combines anchor coupling, a constrained symmetric feature map, norm estimation, adaptive step sizes, iterative state inference, and an approximate gradient. These components yield attractive properties of the inner optimization problem, but uniqueness and convergence do not themselves establish that the resulting representation is more discriminative or generalizable for slide classification.

Table 2 makes the necessity of iterative inference particularly unclear. On BML, the one-step variant achieves macro-F1 0.9217 versus 0.9297 for the full model. On TCGA-KICH, its AUC is higher, 0.6647 versus 0.6337. Moreover, Appendix D.4 states that the one-step variant starts from zero and therefore contains no anchor-mediated cross-patch coupling. This comparison changes both the presence of contextual interaction and the extent of optimization, so it does not isolate the benefit of approaching the energy minimum.

A more informative comparison would retain the same coupling and readouts while using a small fixed number of interaction steps. Without such evidence, it remains unclear whether the optimization-defined endpoint adds substantial value beyond ordinary contextual aggregation. Measured iteration counts and runtime are also needed to assess the practical cost of this formulation.

The empirical advantage is not sufficiently robust or general to resolve this concern. The headline ranking across primary datasets obscures considerable variation in effect size. On AURORA, accuracy ties MammothMIL, while AUC and macro-F1 improve by only 0.002 and 0.001. In the reported paired analysis against MammothMIL, only BML AUC remains significant after Holm correction. The supplementary TCGA-KIRP results do not show an advantage over the strongest baselines.

The primary datasets contain only 109–225 slides each, all experiments use CONCH features, and evaluation uses one training seed and one fixed five-fold partition. These results provide limited evidence that the benefits persist across training variability or representation choices. Patient-level separation also requires clarification, especially for AURORA’s 225 slides from 55 patients, as does the independence of hyperparameter selection from evaluation. Given the added machinery, stronger evidence of a reproducible benefit is needed.

Questions for authors

Necessity of the optimization formulation: How does EnergyMIL compare with a small fixed number of anchor-interaction updates using the same embedding and readouts? Can these controls distinguish the benefit of contextual interaction from the benefit of converging toward a unique energy minimizer?

Practical cost and convergence: What are the observed iteration counts, frequency of reaching the 400-update cap, inference latency, and training memory usage? Does approaching the minimum more closely consistently improve classification? Please confirm whether the reported FLOPs include all solver iterations and matrix operations.

Evaluation reliability: Were folds patient-disjoint, and were hyperparameters selected independently of held-out evaluation folds? Do the gains persist across training seeds and paired comparisons against the strongest relevant baselines? These answers would materially affect my recommendation.

2 comments
AuthorsOpenReview ↗

We added matched depth comparisons, corrected the computational accounting, and clarified EnergyMIL’s memory use and evaluation protocol.

Q1. Necessity of the optimization formulation

Thank you for proposing a matched depth comparison. We held the patch embedding, normalized anchor coupling, readouts, training settings, and folds fixed while changing the number of updates. A budget of KK applies K−1K-1 detached updates and one differentiable update. Every variant with at least two updates therefore contains cross patch interaction, unlike the submitted one update ablation. On TCGA-KICH, ten updates improve ACC and F1 by 2.8 and 7.4 points over one update. A further control replaces the energy state with two trainable residual anchor layers of nearly equal size. Its ACC and F1 are 0.789 and 0.560, compared with 0.816 and 0.638 for EnergyMIL in the same new runs. BML shows no consistent depth advantage. Five to ten updates are competitive with convergence, so approaching the minimum more closely does not consistently improve classification. The optimization formulation supplies a unique contextual reference state and a shared refinement rule. The table gives descriptive percentage point differences from the convergence run within the new experiment. Its absolute scores differ from the submitted Table 2 and will be reconciled in the revision.

UpdatesBML ACCBML AUCBML F1TCGA-KICH ACCTCGA-KICH AUCTCGA-KICH F1
1+2.7+2.8+5.8-1.8-1.8-4.8
20.0+0.3+1.8-1.8-0.8-6.4
5+2.7+2.1+5.8+0.1+3.5+1.6
10+1.8+3.2+3.9+1.0+3.5+2.6
Convergence0.00.00.00.00.00.0

Q2. Practical cost and convergence

Thank you for requesting complete cost accounting. The submitted 2.63G figure came from a module profile that omitted solver matrix products, so it did not include the full refinement. Reprofiling the released implementation across every detached update and the final differentiable update gives 91.0 GFLOPs for a slide with 10,000 patches. In the new convergence runs, mean and maximum detached updates are 233.5 and 262 on BML, and 258.0 and 304 on TCGA-KICH. None of the 220 evaluated slides reaches the 400 update cap. On an RTX 5090, Peak GPU memory is 0.12 GB for inference and 0.17 GB for training. EnergyMIL has 0.535M total parameters, fewer than MammothMIL and TransMIL, while using substantially less peak memory than both. Thus the practical advantage is compact model size and low memory use during global contextual refinement. We will correct the FLOP accounting in Table 3. The table combines submitted parameter counts with the newly measured computational costs, reported as medians across three independent processes.

MethodSubmitted params. (M)Inference GFLOPsInference peak GBTraining peak GB
EnergyMIL0.47427.70.120.17
ABMIL0.336.60.080.14
MammothMIL0.615.80.770.97
TransMIL2.41105.70.551.19

Q3. Evaluation reliability

Thank you for asking about the scope of the evaluation. All methods used the same frozen features and stratified folds. EnergyMIL used one shared setting, M=32M=32, m=0.3m=0.3, and τ=0.1\tau=0.1, across the four datasets, with no dataset specific adjustment on HER2 or AURORA. Baseline architecture settings followed the original papers. The BML and TCGA-KICH sensitivity analyses used the reported held out folds, no nested validation was performed, and the released procedure selected checkpoints by held out accuracy. We will disclose this dependence rather than describe the results as test independent estimates. AURORA was split at the slide level, not by patient, so its results support matched slide level comparison rather than unseen patient generalization. The study used one training seed, and fold deviations do not measure training seed variability. In the paired analysis against MammothMIL, BML AUC remains significant after Holm correction, while the smaller differences do not. We will align the strength of the claims with this evidence.

Thank you for the additional experiments and candid clarifications. The matched-update comparison addresses my concern about the original one-step ablation. However, the results still do not clearly justify the full optimization procedure: 5–10 updates are competitive with convergence, and BML shows no consistent benefit from additional updates.

The revised cost accounting also strengthens my concern about the trade-off. Low memory use is a useful advantage, but the computational cost is substantially higher than originally reported. Please clarify the difference between 91.0 GFLOPs in the text and 427.7 GFLOPs in the table, and provide measured latency.

My main concern remains the evaluation. Using the held-out folds for sensitivity analysis and checkpoint selection limits their value as independent evidence. The slide-level split on AURORA also leaves patient-level generalization unresolved. These issues require additional validation beyond qualifying the claims.

Overall, I appreciate the effort in the rebuttal, but I am still not convinced that the demonstrated gains justify the added complexity. I will maintain my weak reject.

Reviewer_W6QaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper proposes EnergyMIL, a multiple instance learning framework that formulates contextual representation learning for whole-slide image (WSI) classification as a strongly convex energy minimization problem. Unlike conventional contextual MIL methods that define feature interactions through specific network architectures, EnergyMIL explicitly defines the contextual representation as the unique minimizer of a slide-dependent energy function.

The method consists of three main components: anchor-based global coupling for efficient cross-patch information exchange, proximal refinement for approximating the unique contextual state with memory-efficient gradient computation, and complementary focal and global readouts for slide-level classification. The authors also provide theoretical guarantees regarding the uniqueness of the contextual state and the convergence of the optimization procedure.

Experiments are conducted on four primary pathology datasets, with comparisons against 13 MIL baselines. EnergyMIL achieves the highest or tied-highest reported mean performance across all twelve dataset–metric combinations. The paper further includes component ablations, computational complexity analysis, parameter sensitivity studies, and additional evaluations on two supplementary datasets.

Critical strengths

1. Principled formulation of contextual MIL.

The paper introduces an interesting perspective by defining contextual representations through an explicit optimization objective rather than a fixed sequence of neural network operations. The combination of anchor-based global coupling and strongly convex energy minimization provides a structured approach to modeling slide-wide interactions. I find this formulation conceptually meaningful and a potentially valuable direction for contextual MIL.

2. Clear theoretical foundation and memory-conscious design.

The authors provide mathematical analysis establishing the uniqueness of the contextual state and convergence guarantees for the refinement procedure under the stated assumptions. Furthermore, the factorized anchor coupling and detached iterative refinement are designed to accommodate large WSI bags without storing dense patch-to-patch interactions or the full optimization trajectory. These design choices are relevant to the computational challenges of WSI classification.

3. Extensive baseline comparisons and supporting analyses.

The paper compares EnergyMIL with a broad set of established and recent MIL methods under shared feature representations and cross-validation splits. The reported performance is promising across the four primary datasets. I also appreciate the inclusion of component ablations, paired statistical analyses, parameter sensitivity experiments, and additional datasets, including results where EnergyMIL does not outperform competing approaches.

Critical weaknesses

1. Insufficient ablation evidence for the contribution of energy-based optimization.

Although the paper provides several component ablations, I am not fully convinced that the reported improvements can be specifically attributed to the proposed optimization-defined contextual representation. In particular, the one-step variant in Table 2 starts from Z0=0Z_0=0, which eliminates the anchor-mediated cross-patch interaction in that update. Consequently, the comparison between the full model and the one-step variant simultaneously changes the amount of contextual interaction and the degree of iterative refinement. It therefore does not adequately isolate the benefit of converging toward the strongly convex energy minimum.

A more informative comparison would retain the same anchor coupling and contextual interactions while varying the number of refinement steps, or compare against a matched finite-depth contextual model. This would help establish whether the energy-based formulation provides benefits beyond conventional iterative feature aggregation.

2. Insufficient evidence for the claimed computational efficiency.

The paper reports a relatively low computational cost of 2.63G FLOPs. However, the proposed method involves iterative refinement, whose computational complexity depends on the number of iterations. Appendix D.3 describes a cost of O(TNd2)O(TNd^2) under the reported evaluation order, where TT denotes the number of refinement steps. It is not sufficiently clear how the reported FLOPs account for the complete iterative procedure.

Furthermore, the actual iteration counts, end-to-end inference latency, and peak memory usage are not adequately characterized. Since computational scalability is an important motivation for the proposed framework, a more transparent evaluation of its practical computational cost is needed.

3. Concerns regarding experimental protocol and reliability of the reported improvements.

The experimental design leaves several important questions unresolved. First, the reported hyperparameters are selected based on sensitivity analyses on BML and TCGA-KICH, but it is unclear whether this selection was performed independently of the held-out folds used for the main results. Without a clearly separated validation procedure, the reported performance may be affected by model selection bias.

Second, AURORA contains multiple slides from the same patients, while the paper describes stratified five-fold cross-validation without explicitly confirming patient-level grouping. This makes it difficult to rule out potential patient overlap between training and evaluation folds.

Finally, the experiments rely on a single random seed, and although the paper includes paired statistical analyses, only a limited subset of improvements remains statistically significant after multiple-comparison correction. These limitations make it difficult to fully assess the robustness and generalizability of the reported advantages.

Questions for authors

Q1. Contribution of iterative energy minimization

In Table 2, the one-step variant starts from Z0=0Z_0=0, meaning that anchor-mediated cross-patch interaction is absent from its update. Could the authors provide comparisons that preserve the same anchor coupling while varying the number of refinement steps, such as 2, 5, 10, and convergence-based refinement? This would help determine whether the improvement comes from the optimization-defined equilibrium state rather than simply introducing cross-patch interaction.

Q2. Practical computational cost

Could the authors clarify whether the reported 2.63G FLOPs include all detached refinement iterations and the final differentiable update? What are the average and maximum numbers of refinement iterations during inference? Providing end-to-end inference time and peak GPU memory usage, ideally alongside representative MIL baselines under the same hardware and input conditions, would help substantiate the computational efficiency claims.

Q3. Hyperparameter selection and evaluation independence

The paper states that the final values of MM, mm, and τ\tau are selected based on sensitivity analyses on BML and TCGA-KICH. Could the authors clarify whether these analyses used the same held-out folds reported in Table 1? Was an independent validation procedure or nested cross-validation employed? If the reported evaluation folds influenced hyperparameter selection, how would this affect the interpretation of the main results?

Q4. Patient-level separation and result stability

AURORA contains 225 slides from 55 patients. Could the authors confirm whether all slides from the same patient were assigned to the same cross-validation fold? Additionally, given that the evaluation uses a single random seed and several performance improvements are relatively small, could the authors provide further evidence of stability across different random seeds or independent data partitions?

Potential rating update

My current assessment is 2/4. I appreciate the principled formulation and theoretical analysis, but the concerns regarding mechanism isolation, computational cost, and experimental protocol prevent me from fully supporting acceptance. I would consider increasing my rating to 3/4 if the authors can convincingly clarify the evaluation protocol, substantiate the practical computational efficiency, and provide stronger evidence that the iterative energy-minimization mechanism offers benefits beyond matched contextual aggregation baselines.

2 comments
AuthorsOpenReview ↗

We added some key experiments, corrected the computational cost, and clarified the evaluation protocol and its limitations.

Q1. Contribution of iterative energy minimization

Thank you for identifying the confound in the one update ablation. We added controlled five fold runs that hold the embedding, anchor coupling, readouts, training settings, and folds fixed. A budget of KK uses K−1K-1 detached updates and one differentiable update, so every variant from two updates onward includes cross patch interaction. On TCGA-KICH, ten updates improve ACC and F1 by 2.8 and 7.4 points over one update. A further control replaces the energy state with two trainable residual anchor layers of nearly equal size. Its TCGA-KICH ACC and F1 are 0.789 and 0.560, compared with 0.816 and 0.638 for EnergyMIL in those runs. BML shows no consistent depth advantage. The evidence supports structured anchor refinement, while a finite budget can match convergence in classification. The energy minimum supplies a unique reference state. The table gives descriptive percentage point differences from the convergence run. These new runs use a separate implementation snapshot whose absolute scores differ from the submitted Table 2. We will reconcile the discrepancy in the revision.

UpdatesBML ACCBML AUCBML F1TCGA-KICH ACCTCGA-KICH AUCTCGA-KICH F1
1+2.7+2.8+5.8-1.8-1.8-4.8
20.0+0.3+1.8-1.8-0.8-6.4
5+2.7+2.1+5.8+0.1+3.5+1.6
10+1.8+3.2+3.9+1.0+3.5+2.6
Convergence0.00.00.00.00.00.0

Q2. Practical computational cost

Thank you for requesting complete cost accounting. EnergyMIL's real advantage lies in its low number of model parameters and low peak GPU memory usage during context refinement. With 10,000 patches and 512 dimensional features on an RTX 5090, EnergyMIL uses 0.12 GB at inference and 0.16 GB during training. These values are close to ABMIL and below the contextual baselines shown in the table. Detached refinement keeps training memory low even when many updates are required. The submitted 2.63G figure omitted solver matrix products. Counting all matrix products, including detached updates and the final differentiable update, gives 91.0 GFLOPs for the released implementation. We will correct Table 3 and frame the efficiency claim around memory and storage scaling. The table reports medians across three independent processes.

MethodParamInference GFLOPsInference peak GBTraining peak GB
EnergyMIL0.4791.00.120.16
ABMIL0.336.60.080.14
MammothMIL0.615.80.770.97
TransMIL2.41105.70.551.19

Q3. Hyperparameter selection and evaluation independence

Thank you for asking us to distinguish tuning from evaluation. EnergyMIL used one shared setting, M=32M=32, m=0.3m=0.3, and τ=0.1\tau=0.1, across all four datasets. HER2 and AURORA therefore provide a transfer check without dataset specific parameter tuning. Baseline architecture settings followed the original papers without an additional search, and all methods used the same features, folds, loss, optimizer, and maximum training budget. The EnergyMIL sensitivity analyses on BML and TCGA-KICH used the reported held out folds. No independent validation or nested cross validation was used, and the released procedure selected each checkpoint by held out accuracy. We will disclose these selection steps and interpret the submitted results as comparisons under the shared protocol rather than test independent estimates.

Q4. Patient level separation and result stability

Thank you for raising patient identity and training variability. Every method used the same stratified AURORA partitions, providing a matched slide level comparison. The folds were not grouped by patient, so slides from one patient could enter different folds. We will state this scope explicitly and reserve unseen patient generalization for a grouped evaluation. In addition, our study used one training seed, and fold standard deviations describe partition variation rather than independent training runs. The paired analysis provides complementary evidence, with BML AUC remaining significant against MammothMIL after Holm correction. We will limit claims about the smaller gains accordingly.

Thanks to the authors for the honest response and the extra experiments.

After reading the rebuttal and the other discussions, my main concerns are still there. First, selecting checkpoints by held-out test accuracy (along with running sensitivity checks on the held-out folds and splitting AURORA at the slide level rather than by patient) makes it hard to trust the main comparisons without a full re-evaluation under a clean train/val/test split. Second, the new step-budget table shows that 5-10 updates already match or beat running to convergence, while full refinement actually costs 91.0 GFLOPs instead of the 2.63G reported in the paper (and runtime latency is still missing). This makes it less clear why solving all the way to the energy minimum is needed in practice.

That said, the energy formulation itself is still an interesting direction, so I will keep my score at 2 and hope these comments are useful for a future revision.

Comments on the paper

1 comment
Area_Chair_vLdMOpenReview ↗

Gentle reminder: discussion closing soon

Dear Reviewer,

As the discussion period closes soon, we kindly encourage you to engage in the discussion and share any remaining thoughts or feedback.

Thank you for your time and valuable contributions!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.