acceptodds
Under review as a conference paper at ICLR 2027

Representation Composition Matters: Analyzing and Fusing Pathology Foundation Models

Abstract

The emergence of pathology foundation models (FMs) provides diverse visual representations for computational pathology, yet when and how to effectively combine multiple FMs remains unclear. Existing fusion approaches typically aggregate multiple foundation models using fixed aggregation or learned fusion mechanisms, without explicitly considering how encoder utility depends on representation composition and downstream task requirements. To address this gap, we conduct a three-perspective analysis of multi-foundation-model fusion, examining representation geometry, feature-wise statistical association, and downstream functional contribution. Through this analysis, we find that representation similarity alone is insufficient to characterize functional redundancy across the evaluated foundation-model compositions. Moreover, no single fusion strategy among those evaluated consistently dominates across compositions, and fusion performance does not improve monotonically as the number of integrated encoders increases, indicating that fusion effectiveness varies with representation composition. Motivated by these findings, we propose CoR-Fuse, a lightweight, plug-and-play consensus-aware fusion module for adaptive multi-foundation-model integration. Experimental results on TCGA-UVM and TCGA-BLCA show that CoR-Fuse achieves competitive performance across the evaluated foundation-model compositions, while requiring fewer parameters and FLOPs than attention-based fusion strategies. Code is provided in the supplementary material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
computational pathology, pathology foundation models, representation learning, whole slide image

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_g8h9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates how the composition of pathology foundation-model representations affects downstream fusion. It analyzes 13 encoders on TCGA-UVM through linear CKA, cross-feature Spearman correlations, and mean-baseline encoder ablations. The reported results suggest that adding encoders does not consistently improve performance and that fusion-method rankings vary across encoder compositions.

Motivated by these observations, the authors propose CoR-Fuse. Encoder-specific projections are initialized through mean-fusion training, after which a shared MLP conditions on each encoder’s slide descriptor and the cross-encoder mean to generate slide-specific weights. These weights fuse coordinate-aligned patch embeddings before downstream aggregation.

Evaluation covers UVM D3/M3 classification and BLCA TP53 mutation prediction, using patient-level five-fold cross-validation, multiple encoder compositions, baseline comparisons, and component ablations. CoR-Fuse achieves competitive performance with low additional fusion overhead and lower observed dispersion across the evaluated UVM compositions.

Critical strengths
  1. Useful comparison across representation compositions. Evaluating 13 pathology encoders and comparing fusion strategies under three fixed-size compositions helps separate encoder count from encoder identity. The changes in method ranking caution against assuming that more encoders or more complex fusion automatically improves performance. This is useful evidence for practitioners selecting among available pathology representations.

  2. A simple fusion mechanism with low additional overhead. CoR-Fuse generates one set of weights per slide while retaining patch correspondence for downstream aggregation. In Table 1, its router adds approximately 0.06M trainable parameters over mean fusion and has lower reported downstream FLOPs than the evaluated attention alternatives. This makes the method practically attractive when foundation-model embeddings are already available; the reported costs exclude backbone feature extraction.

Critical weaknesses
  1. The mean-baseline LOO analysis does not sufficiently establish encoder utility or synergy. Equations (5)–(6) quantify a fixed classifier’s sensitivity to replacing an encoder representation with a mean baseline. This does not by itself establish uniquely supplied information or the predictive benefit of removing an encoder and retraining. Mean substitution may disrupt correlations among encoder representations, while classifier nonlinearity can produce nonzero interactions even for redundant information. The manuscript insufficiently specifies baseline estimation, replacement space, the reference predictor, and the target-class convention, and does not validate its conclusions against alternative baselines or reference models.

  2. The incremental contribution of consensus routing is not sufficiently isolated or demonstrated against closely related methods. The paper does not directly compare CoR-Fuse with the cited AdaFusion and Shazam, both of which investigate adaptive integration of pathology foundation models. Component ablations cover only one five-encoder composition and omit a full consensus-by-warm-start factorial comparison. The interpretation of the cross-encoder mean as semantic consensus also lacks direct validation. Moreover, the BLCA AUC gain over mean fusion is only 0.01, Table 1 reports dispersion across compositions rather than uncertainty in paired improvements, and the full-model BLCA results in Tables 2 and 3 are inconsistent. Parameters and downstream FLOPs support low fusion overhead, but unreported latency and peak memory leave practical efficiency incompletely substantiated.

Questions for authors
  1. Can the authors explicitly define the mean-baseline LOO protocol, including baseline estimation, replacement space, reference predictor, target class, and evaluation split, and substantiate its interpretation as encoder utility rather than intervention sensitivity? Supporting evidence could include robustness to alternative baselines/reference models or duplicated-encoder and removal-and-retraining controls. Please also provide held-out slide-level LOO distributions and signed pairwise interaction summaries with uncertainty, including the signed results mentioned in Figure 6.

  2. Can the authors explain and experimentally validate why the cross-encoder mean provides useful semantic consensus, using a capacity-matched nonlinear gate without cross-encoder context and a consensus-by-warm-start factorial ablation? Please match nonlinearities, encoder biases, initialization, and total optimization budgets, evaluate across the three composition regimes, and provide training curves or failure-rate evidence for the claimed optimization-stability benefit.

  3. Can the authors add direct experimental comparisons with both AdaFusion and Shazam on the evaluated tasks, using identical patient splits and matching encoder compositions, feature availability, downstream heads, and tuning budgets where applicable? Please document necessary method-specific adaptations and report predictive performance alongside computational cost, so that the contribution beyond existing adaptive pathology-foundation-model fusion can be assessed.

  4. Can the authors provide paired performance differences and uncertainty against mean/global fusion and an inner-validation-selected single encoder, clarify the aggregation of folds, seeds, and compositions, and reconcile the BLCA full-model results in Tables 2 and 3? Please also report per-slide latency/throughput, peak GPU memory, and total training time including warm start under matched hardware, precision, and patch counts, distinguishing cached-feature fusion from backbone extraction, to substantiate the practical lightweight claim.

4 comments
AuthorsOpenReview ↗

Response to W1

We thank the reviewer for this important and technically well-founded comment. We agree that our original interpretation of the mean-baseline LOO and pairwise interaction analyses was too strong.

We now describe the LOO results as baseline-dependent, fixed-predictor sensitivity analysis, rather than as a direct measure of encoder-specific predictive utility. Similarly, we clarify that the pairwise inclusion–exclusion score captures the non-additive response of the fixed predictor to pairwise baseline replacement. We no longer interpret definitive evidence of intrinsic synergy or interference between encoders.

Importantly, we emphasize that this analysis is an auxiliary diagnostic rather than the principal evidence for the proposed fusion method. The motivation for adaptive routing is supported more broadly by the observed variation in fusion performance across FM compositions and by the sample-dependent behavior of the fixed-predictor sensitivity analysis. We have revised the manuscript to ensure that the LOO and pairwise interaction results are not used to make stronger claims than their experimental design supports.

AuthorsOpenReview ↗

Response to W2

We thank the reviewer for the constructive and detailed comments. We have revised the manuscript and expanded the experiments accordingly.

We acknowledge the relevance of AdaFusion and Shazam, which also investigate the adaptive integration of multiple pathology foundation models.

Regarding implementation, the publicly accessible AdaFusion repository indicating that the code will be released in the future.

Shazam, in contrast, require extracting intermediate-layer representations from the teacher backbones and implementing its multi-scale aggregation and distillation pipeline, rather than simply replacing the fusion module while keeping the existing feature inputs and downstream protocol unchanged. We therefore distinguish Shazam as a complementary framework-level approach rather than treating it as a directly interchangeable fusion-head baseline.

Our evaluation includes several categories of fusion baselines: Fixed aggregation: Mean fusion, Concatenation, Global learned weighting: Global weighting fusion Slide-specific weighting: Gated fusion and Attention-based fusion: Cross attention and Self attention.

We also compare CoR-Fuse with three pathology-specific methods: Meta-Encoder(Gaoetal.,2026), FuseCPath(Yangetal.,2025) and ELF(Luoetal.,2026) in pathology research field. To further extend the comparison to general multimodal feature fusion, we also compared EmbraceNet (Choi&Lee,2019).

Following the reviewer's suggestion, we add the baseline results in ablation study:

Table. Ablation study of CoR-Fuse components.

MethodTCGA-UVM AUCTCGA-UVM AUPRCTCGA-BLCA AUCTCGA-BLCA AUPRC
w/o Both0.72±0.080.72 \pm 0.080.80±0.160.80 \pm 0.160.69±0.050.69 \pm 0.050.70±0.040.70 \pm 0.04
w/o Consensus0.79±0.220.79 \pm 0.220.83±0.200.83 \pm 0.200.72±0.030.72 \pm 0.030.72±0.100.72 \pm 0.10
w/o Warm Start0.73±0.230.73 \pm 0.230.77±0.220.77 \pm 0.220.72±0.020.72 \pm 0.020.71±0.110.71 \pm 0.11
CoR-Fuse0.81±0.090.81 \pm 0.090.84±0.100.84 \pm 0.100.75±0.070.75 \pm 0.070.76±0.120.76 \pm 0.12

The cross-encoder mean is constructed by averaging the projected representations of different foundation models in a shared dimensional space. In our method, it serves as a reference representation that provides shared cross-encoder context to the routing network.

We agree that the AUC improvement over Mean Fusion on TCGA-BLCA is modest (0.01). We do not intend to claim that CoR-Fuse consistently produces large predictive performance gains over simple fusion strategies across all datasets.

Our main objective is to investigate whether lightweight consensus-conditioned routing can achieve competitive performance across different foundation-model compositions without introducing the substantial additional parameters and computational cost associated with more complex attention-based fusion mechanisms.

AuthorsOpenReview ↗

Response to Q1-Q4

We thank the reviewer’s questions.

Response to Q1

In our analysis section, the reference predictor is trained using the complete set of encoders and is kept fixed throughout the LOO evaluation. For each held-out slide, we replace the representation of one encoder with its corresponding encoder-specific mean baseline, while keeping the representations of all other encoders unchanged. The prediction difference is then evaluated using the target-class one-vs-rest logit margin, defined as the logit of the ground-truth class minus that of the alternative class. Importantly, no model retraining is performed after representation replacement.

Overall, we have narrowed the role of this analysis to an auxiliary diagnostic of sample-dependent predictor sensitivity, rather than using it as direct evidence of encoder-specific predictive utility.

Response to Q2

We clarify that the cross-encoder mean is computed by averaging the projected slide-level descriptors of all encoders and is used as a shared cross-encoder reference to condition the routing network. We do not claim that this arithmetic mean constitutes a semantically validated consensus representation. Rather, its purpose is to provide cross-encoder context when estimating sample-specific routing weights.

Regarding the component contributions, we have included a factorial ablation of consensus and warm-start, as detailed in our response to Weakness 2 and the revised ablation table.

Then we narrowed the corresponding claims and avoid interpreting the cross-encoder mean as intrinsically semantic or the warm-start strategy as a demonstrated optimization-stability guarantee.

Response to Q3

In response to W2, we discuss the reasons why we did not select these two methods for comparison.

In particular, we clarify that AdaFusion and Shazam operate under different fusion paradigms and feature requirements from CoR-Fuse. Consequently, reproducing these methods using only the final-layer patch embeddings employed by our current pipeline would not constitute a faithful implementation, particularly for Shazam, which requires additional intermediate features and its associated multi-scale training procedure.

Response to Q4

Regarding peak GPU memory, we agree that memory consumption is an important aspect of practical computational efficiency. However, our efficiency analysis is specifically designed to characterize the additional computational overhead of the cached-feature fusion stage, rather than the resource requirements of the entire multi-encoder pipeline.

The measured GPU memory usage would depend not only on the fusion module itself but also on the handling and residency of these shared input feature tensors.

I thank the authors for the response. The clarified LOO protocol and the narrowed interpretation of the LOO, interaction, and cross-encoder mean analyses make the claims more accurate, and I accept that AdaFusion's code is not yet public. However, the rebuttal adds little new empirical evidence. The factorial ablation table contains no numbers, and no capacity-matched gate without cross-encoder context is provided. Since the authors no longer interpret the mean as semantic consensus, such a comparison is needed to show that consensus conditioning adds anything over standard context-conditioned gating. With the contribution now framed as competitive, lightweight fusion and a 0.01 AUC gain on BLCA, paired fold-level differences with confidence intervals, a comparison against the best single encoder, and runtime measurements (fusion latency and total training time, excluding backbone extraction) are needed, but none were provided. The inconsistency in the BLCA results between Tables 2 and 3 was also not addressed. I therefore maintain my score.

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

This paper investigates the relationship between representation composition in computational pathology. The authors analyze 13 pathology foundation models from three perspectives: representation geometry using linear CKA, feature-wise statistical association using Spearman correlation, and task-level contributions using leave-one-encoder-out attribution and pairwise interactions.

Based on these analyses, the authors propose CoR-Fuse, a consensus-aware fusion framework. The method first trains encoder-specific projection heads and a downstream classifier using mean fusion, then introduces a slide-specific routing network conditioned on individual encoder descriptors and their average representation. The learned routing weights are applied to patch-level features before downstream prediction.

Experiments are conducted on TCGA-UVM and TCGA-BLCA, evaluating different encoder combinations, fusion strategies, downstream classifiers, and computational costs. The proposed method achieves competitive performance, with relatively small differences compared with conventional fusion baselines.

Critical strengths
  1. Exploratory comparison of multiple pathology foundation model representations.

The paper analyzes representations from 13 pathology foundation models using coordinate-aligned patches and examines several aspects of their relationships. The inclusion of different encoder compositions and standard fusion baselines provides descriptive evidence of how fusion performance can vary across model combinations. This analysis has some practical value as an exploratory comparison, although its implications for the proposed method remain insufficiently established.

Critical weaknesses
  1. Limited methodological novelty and disconnect from diagnostic analysis

The paper fails to establish a sufficiently distinct contribution relative to existing dynamic fusion and mixture-of-experts literature, nor does it bridge its diagnostic findings with the proposed design. Although the authors analyze CKA similarity, feature correlations, and leave-one-out contributions, none of these metrics are operationalized in the routing objective or consensus mechanism. Instead, CoR-Fuse relies on a standard context-conditioned gating network that concatenates a mean-consensus descriptor with individual features before an MLP and softmax—an approach closely resembling conventional dynamic fusion. Furthermore, the paper provides no evidence that the learned routing weights align with the functional complementarity identified in the analysis, leaving the connection between the motivation and methodological design unsubstantiated.

  1. Insufficient empirical validation and marginal performance gains

The experimental evaluation fails to demonstrate a compelling advantage over simpler alternatives. In main results, CoR-Fuse improves AUC by only about 0.010.01 on TCGA-UVM and TCGA-BLCA, with competitive methods often matching or outperforming it depending on encoder composition. Ablations show that removing the consensus component causes only negligible drops in performance, and routing Gini coefficients remain nearly unchanged, suggesting the gains may stem from parameter capacity or optimization setup rather than consensus conditioning. Additionally, the evaluation relies heavily on a small cohort (TCGA-UVM with 80 cases), uses nested rather than independently sampled encoder sets for scaling experiments, and omits end-to-end efficiency comparisons that account for multi-backbone feature extraction costs.

Questions for authors
  1. What is the precise methodological distinction between CoR-Fuse and existing context-conditioned gating or mixture-of-experts fusion approaches?

The paper argues that previous learned fusion methods insufficiently consider representation composition. However, the proposed router is also learned from downstream classification supervision and uses the average of encoder descriptors as additional context. Could the authors clarify what representation-composition property is explicitly modeled by CoR-Fuse that cannot be captured by a conventional gating network with comparable capacity and global context?

  1. How do the representation analyses contribute to the proposed routing mechanism?

Can the authors quantitatively demonstrate that the learned routing weights correspond to leave-one-encoder-out contributions or task-relevant encoder complementarity? Do CKA similarity, feature-wise correlation, or pairwise interaction measures predict when CoR-Fuse improves over mean fusion or other learned fusion methods? Without this evidence, the relationship between the diagnostic analysis and the proposed method remains unclear.

  1. Are the performance improvements statistically reliable and attributable to consensus conditioning?

Could the authors provide paired fold-level comparisons and confidence intervals against mean fusion, gated fusion, and cross-attention? In addition, please compare CoR-Fuse against capacity-matched MLP gating and other adaptive routing baselines trained using the same warm-start procedure and optimization budget. These controls are important for distinguishing the contribution of consensus conditioning from model capacity and optimization effects.

  1. How robust are the findings across genuinely different encoder compositions and independent datasets?

The paper evaluates only a small number of representative fixed-size compositions, while the encoder-count analysis uses a single ranked sequence of nested sets. Could the authors evaluate a larger sample of independently selected encoder subsets at fixed cardinality, controlling for the predictive strength of individual encoders? Evaluation on additional datasets would also help determine whether the observed composition-dependent behavior generalizes beyond the current cohorts.

  1. What is the end-to-end performance-efficiency trade-off relative to using a single strong foundation model?

The reported FLOPs exclude the frozen backbone computations. Could the authors provide inference time, total computational cost, and predictive performance comparisons that account for all foundation encoders required by CoR-Fuse? This would clarify the practical benefit of multi-encoder fusion given its relatively small improvements over simple baselines.

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

This paper observes that existing methods for combining and integrating visual representations for computational pathology have significant limitations. The authors argue that representation similarity alone is insufficient to characterize functional redundancy across the evaluated foundation-model compositions. To address this, the authors propose CoR-Fuse, a lightweight, plug-and-play consensus-aware fusion module for adaptive multi-foundation-model integration. Extensive experimental results demonstrate that the proposed method can achieve competitive performance.

Critical strengths
  1. The proposed method is more effective than the naive method.
  2. The authors explore a specific problem in a specific scenario, which is of research significance.
Critical weaknesses
  1. In fact, the studied problem is a multi-view/multi-modal fusion classification issue, which lacks extensive discussion of relevant works, such as low-quality fusion methods, imbalanced fusion methods, etc.
  2. The comparison baselines are insufficient, lacking more convincing baselines. This is actually a downstream task branch of multi-modal fusion.
  3. In addition, the ablation study lacks baseline results. Warm Star seems important for performance improvement. similar strategies have been applied in many existing studies, with limited innovation.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to W1

Thank you for pointing out the connection to broader fusion research. Our setting is more specifically defined as multi-view representation fusion for pathology foundation models (PFMs).

Unlike low-quality fusion methods, we do not explicitly model input corruption or missingness; unlike imbalanced-learning methods, we do not aim to correct branch-level optimization imbalance.

These distinctions also inform our baseline selection. For example, QMF(Zhang et al., ICML 2023) performs prediction-level fusion, whereas OGM-GE(Peng et al., CVPR 2022) intervenes in branch optimization; neither is a direct replacement for a module that outputs a fused patch bag for ABMIL. EmbraceNet(Choi&Lee, 2019) supports feature-level integration and has therefore been implemented as an additional baseline under the same FM inputs and ABMIL head. The experiments use the same encoder composition and random seed, results are as follows:

Table 1. Fusion performance under a representative five-encoder composition on TCGA-BLCA.

MethodAUCAUPRCF1Params (M)FLOPs (G)
Mean Fusion0.74±0.060.74 \pm 0.060.74±0.100.74 \pm 0.100.60±0.100.60 \pm 0.104.0174.81
Concatenation0.74±0.060.74 \pm 0.060.75±0.080.75 \pm 0.080.57±0.230.57 \pm 0.235.3299.31
Global Weight0.74±0.050.74 \pm 0.050.75±0.090.75 \pm 0.090.65±0.070.65 \pm 0.074.0174.81
Gated Fusion0.73±0.060.73 \pm 0.060.74±0.100.74 \pm 0.100.68±0.060.68 \pm 0.064.0174.81
Cross Attention0.74±0.030.74 \pm 0.030.73±0.060.73 \pm 0.060.57±0.170.57 \pm 0.176.11153.62
Meta-Encoder0.64±0.060.64 \pm 0.060.63±0.160.63 \pm 0.160.63±0.110.63 \pm 0.1124.192235.86
FuseCPath0.67±0.060.67 \pm 0.060.67±0.160.67 \pm 0.160.59±0.150.59 \pm 0.157.163.21
EmbraceNet0.74±0.060.74 \pm 0.060.75±0.110.75 \pm 0.110.60±0.070.60 \pm 0.074.01144.63
CoR-Fuse (Ours)0.75±0.070.75 \pm 0.070.76±0.120.76 \pm 0.120.68±0.080.68 \pm 0.084.0474.81

Table 2. Fusion performance under the five-encoder composition {UNI, Virchow2, Midnight-12k, GigaPath, Phikon} on TCGA-UVM.

MethodAUCAUPRCF1Params (M)FLOPs (G)
Mean Fusion0.77±0.120.77 \pm 0.120.76±0.250.76 \pm 0.250.71±0.150.71 \pm 0.153.6140.19
Concatenation0.81±0.130.81 \pm 0.130.80±0.290.80 \pm 0.290.73±0.150.73 \pm 0.154.9354.77
Global Weight0.77±0.120.77 \pm 0.120.75±0.250.75 \pm 0.250.69±0.170.69 \pm 0.173.6140.19
Gated Fusion0.79±0.130.79 \pm 0.130.79±0.240.79 \pm 0.240.66±0.260.66 \pm 0.263.6140.19
Cross Attention0.76±0.160.76 \pm 0.160.79±0.210.79 \pm 0.210.67±0.140.67 \pm 0.145.7287.12
EmbraceNet0.71±0.210.71 \pm 0.210.77±0.160.77 \pm 0.160.69±0.190.69 \pm 0.194.0186.13
CoR-Fuse (Ours)0.81±0.090.81 \pm 0.090.84±0.100.84 \pm 0.100.72±0.150.72 \pm 0.154.0444.55
AuthorsOpenReview ↗

Response to W2

Thank you for raising this concern. Our evaluation includes several categories of fusion baselines: Fixed aggregation: Mean fusion, Concatenation, Global learned weighting: Global weighting fusion Slide-specific weighting: Gated fusion and Attention-based fusion: Cross attention and Self attention.

These baselines assess whether adaptive weighting and interaction modeling provide benefits over simpler aggregation strategies.

We also compare CoR-Fuse with three pathology-specific methods: Meta-Encoder(Gaoetal.,2026), FuseCPath(Yangetal.,2025) and ELF(Luoetal.,2026) in pathology research field. To further extend the comparison to general multimodal feature fusion, we also compared EmbraceNet (Choi&Lee,2019).

AuthorsOpenReview ↗

Response to W3

We thank the reviewer for pointing out the insufficient baseline comparison and the need to better clarify the role and novelty of the warm-start strategy.

We agree that warm-start initialization is an established optimization strategy and do not claim it as an independent algorithmic novelty. Although it have been applied in many existing studies, but it is useful. Its role in CoR-Fuse is to facilitate the optimization of adaptive routing.

Specifically, Stage 1 learns encoder-specific projection heads and the downstream prediction head under uniform mean aggregation. Their parameters are then transferred to Stage 2, where slide-specific routing is introduced. This initialization provides a task-aligned starting point, allowing Stage 2 to focus more directly on refining encoder contributions rather than simultaneously learning feature projection, prediction, and routing from random initialization.

Then we add the baseline results in ablation study:

Table. Ablation study of CoR-Fuse components.

MethodTCGA-UVM AUCTCGA-UVM AUPRCTCGA-BLCA AUCTCGA-BLCA AUPRC
w/o Both0.72±0.080.72 \pm 0.080.80±0.160.80 \pm 0.160.69±0.050.69 \pm 0.050.70±0.040.70 \pm 0.04
w/o Consensus0.79±0.220.79 \pm 0.220.83±0.200.83 \pm 0.200.72±0.030.72 \pm 0.030.72±0.100.72 \pm 0.10
w/o Warm Start0.73±0.230.73 \pm 0.230.77±0.220.77 \pm 0.220.72±0.020.72 \pm 0.020.71±0.110.71 \pm 0.11
CoR-Fuse0.81±0.090.81 \pm 0.090.84±0.100.84 \pm 0.100.75±0.070.75 \pm 0.070.76±0.120.76 \pm 0.12

Thank you to the authors for the clarification and the new experiments. The results appear that the warm-up phase accounts for the majority of the performance gains. Consequently, I remain skeptical regarding the core innovation of the proposed method, and I therefore maintain my current rating.

Comments on the paper

1 comment
Area_Chair_NxqhOpenReview ↗

A Gentle Nudge: Let's Keep the Discussion Going!

Dear Reviewers,

The authors have now posted their responses (except to Reviewer sqex), addressing your concerns and incorporating your suggestions. We are already halfway through the discussion period!

One of the distinguishing features of ICLR is its interactive review process, where constructive discussions not only help strengthen individual papers but also contribute valuable insights to the broader scientific community.

So, let's not let the discussion be a one-way street! I encourage you to revisit the authors' responses, acknowledge their efforts, and engage in further discussion where needed. After all, a good scientific discussion is often as valuable as the final verdict!

-AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.