acceptodds
Under review as a conference paper at ICLR 2027

CORD: Learning Reusable Degradation Representations Across Heterogeneous Physical Systems

Abstract

Can heterogeneous physical degradation systems benefit from joint pretraining and move beyond system-specific prognostics toward reusable cross-system representation learning? CORD combines type-specific observation interfaces with a shared degradation backbone. Its two self-supervised objectives learn at complementary scales: Intra-Observation Structure Modeling (ISM) captures structure within observations, while Inter-Observation Dynamics Modeling (IDM) captures latent degradation evolution across observation histories. We evaluate CORD under two transfer boundaries: Pretraining-Included System Types, where downstream datasets and held-out units are unseen but their system types are represented during source pretraining, and Pretraining-Excluded System Types, where the entire turbofan-engine type is absent from pretraining. Across bearings, batteries, and cutting tools, CORD (Multi-domain) consistently improves over CORD (Single-domain) under Frozen adaptation, provides further gains under Full FT in most settings, and remains competitive with representative external baselines. Source-pretrained initialization also improves low-label adaptation to the pretraining-excluded engine type. Frozen-representation analysis further shows improved cross-unit lifecycle consistency after multi-domain pretraining. Joint pretraining across heterogeneous physical systems thus produces degradation representations reusable across devices, datasets, and system types.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time-Series Representation Learning, Self-Supervised Pretraining, Cross-System Generalization, Heterogeneous Physical Systems, Degradation Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MkwPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Physical degradation measurements differ between systems in terms of channel semantics, physical units, sampling cadence and operating context. For bearings it's vibration, for batteries electrochemical cycling and for turbofans multivariate flight histories. Therefore, industrial prognostics is typically developed per-system, with models that learn system-specific degradation patterns.

While there are transfer learning techniques that are used to adapt knowledge across operating conditions or related domains, they do not investigate if degradation knowledge can be learned jointly across different semantically disjoint physical systems.

The authors introduce CORD, which keeps observation type specific interfaces but shares a degradation backbone across systems. The type-specific interface is split into a global stem that maps each channel across the full sequence length to a global token and a local stem that maps each window-channel pair to a local token. both are pooled across channels and passed to the shared encoder. The shared encoder alternates with a type-specific adapter. The final output projection is shared as well.

It is trained self-supervised with two objectives, Intra-Observation Structure Modeling (ISM), which comes down to masked reconstruction of the local tokens utilizing type-specific decoder and Inter-Observation Dynamics Modeling (IDM), which forecasts the next observation embedding given the past 6 observation embeddings. Different prediction heads utilize the resulting representation on different RUL task.

They test transfer in two stages. Level 1 tests a representation that was pretrained on all available systems on a holdout of one of those systems. Level 2 tests a representation which was pretrained completely without the target system and introduces it only during downstream adaptation. The task they solve is remaining useful lifetime estimation, mapping a (multivariate) time series of the respective system to a scalar RUL value. They test both, probing the frozen representation and full fine-tuning. The evaluations show that CORD is better than the selected baselines and that the pretraining is especially beneficial in the low (labeled) data regime.

The idea to train a RUL estimator jointly on diverse domains to improve its performance is compelling and the results look strong. The authors have properly seeded the experiments, tested probing and full fine-tuning, and have performed interesting ablations like scaling the available labeled data. The paper is very dense and notation-heavy. I think overhauling the notations and aligning the figures with it could greatly improve the overall state of the manuscript. Especially, consistent notation of the tensor shapes in Figure 1 and more details on the observation projection and loss calculation could help to improve clarity. The baselines seem not to be related to prior work that has performed RUL estimation on the used datasets and an assessment of statistical significance is missing. The code is not published. If my questions are adressed properly, I am willing to increase my score to weak acceptance.

Critical strengths
  • The improvement looks good and the idea is interesting
  • The evaluation across label budget, scratch, single domain and multi-domain for each system is extensive
Critical weaknesses
  • Clarity & Notations (see summary & questions)
  • Justification of baselines (see questions)

minor:

  • I think contribution 2 and 3 could be combined
Questions for authors
  • Could you clarify the computational graph? Especially, the part about the observation projection seems to miss an in-detail explanation in the paper. What is the global-state and valid-local mean in Figure 1? Could you additionally clarify the loss calculations by including the model in formula 2 and 3? I'd like to better understand that.
  • Is the improvement shown in table 3 statistically significant? A rank-based evaluation would be interesting.
  • How where the baselines selected? I guess there is literature available that trains RUL estimators on the respective datasets?
  • Why have you decided to use TCN instead of ModernTCN?
  • Have you ablated the training of the task specific heads without the backbone on the RUL task?
4 comments
AuthorsOpenReview ↗

Response to Reviewer MkwP: Methodological Clarifications and Extended Evaluation

We thank the reviewer for the careful assessment and constructive questions. We have revised the methodological presentation and expanded the experimental evaluation. Further details are provided in our newly submitted revised manuscript. We address the concerns below.

1. Computational graph, notation, and loss formulation.

We have redesigned Figures 1–2 and revised Sections 3.2–3.3 to clarify the full computation and unify the notation and tensor dimensions. In particular, Figure 2 now shows how type-specific stems and validity-aware channel pooling produce one global token and local tokens. After the shared Transformer and adapters, the global-token state h0h_0 and the mean of valid local-token states hˉ\bar{h} are concatenated and mapped by the shared observation projection to ete_t (Eq. 1).

Section 3.3 and Eqs. (2)–(3) now specify the two pretraining computations and their losses: ISM reconstructs masked local descriptors using a type-specific reconstruction head, whereas IDM uses a shared GRU to predict the current unmasked observation embedding from its history, with a stop-gradient target. Appendix B provides the network dimensions, masking rules, and optimization details.

2. Statistical analysis and rank-based evaluation.

Following the reviewer's suggestion, we have added paired seed-wise statistical analyses, including 95% confidence intervals, exact Wilcoxon signed-rank tests, and Holm correction (Section 4.4 and Appendix G). These include comparisons between CORD (Multi-domain) and CORD (Single-domain) (Table 32), as well as supplementary comparisons with XGBoost (Table 33), complementing the baseline evaluation in the revised Table 5.

We have also introduced an average RMSE rank in Table 5. Among the eleven methods evaluated across all nine Level-I settings, CORD (Multi-domain) achieves the best average rank (1.33), providing an additional cross-setting comparison beyond individual RMSE values.

3. Baseline selection, RUL-specific methods, and ModernTCN.

We have reorganized the RUL prediction discussion in Section 2 and expanded the baseline comparisons in Section 4.3. We have added four methods: ModernTCN, Regression Shapalet–GNN for bearings, Battery Transformer for batteries, and TCN-BiLSTM-Attention for cutting tools (Section 5.3, Table 5, Appendix E).

TCN remains as a conventional temporal-model baseline, while ModernTCN is now explicitly evaluated as suggested by the reviewer. RUL-specific models retain their native input processing and backbone architectures; all methods use matched downstream splits, label budgets, seeds, and metrics.

We report cases where these specialized predictors outperform CORD, as well as CORD's overall performance across system types.

4. Head-only ablation.

We have added a Head-only experiment in Section 5.4, Table 7, and Appendix F.1. It removes the learned representation network and trains the target-specific RUL heads directly on parameter-free aggregates of observation descriptors, using matched data splits and 10% labels under the same 10%-label setting as the Frozen experiments.

Compared with Head-only, the complete Multi-domain encoder reduces RMSE from 0.1160 to 0.1087 for bearings, 0.0707 to 0.0704 for batteries, and 0.2030 to 0.0864 for cutting tools.

The contribution is particularly pronounced for cutting tools, while the differences are smaller for bearings and batteries.

5. Code availability and organization of contributions.

We have added an anonymous code repository link in the Abstract, with implementations, experiment configurations, and retained records described in the Reproducibility Statement.

We have also reorganized the Introduction's contributions into two points, combining the closely related contributions identified by the reviewer: the cross-system representation-learning framework and empirical evidence of transferability across datasets and system types.

We thank the reviewer again for the feedback, which helped improve both the clarity and experimental support of the paper.

AuthorsOpenReview ↗

Response to Reviewer MkWP: Additional Stage-wise Analysis

As an additional update, we have included a stage-wise RMSE analysis in Appendix E.4 of the revised manuscript, comparing CORD with existing baselines across three physical systems and different label budgets to provide further insights into prediction performance throughout the degradation process.

Thank you for the revision, which addresses all my points and makes both the improvements and the limitations much clearer, so I am raising my score to 3.

AuthorsOpenReview ↗

Thank You for Your Feedback and Reassessment

Dear Reviewer,

Thank you for taking the time to review our revised manuscript and responses. We sincerely appreciate your constructive feedback, your recognition of the improvements, and your willingness to reconsider your assessment.

Your comments have been valuable in helping us improve the clarity and quality of our work.

Thank you again for your time and consideration.

Reviewer_rc3sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Physical systems undergo degradation over time and CORD addresses the question of whether degradation knowledge can be learned jointly across physically different systems. The main idea is to build a shared representation of degradation that can be used across multiple heterogenous physical systems without system-specific training. The proposed methodology introduces a shared backbone across systems while keeping observation interfaces type-specific, where they preserve its systems’ sensing semantics. There are two self-supervised objectives used: first is to capture structural relationships within a health-state observation (ISM) and the second is to capture how latent health states evolve (IDM). Validation uses two transfer boundaries: the first evaluates generalisation under pretraining included system types while the second evaluates generalisation under pretraining-excluded system types. Overall, results indicate that multi domain pretraining improves the reusability of learned degradation representations across evaluated datasets.

Critical strengths
  1. The problem CORD addresses is a realistic industrial problem and addresses a limitation of system specific prognostics models. The usage of 2 complementary self supervised objectives (ISM and IDM) is an interesting approach to utilise both structural relationships within a health state observation and dynamics across the history. Table 6 shows substantial improvements when IDM is incorporated.
  2. This paper also shows strong frozen transfer results. Multi-domain pre-training outperforms single-domain pretraining in all 9 Level I frozen configurations (Table 2). This is good evidence that resulting shared representation is useful without encoder updates
Critical weaknesses
  1. The major limitation is that limited generalisation evidence across unseen physical systems. The paper claims cross-system evaluation but Level II only evaluates one excluded system type and one held out engine which isn’t sufficient.
  2. The ablation study lacks an analysis of architectural contributions. The framework combines residual adapters, shared transformer layers, two pretraining objectives. While the paper studies structured descriptors vs raw-resampled inputs and ISM vs ISM+IDM, it does not isolate the contributions of architectural components.
Questions for authors
  1. Clarification: Would the proposed model retain its advantage if each physical type were excluded from source pretraining in return (table 1)? Example: Could pretraining on batteries and cutting tools improve predictions for unseen bearing systems? This question is related to the first weakness listed above.
  2. Clarification: Could a component-wise ablation be provided for important parts of the framework? How much of the improvement is retained with simpler multi-domain training? This question is related to the second weakness listed above.
  3. Clarification: Why did you set H as 6? Using limited history might miss degradation patterns. So you can also consider tuning it based on the system rather than using same value across systems.
  4. Suggestion for future work: Currently these experiments remain offline. Evaluation is done by pre training and then downstream adapting. How about extending this to continual degradation/ online learning?
5 comments
AuthorsOpenReview ↗

Response to Reviewer rc3s: Extended Transfer Evaluation and Ablation Studies

We thank the reviewer for the constructive feedback and for recognizing the relevance of cross-system degradation modeling and the strong frozen-transfer results. We have expanded the experiments and revised the methodological explanations to address the concerns. Further details can be found in our newly submitted revised manuscript.

1. Generalization to unseen physical system types.

We agree that evaluating only one pretraining-excluded system type was insufficient to establish broader cross-system transferability. We have therefore added three leave-one-system-type-out (LOSO) experiments, in which bearings, batteries, and cutting tools are each excluded from source pretraining in turn. For each excluded type, CORD is pretrained on the other two system types and then adapted to the unseen target type.

Together with the original N-CMAPSS experiment, Level II now evaluates four unseen target system types (Sections 4.2 and 5.2, Table 4, and Appendix D). Multi-domain pretraining improves RMSE over training from scratch in most target–label-budget settings, although the benefits vary across systems and label budgets. These additional experiments provide a broader evaluation of transferability beyond the original turbofan-engine setting.

2. Architectural contributions and component-wise ablations.

We appreciate the suggestion to examine the contributions of individual architectural components. We have added component-wise ablations in Section 5.4, Table 7, and Appendix F.1, evaluating simplified configurations of the multi-domain framework. Specifically, we examine removing type-specific adapters, replacing type-specific stems with shared stems, replacing attention-based channel pooling with mean pooling, and removing the global token. We also include a Head-only comparison without the learned representation network.

The component-wise ablations quantify the performance retained by simpler architectures under the same multi-domain pretraining protocol, while Head-only provides a separate comparison without the learned representation network. For example, removing adapters increases RMSE by 7.1%, 12.2%, and 18.6% for bearings, batteries, and cutting tools, respectively. Shared stems improve battery performance but reduce performance on bearings and cutting tools. Similarly, simplifying channel pooling or removing the global token increases RMSE across all three system types. These results clarify the contributions of the architectural components and show that their effects are system-dependent.

3. Choice of the IDM history length H = 6.

We agree that the choice of history length requires further experimental support. We have added a sensitivity analysis comparing H = 3, 6, and 12 under Frozen adaptation with 10% labeled target data (Section 5.4 and Appendix F.3, Table 30).

The results show that H = 6 achieves the lowest RMSE for bearings and cutting tools, while H = 3 performs best for batteries. Across the three system types, H = 6 obtains the best average RMSE rank (1.67), compared with 2.00 for H = 12 and 2.33 for H = 3. We therefore retain H = 6 as a common pretraining setting, while recognizing that the preferred history length may differ across physical systems.

4. Continual and online learning.

We appreciate the suggestion to extend CORD toward continual or online degradation learning. The current study focuses on offline self-supervised pretraining across heterogeneous physical systems and subsequent downstream adaptation. Extending the framework to update degradation representations as new observations become available is a meaningful direction for future research, particularly for adapting to evolving operating conditions and degradation patterns.

We thank the reviewer again for the constructive suggestions, which have helped strengthen the experimental evaluation and clarify the architectural design choices.

AuthorsOpenReview ↗

Response to Reviewer rc3s: Additional Stage-wise Analysis

As an additional update, we have included a stage-wise analysis in Appendix E.4 of the revised manuscript, comparing CORD with existing baselines across three physical systems and different label budgets to provide further insights into prediction performance throughout the degradation process.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revised Manuscript

Dear Reviewer,

Thank you again for your time and constructive feedback.

We have carefully addressed your comments through substantial revisions and additional experiments in our updated manuscript.

As the discussion period is approaching its end, we would greatly appreciate it if you could review our responses and let us know whether any major concerns remain.

Thank you for your time and consideration.

Reviewer_Re3DOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CORD, a pretraining method for modeling degradation in physical systems. CORD combines Intra-Observation Structure Modeling (ISM), which captures structure within individual observations, with Inter-Observation Dynamics Modeling (IDM), which captures latent degradation dynamics across observation histories. Results from in-domain and out-of-domain evaluations show that CORD outperforms representative time-series baselines.

Overall, the paper addresses an interesting and relevant problem. However, I believe the manuscript requires substantial revision before it is ready for further review. The introduction is difficult to follow: several abbreviations are introduced without their full names, and the motivation for ISM and IDM is not clearly established. The authors also introduce two evaluation levels that both emphasize generalization, but their relationship to evaluating degradation modeling is unclear. The discussion of related work and the baseline comparisons should include methods developed specifically for predictive maintenance, particularly bearing remaining useful life (RUL) prediction methods such as [1]. Section 3 needs substantially more methodological detail and clearer justification for the design choices. Finally, the experimental descriptions require clarification—for example, what do the labels “2_2–2_5” in Table 1 represent?

[1] X. Yang, Y. Zheng, Y. Zhang, D. S.-H. Wong, and W. Yang, “Bearing Remaining Useful Life Prediction Based on Regression Shapalet and Graph Neural Network,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–12, 2022, Art. no. 3505712, doi: 10.1109/TIM.2022.3151169.

Critical strengths

The problem is interesting to the domain. And the motivation to provide a better pretrained model to estimate the RUL is important for the manufacturing domain.

Critical weaknesses
  1. The motivation of the proposed method is not clear. And where does the structure part come from? There are too many details missing in the paper, and the baselines are not competitive enough to prove its effectiveness.
Questions for authors
  1. Full-FT, PHM, and RUL should be given their full name the first time they are shown in the paper.
  2. More baselines, like the graph-based method, should also be considered.
5 comments
AuthorsOpenReview ↗

Response to Reviewer Re3D: Clarifications and Additional Experiments

We thank the reviewer for the detailed and constructive assessment. We have revised the motivation and methodological presentation, expanded the experiments, and clarified the evaluation protocols. Further details of these changes can be found in our newly submitted revised manuscript. We address the concerns as follows.

1. Motivation and the meaning of “structure” in ISM.

We agree that the original explanation did not sufficiently establish the roles of ISM and IDM. We have reorganized the Introduction (Section 1) to distinguish system-specific RUL modeling from learning reusable representations jointly across physically heterogeneous systems. Section 3.1 now clarifies the system type–dataset–unit–observation hierarchy. Figure 1(a) illustrates how measurements from an observation are converted into local and global descriptors, and Section 3.2 and Figure 2 detail their encoding. Section 3.3 and Figure 1(b) explain that “structure” means relationships among local descriptors within one observation: ISM learns these through masked reconstruction, whereas IDM predicts the current embedding from an unmasked history of observations of the same unit. Table 8 provides complementary evidence of IDM’s contribution.

2. Methodological details and architectural justification.

We have reorganized Section 3 and redesigned Figures 1–2 to distinguish type-specific stems and adapters from shared channel pooling, Transformer encoding, and projection; Appendix B provides implementation details. More importantly, we have added architectural ablations (Section 5.4, Table 7, Appendix F.1) evaluating adapters, type-specific versus shared stems, channel pooling, and the global token, alongside a Head-only comparison. Removing adapters increases RMSE for all three represented system types, whereas sharing stems has mixed effects, making the roles and tradeoffs of these components clearer. We have also added an IDM history-length sensitivity study comparing H = 3, 6, and 12 (Appendix F.3).

3. Purpose of the two evaluation levels.

We agree that their distinct roles in assessing degradation representations needed clarification. Sections 3.1, 3.4, and 4.2, Figure 1(c), and Table 2 now make the transfer boundaries explicit. Level I tests reuse on held-out datasets of system types represented in pretraining; Frozen adaptation assesses whether the learned representation network is useful without parameter updates. Level II tests whether the shared backbone provides a useful initialization for an entirely excluded system type, with newly initialized type-specific modules. RUL prediction evaluates downstream utility, while the frozen-feature lifecycle retrieval and visualization in Section 5.5 provide complementary representation-level evidence. We have also added three leave-one-system-type-out experiments, excluding bearings, batteries, and cutting tools in turn; together with N-CMAPSS, Level II now evaluates four unseen target types (Table 4, Appendix D).

4. Related work and stronger RUL baselines.

We agree that RUL-specific methods were insufficiently represented. We have expanded Section 2 to include graph-based RUL prediction and, following the reviewer’s specific suggestion, implemented Regression Shapalet–GNN (Yang et al., 2022) for the bearing task. We additionally evaluated Battery Transformer, TCN-BiLSTM-Attention, and ModernTCN. Sections 4.3 and 5.3, Table 5, and Appendix E report these comparisons, preserving each baseline’s native input processing and architecture while aligning target splits, label budgets, seeds, and metrics. CORD (Multi-domain) has the best average RMSE rank (1.33) among methods evaluated on all nine Level-I settings; we also report where RUL-specific methods perform better. We further added paired-seed confidence intervals, exact Wilcoxon tests, and Holm correction for the primary Level-I comparisons (Section 4.4, Appendix G).

5. Abbreviations and dataset identifiers.

We have defined Prognostics and Health Management (PHM) and Remaining Useful Life (RUL) at first mention in Section 1, and partial fine-tuning (Partial FT) and full fine-tuning (Full FT) in Section 3.4. Table 2 and Appendix A clarify device identifiers and adaptation/test splits. For example, Bearing2_2–Bearing2_5 are four XJTU-SY bearings under operating condition 2 used for downstream adaptation, while Bearing2_1 is held out for testing. Corresponding identifiers and splits are also documented for batteries, cutting tools, and engines.

We appreciate the reviewer’s suggestions, which helped us make the methodology clearer and the evaluation more comprehensive.

Thanks to the reviewer for the reorganization and for adding more experimental results. The revision makes the paper clearer and easier to understand. However, the evaluation metrics mainly focus on RMSE and MAE, which are static. So, what is the dynamic performance of CORD compared to other baselines? Is it stable across all time steps? And does CORD perform better than other baselines at 10%, 50%, or 100% RUL?

AuthorsOpenReview ↗

Response to Reviewer Re3D: Additional Analysis of Dynamic RUL Prediction Performance

We thank the reviewer for this thoughtful follow-up and for acknowledging the improvements in the revised manuscript. We agree that aggregate RMSE and MAE alone do not fully characterize how prediction accuracy evolves throughout the degradation process. The reviewer's suggestion provides a valuable opportunity to examine CORD's performance at different RUL stages.

To address this concern, we have conducted an additional stage-wise RMSE analysis on three heterogeneous physical systems: bearings (XJTU-SY), batteries (CALCE CS2), and cutting tools (PHM2010). We divide the normalized true RUL into ten intervals and evaluate the prediction errors of CORD and the existing baselines under 10%, 20%, and 100% label budgets, using five random seeds. This provides a more comprehensive view of prediction performance throughout the evaluated lifecycle, including the regions around 10%, 50%, and 100% RUL. Due to space limitations in the main manuscript, we have included the complete results in Appendix E.4.

The results show that CORD remains highly competitive across different degradation stages, physical systems, and label budgets, although its relative advantage varies throughout the lifecycle. In particular, CORD achieves relatively low prediction errors in many low-RUL intervals, especially for bearings and cutting tools, where accurate predictions near the end of life are particularly relevant to maintenance planning. Compared with several baselines that exhibit pronounced error peaks at certain stages, CORD generally maintains a more moderate error profile. However, some baselines outperform CORD in specific intervals, particularly during intermediate or early degradation stages.

These differences can be understood in relation to the characteristics of the three physical systems. For bearings, vibration measurements exhibit local variability during degradation, making the relationship between observed signals and remaining life difficult to characterize consistently. This may contribute to the error fluctuations observed at intermediate RUL stages. For batteries, degradation evolves gradually over charge–discharge cycles, and early-life observations may contain less distinctive information about the eventual lifetime. This is consistent with the increasing errors toward higher RUL levels observed for several methods, although battery-specific models remain competitive in certain intervals. For cutting tools, wear progression and machining-signal variations are not necessarily uniform across successive cutting passes. This may explain why the relative performance changes across degradation stages, with CORD performing favorably in several low-RUL regions while TCN-BiLSTM-Attention achieves lower errors in some intermediate and high-RUL intervals. These interpretations are consistent with the characteristics of the respective datasets, although the analysis does not establish the individual causes of these differences.

We acknowledge that CORD does not consistently achieve the lowest error at every RUL stage, and its prediction errors also vary throughout the degradation process. The stage-wise results characterize how prediction accuracy changes with RUL rather than directly measuring the smoothness of predictions between consecutive time steps. Nevertheless, the additional evaluation demonstrates that CORD achieves competitive stage-wise accuracy across heterogeneous systems, complementing its favorable aggregate RMSE results. More importantly, CORD's central contribution is to learn transferable degradation representations across physically heterogeneous systems, rather than to guarantee uniformly optimal predictions at every individual lifecycle stage.

In future work, we will further investigate system-specific degradation characteristics and explore improvements in temporal representation learning to achieve more accurate and consistent RUL predictions across the entire degradation process.

We sincerely thank the reviewer again for encouraging this additional analysis. It provides a more complete and balanced understanding of CORD's prediction behavior and helps clarify both its strengths and the stage-dependent differences that are not apparent from aggregate metrics alone.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.