acceptodds
Under review as a conference paper at ICLR 2027

A Feature-Based Explanation Framework for Time-Dependent Outputs

Abstract

Feature-based explanations quantify features' influence on model predictions, but are primarily designed for scalar outputs. In many applications, however, outputs are functional or multivariate, such as time-dependent trajectories in demand forecasting. Consequently, existing approaches typically explain each output location independently, ignoring dependencies across the output components. We address this limitation by developing a unified framework for feature-based explanations of time-dependent outputs. Specifically, we generalize functional decomposition to Hilbert-valued prediction functions and extend an existing feature-based explanation framework to this setting. Our framework introduces kernel-based output representations that enable time-dependency-aware explanations at multiple levels of temporal granularity, including *time-specific*, *time-resolved*, and *time-aggregated*, while providing a unified view in which existing methods arise as special cases.

Primary area:
interpretability and explainable AI
Keywords:
Explainable Artificial Intelligence, XAI, Explainable AI, Interpretability, Functional Decomposition, Time dependent Outputs

Reviews

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

This paper extends the scalar operator-based feature explanation framework of Fumagalli et al. (2025) to Hilbert-valued (i.e., time-dependent or functional, outputs). The central new component is an output behavior operator BoutB_{out}, a temporal aggregation and smoothing mechanism parameterized by a kernel KK, which yields explanations at three levels of temporal granularity: time-specific, time-resolved, and time-aggregated. The paper also gives guidance on how to select kernels. The work addresses a real shortcoming of current practice: applying scalar explanations pointwise at each output location ignores dependencies across the output domain and provides no principled mechanism for aggregating feature effects over time. The framework is relevant to applications such as time-series forecasting and dose-response analysis.

Critical strengths
  • The synthetic recovery study (App. D.1) and the temporal localization experiment with its anti-causal and placebo controls (App. D.5) are carefully designed, and reproducibility is strong.
  • The kernel-to-problem mapping in Table 5, along with the recommendation to assign kernels per feature, has practical value.
  • The paper is clearly written and well organized.
  • Theorem 2 links Shapley-style XAI with time-dependent Sobol indices.
Critical weaknesses
  • Limited novelty. Relative to Fumagalli et al. (2025), the main addition is a new operator, BoutB_{out}, which is a kernel integral operator (Eq. 6) and mathematically a linear smoother over the output domain. The premise of Theorem 1, that νinp(i)(s)≥νinp(j)(s)ν_{inp}({i})(s) ≥ ν_{inp}({j})(s) for all s ∈\in T, is a strong pointwise-dominance condition that is rarely verifiable in practice.
  • The real-data experiments (SPY, IHEPC, NESO) are largely qualitative case studies based on a single instance or a median profile. The regime-shift narrative in the SPY analysis rests on (a) an interaction effect, precisely the component that the authors' own synthetic study finds hardest to recover; (b) a single trajectory with no confidence intervals; and (c) results that are sensitive to user-specified kernel parameters. I suggest measuring the divergence between kernel-before-interaction and post-hoc smoothing for the sensitivity/risk games on real data, reporting the magnitude of ranking changes and trajectory differences.
  • The claim of "matching traders' interpretation of market events" (Sec. 5.2) is not supported by evidence from traders, and the statement that hourly reserve adjustments "mis-target the structure of forecast variance" (Sec. 5.2, Fig. 4) is not supported by any downstream reserve simulation.
  • The Figure 9 caption says "two synthetic examples" but lists three (ICU, price pulse, 3-day medication) and contains a duplicated "and and"; the 1.47×/1.61× partial/pure ratios in Figure 3, col. 3 should be disambiguated.
  • The special-case equivalences would be more convincing if verified empirically, for instance by numerically comparing the identity-kernel implementation against the published TsSHAP/ShapTime outputs on a shared dataset.
Questions for authors

Please refer to the 'critical weakness'.

2 comments
AuthorsOpenReview ↗

Response to Reviewer kdTZ

We thank the reviewer for their detailed review. We appreciate that they find the paper clearly written, the synthetic and localization experiments carefully designed, and the kernel selection guide of practical value. We have uploaded a revised manuscript in which all changes are highlighted in blue, and we refer to the corresponding sections in our responses to any open points below.

W1: We thank the reviewer for these observations. Regarding novelty, we agree that the proposed output behavior operator is linear. However, the contribution lies not in this operator alone, but in the generality of the resulting construction. By introducing a Hilbert-valued functional decomposition and extending the unifying framework of Fumagalli et al. (2025), we establish a systematic framework for feature-based explanations of Hilbert-valued outputs, including time-dependent predictions.

Importantly, this connects two previously separate research directions: time-dependent XAI, which primarily explains individual time points or temporal summaries without explicitly accounting for cross-time dependencies, and functional sensitivity analysis, which incorporates such dependencies but focuses largely on global variance-based importance. Our framework extends dependency-aware explanations to local prediction, global risk, and global sensitivity, covering pure, partial, and full operators; individual, interaction, and joint effects; and different temporal granularities. This enables, for instance, cross-time-dependency-aware global risk attribution.

Rather than merely smoothing explanations, kernels emphasize different relationships between output locations, providing distinct, interpretable perspectives on feature influence. We provide guidance on when and how to select kernels, combining this flexibility with a broad class of previously unavailable dependency-aware explanations through a single, coherent mathematical construction.

We have revised the introduction and related work (Sec. 1 and 2) to emphasize these contributions and added Table 3 in Appendix B highlighting the explanation settings enabled beyond existing pointwise XAI and functional sensitivity approaches.

Regarding Theorem 1, we agree that pointwise dominance is a strong sufficient condition that may not hold in practice. However, it is not a requirement for the reliability or applicability of our explanations. Different kernels emphasize different temporal relationships and therefore answer different explanatory questions; differing rankings do not inherently indicate inconsistency or reduced reliability. The theorem establishes a sufficient condition for ranking preservation under temporal aggregation, which can be checked directly from computed time-resolved effects. Moreover, App. D.1 (Table 6) shows that rankings may remain stable even when the condition fails, as long as features are well separated in aggregate magnitude. Neither pointwise dominance nor ranking preservation is required to apply the framework, as clarified in Sec. 4.2 and App. D.1.

AuthorsOpenReview ↗

Additional Response to Reviewer kdTZ

W2: Our real-world experiments serve complementary purposes: the local SPY example deliberately explains a particular model prediction to illustrate how different kernels reveal its temporal feature effects, while other examples investigate global explanations. These case studies are complemented by synthetic recovery experiments and quantitative temporal-localization analysis (App. D.5). The SPY example is therefore not intended to establish a population-level pattern.

(a) Interaction reliability. Feature-based explanations characterize the fitted model rather than the underlying data-generating process. Different model classes may learn different interaction structures, so variability in interaction recovery across models does not necessarily indicate unreliable attribution for a particular fitted model. Moreover, the SPY interaction illustrates how model-specific interactions can be represented across time and is not the main finding of the example.

(b) Confidence intervals. For a fixed model and instance, the exact explanation is deterministic. Confidence intervals across observations would address population-level variability rather than uncertainty in the explanation of this particular instance. Numerical approximation error is a separate concern, investigated in our synthetic recovery experiments, and can be reduced by increasing the number of masking samples.

(c) Kernel sensitivity. Dependence on kernel choice is intentional: different kernels define different notions of temporal contribution and therefore answer different explanatory questions (Sec. 5.1, Sec. 6, Tab. 5). Our framework provides guidance for selecting kernels accordingly. Theorem 1 establishes sufficient conditions for preserving feature rankings across kernels, while App. D.1 (Table 6) illustrates when rankings remain stable or change.

Additional divergence measurements. With the extension of Theorem 1, we have clarified that kernel aggregation and linear interaction operators commute for prediction-, risk-, and sensitivity-based explanations. Consequently, applying the same fixed kernel before or after the interaction operator yields mathematically identical results. Additional divergence measurements would therefore only reflect numerical approximation error rather than a methodological difference. We have clarified this in Appendix D.6.

W3: The first statement was meant to refer to the established interpretation of scheduled macro announcements in the market microstructure literature, not to evidence collected from traders. We have rephrased it accordingly and added the corresponding references. The second statement describes a hypothetical use case motivating the sensitivity analysis; a downstream reserve simulation is outside the scope of this paper. We have reworded it to make its hypothetical nature explicit.

W4: Thank you, we have corrected the Figure 9 caption (now Figure 8 in revised manuscript) and clarified in the Figure 3 caption and corresponding main text that the two ratios refer to the identity kernel (1.47×) and the causal kernel (1.61×), respectively.

W5: The special-case relationships in Sec. 4.3 are mathematical equivalences of attribution definitions under matching coalition constructions, rather than claims that independently implemented pipelines produce identical numerical outputs. We have therefore added explicit derivations for TsSHAP and ShapTime in Appendix C, showing how their attribution formulas follow from the corresponding operator choices. These derivations establish the correspondence independently of any particular model or dataset. Comparing separate implementations would primarily introduce differences in surrogate modeling, coalition evaluation, and numerical approximation, rather than test the claimed mathematical relationship. Our identity-kernel experiments already illustrate the corresponding pointwise and aggregated explanation settings (e.g., TsSHAP in Fig. 3, row 1, col. 3).

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

The authors present a framework for explaining multi-step time series forecasting models. They extend a feature-based explanation framework to include hilbert-valued outputs. This allows thh feature contribution to be represented as hilbert curves and then separate the explanation into prediction, risk, or variability contributions separate from the time-dependency of the explanation. A kernel can then combine information across output times or aggregate it into an importance score. The authors evaluate this approach with synthetic effect-recovery experiments and standard benchmarks.

Critical strengths

S1. The choice of explanation target is helpful and explicit. Explaining a particular forecast, prediction error, or variation across inputs can often lead to different feature rankings, so bringing these choices into one framework gives researchers more control over specifying exactly what they want to explain. The covariance formulation also captures relationships across output times that individual formulations would miss.

S2. The experiments give a concrete way to see if the framework recovers known effects on synthetic and real-data examples.

Critical weaknesses

W1. The main limitation of this work is in it’s novelty or marginal contribution. The framework extends what Fumagalli et al [1] proposed and areas like functional-output sensitivity and covariance sensitivity already address substantial parts of the problem. There is some value in the synthesis here but the paper can do more to identify exactly what is enabled by this synthesis. For example, do researchers find the ability to disambiguate effects more meaningful in qualitative surveys? The authors can do more explain what parts are truly meaningful in their new framework that couldn’t have been reached before.

W2. Some of the sensitivity results also seem a little challenging to interpret. For example the calendar inputs on the electricity benchmark complicate the results. Season can be determined by month/day but marginal masking can end up evaluating combinations that wouldn’t make sense. Given the masking scheme I think the scores may be fine but need some more attention paid to how they can be interpreted and limit some of the practical conclusions that can be drawn.

[1] Fumagalli et al., 2025, Unifying Feature-Based Explanations with Functional ANOVA and Cooperative Game Theory

Questions for authors

Q1. Which parts of the construction do the authors consider novel compared to the explanation and functional-sensitivity frameworks that already exist?

Q2. Where does the risk calculation prevent the attribution rule and kernel from being applied in either order (apdx D)? The loss is nonlinear but it is evaluated before these two operations. With a fixed kernel and the given attribution rules I would expect agreement with filtered pointwise risk attributions. Could the authors identify the step that makes them different? I think this might affect the overall value.

4 comments
AuthorsOpenReview ↗

Response to Reviewer YQCT

We thank the reviewer for their thoughtful review. We appreciate that they find the explicit choice of explanation target helpful and recognize that the covariance formulation captures cross-time relationships missed by pointwise approaches. We have uploaded a revised manuscript in which all changes are highlighted in blue, and we refer to the corresponding sections in our responses to any open points below.

W1 & Q1: We thank the reviewer for this comment and agree that the novelty and implications of our framework could be stated more clearly.

Our main contribution is bringing temporal-dependency-aware explanations into general feature-based XAI. While existing XAI methods and unifying frameworks primarily address scalar outputs, extensions to time-dependent outputs, such as forecasting trajectories and survival curves, have largely focused on individual methods, particularly pointwise, prediction-based SHAP without explicitly accounting for dependencies between output times. Functional-output sensitivity analysis, by contrast, has long considered such dependencies through covariance-based measures, but primarily outside ML/XAI and for global variance-based importance.

Our framework connects these previously separate research directions. Inspired by functional sensitivity analysis, we extend the general operator framework of Fumagalli et al. to Hilbert-valued outputs, transferring the treatment of temporal dependencies to ML model explanations. This unifies established functional sensitivity measures within feature-based XAI and, importantly, enables dependency-aware explanations beyond variance-based sensitivity, including local prediction-based and global risk-based explanations, without separate specialized methods. Examples include dependency-aware global risk attributions (SAGE- and PFI-type, Sec. 5.2) and temporally localized event effects through causal kernels (Sec. 5.2, App. D.5), which neither existing pointwise XAI nor functional sensitivity measures provide on their own.

Beyond the construction, we provide guidance on when and how temporal dependencies should be considered, illustrating why pointwise explanations may be insufficient. Our framework distinguishes whether feature effects are concentrated at individual times, extend across related temporal phases, or contribute to the overall trajectory, across different explanation types. Existing pointwise XAI methods and functional sensitivity measures also emerge as special cases, clarifying their relationship within a common framework.

We agree that a practitioner study assessing the usefulness of these distinctions would be valuable, but consider it beyond the scope of this methodological paper and an important direction for future work. We have revised the introduction and related work to emphasize the novelty and explanatory capabilities of our framework, and added Table 3 in Appendix B to highlight the explanation settings enabled beyond existing XAI and functional sensitivity approaches.

W2: We thank the reviewer for this point. We agree that marginal masking can evaluate implausible calendar combinations, since season is determined by month and is_weekend by day_of_week. We therefore repeated all energy analyses with (i) grouped masking, treating {month, season} and {day_of_week, is_weekend} as single players, and (ii) conditional masking, drawing background days consistent with the kept features. Results are reported in a new section of Appendix D.4 (paragraph “Robustness to masking scheme”).

Under grouped masking, rankings are essentially unchanged (Kendall’s τ = 1.0 in 6 of 8 settings; the remaining two differ by a single swap of adjacent features), and the main conclusions hold: NESO variance is dominated by the calendar features, IHEPC by lag_daily_mean. Under conditional masking, the dominant drivers remain the same, while lower-ranked features may change order (τ between 0.47 and 1.0), as expected for an observational rather than interventional question. The temporal structure of the effects is preserved under all schemes. Accordingly, we now interpret the month × season interaction in the NESO network plots as a joint calendar effect reflecting their deterministic dependence, and have revised Sec. 5.2 and App. D.4.

AuthorsOpenReview ↗

Additional Response to Reviewer YQCT

Q2: We thank the reviewer for this observation. Our original formulation of Theorem 1 focused on prediction-based explanations, where the dominance condition has a direct interpretation: if one feature contributes more strongly to the predicted output at every time point, its ranking is preserved under any nonnegative kernel. This emphasis was also motivated by the linearity of prediction-based input behavior in the model output, which allows temporal reweighting before computing feature effects. Remark 1 was intended to highlight this distinction from risk- and sensitivity-based explanations, whose input behaviors are nonlinear in the model output.

We agree that kernel aggregation and attribution commute more generally. For any fixed kernel, aggregation is linear in coalition values and therefore commutes with our linear interaction operators, independently of the input behavior. Thus, aggregating risk attributions is equivalent to aggregating risk-based coalition values before attribution; the nonlinearity of the loss does not affect this equivalence, as it is evaluated beforehand. The same holds for sensitivity when retaining full pairwise covariance contributions, including cross-time terms. Kernel nonnegativity is required for ranking preservation, not commutativity.

We have therefore generalized Theorem 1 and its proof (App. A.1) to establish ranking preservation under sufficient dominance conditions for prediction, risk, and sensitivity. We have also revised Remark 1 to clarify that only prediction-based explanations permit temporal reweighting before computing the input behavior, enabling scalar-output explanation methods. Appendix D.6 now explicitly states the general commutativity, while Appendix D.1 provides additional simulations illustrating ranking preservation.

I appreciate the detailed responses and revisions. In particular the masking analyses address my concern about the calendar features in W2, and the revised interpretation of month/season as a joint effect makes sense.

On Q2, I appreciate the clarification, but that said, I think this strengthens my concern on novelty/W1. If risk-based aggregation is equivalent to applying the kernel to standard pointwise attributions, it’s not clear which of the explanations in the paper couldn’t be obtained by kernel-filtering existing pointwise methods, or from classical functional Sobol indices in the sensitivity case. If there is a concrete example where the framework gives something these can’t, that would help clarify the contribution. For now I’ll keep my score.

Thank you for this follow-up. The equivalence you highlight concerns the computation of risk-based explanations, whereas our contribution lies in introducing a general methodology for constructing and studying feature-based explanations of time-dependent outputs that has not previously been developed in XAI. We highlight three key advances over existing approaches.

First, the proposed alternative is itself novel. To the best of our knowledge, incorporating dependencies across output time points through kernels has not been introduced as an explanation methodology in XAI, neither as post-hoc filtering nor as part of an explanation method. Our framework introduces this construction and establishes its theoretical properties, including the equivalence raised by the reviewer and conditions for feature-ranking invariance.

Second, kernels enable explanations that answer different questions, rather than merely modifying existing attributions. For example, in our SPY experiment, the question is which features contribute to predicted volatility while respecting the temporal ordering of macroeconomic announcements. Our asymmetric causal kernel addresses this question, whereas conventional pointwise explanations do not explicitly account for this temporal structure. More generally, our framework provides a mathematical foundation and practical guidance for specifying kernels according to the intended explanation question.

Third, our framework enables explanation concepts beyond those addressed by existing temporal XAI methods or classical functional Sobol indices. In particular, global risk-based explanations quantify how features and their interactions contribute to predictive performance across time, rather than prediction variability. To the best of our knowledge, such explanations have not previously been studied for time-dependent outputs in XAI. Table 3 explicitly highlights these and other previously unaddressed explanation concepts enabled by our framework. Moreover, our operator-based construction enables efficient comparison across explanation concepts: pure, partial, and full effects can be obtained from the same aggregated set function without separately computing and filtering their pointwise counterparts.

Thus, our contribution is not simply a new computational route to existing explanations, but a general methodology that introduces new explanation concepts, establishes their theoretical foundations, and enables their question-driven construction and efficient comparison.

Reviewer_qfQhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends an operator framework for feature attribution to vector and function valued predictions. After masking features, it constructs a prediction, risk, or covariance game, applies a kernel over the output domain, and derives pure, Shapley, or full effects. This gives explanations at individual times, along a trajectory, or aggregated over time. The paper connects these choices to existing explanation methods and Sobol indices, then illustrates kernel selection with synthetic examples, intraday volatility, and electricity demand. The appendix includes recovery experiments and a quantitative study of event localization.

Critical strengths
  • The separation of input behavior, output representation, and interaction allocation is useful. It makes explicit that smoothing an attribution changes the question being answered, and brings local predictions and global importance into one notation. The examples of inappropriate periodic or symmetric kernels are particularly helpful for understanding that choice.
  1. The appendix makes a serious effort to support the examples. It reports model accuracy, chronological splits, masking sample sizes, recovery errors over 30 runs, and computational costs. The localization study includes both an anti-causal control and a placebo cutoff, which are more informative than visual comparisons alone
Critical weaknesses
  1. Claimed connections to existing methods are mathematically reliable. Theorem 2 and Appendix D.2 identify the constant-kernel double covariance integral with classical functional Sobol indices. However, the cited Alexanderian et al. formulation uses a covariance trace ratio, equivalently a ratio of integrated pointwise variances. This should be further clarified.

  2. The main empirical example does not establish the claimed temporal interpretation. Section 5.2 says that the causal kernel localizes the July 13, 2022 announcement effect without pre-event leakage. Yet Table 7 reports a pre-event share of 0.316 for the identity kernel, 0.329 for the matched causal kernel, and 0.327 for the nominal causal kernel on that very day. The aggregate improvement over other days is useful evidence, but it does not support the stronger claim made for Figure 3.

Questions for authors
  1. Which precise Sobol quantity is recovered in Theorem 2 and Figure 7: a trace-based functional index or an index for the integrated scalar output? Please reconcile the formula, the cited definition, and the analytical values in that figure.

  2. Please reconcile the announcement feature definition, the inclusion of July 13 in the intraday-release set, and the nonzero pre-event shares with the claims around Figure 3. Would the localization conclusion still hold with the evaluation restricted to correctly identified afternoon events and an explicit justification for treating pre-event attribution as an error?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qfQh

We thank the reviewer for their careful reading. We appreciate that they find the separation of input behavior, output representation, and interaction allocation useful, and value our appendix, including the localization study with its controls. We have uploaded a revised manuscript in which all changes are highlighted in blue, and we refer to the corresponding sections in our responses to any open points below.

W1 & Q1: We thank the reviewer for highlighting that the relationship between our covariance-based formulation and existing functional Sobol indices could be stated more explicitly. We have clarified this distinction in Section 3.1 of the revised manuscript.

Our framework generalizes scalar variance-based Sobol indices to time-dependent outputs by incorporating cross-time covariance contributions, motivated by Shi et al. (2018). Theorem 2 concerns the covariance-based time-specific and time-aggregated sensitivity measures defined in Section 3.1, recovered through our sensitivity-based input behavior, interaction operators, and normalization. It does not claim equivalence between the constant-kernel formulation and the trace-based functional Sobol indices of Alexanderian et al. (2020).

Specifically, the kernel choices yield different aggregations:

  • Constant kernel ((K(t,s)=1)): Integrates covariance contributions over all time pairs, including cross-time dependencies. After normalization, this yields our covariance-based Sobol indices (as defined in Sec. 3.1).

  • Identity kernel ((K=I)): Retains diagonal covariance contributions, corresponding to integrated pointwise variances or the trace of the covariance operator. After normalization, this recovers the trace-based functional Sobol indices of Alexanderian et al. (2020) and the identity-weighted indices of Gamboa et al. (2014).

Alexanderian et al. motivate their indices through covariance operators, but their trace-based definition does not explicitly incorporate off-diagonal covariance contributions, although the full covariance structure enters their Karhunen–Loève-based computation.

Regarding Figure 7, the comparison uses covariance-based indices under the constant kernel, not Alexanderian et al.'s trace-based indices. We have clarified this in the figure caption, in Appendix D.2, and in Theorem 2 accordingly. These changes do not affect the underlying construction or derivation.

W2 & Q2: We thank the reviewer for this observation. Our wording conflated two distinct statements, and we have revised it. The causal kernel K(t, s) = exp(−(t − s)/ℓ) 1[t ≥ s] averages the pointwise attributions only over past and current times s ≤ t. This has two consequences. First, the response after the 14:00 release cannot be shifted to earlier times, whereas symmetric kernels such as OU spread part of it backward across the bandwidth ℓ. Second, attribution that is already present before 14:00 in the pointwise effect is retained, not removed. Our claim is the first statement, not the absence of any pre-release attribution.

This explains the nonzero pre-event shares in Table 7. The share ρ measures the fraction of absolute attribution mass before 14:00, and the pre-release window covers 54 of the 78 bars, so even a small persistent offset in the morning accumulates to a substantial share, as seen for July 13th. Moreover, ann_indicator is a day-level feature pooling morning and afternoon releases, so part of the observed morning attribution reflects the model’s general response to announcement days and is not necessarily spurious.

The evidence for temporal localization therefore rests on the relative comparisons in App. D.5 rather than on absolute shares for a single day: the reduction in ρ relative to the pointwise baseline, the excess over a placebo cutoff, the anti-causal control, and the control feature vix_prev, for which the causal kernel increases ρ. We have revised Sec. 5.2 and App. D.3 accordingly; Fig. 3 is intended to serve as an illustration of how kernel choice shapes the attribution profile, while App. D.5 provides the quantitative evidence.

Comments on the paper

1 comment

As the rebuttal period is coming to an end, we would like to sincerely thank the Area Chair and all reviewers for their time, careful evaluation, and constructive feedback. We hope that our responses and revisions to the manuscript have adequately addressed the remaining questions and concerns. We would greatly appreciate it if the Area Chair and reviewers could take our rebuttal and revised manuscript into consideration when reassessing our work and reaching their final evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.