acceptodds
Under review as a conference paper at ICLR 2027

RExU: Recovery-Expert Unlearning for Mixture-of-Experts

Abstract

Machine unlearning aims to remove target knowledge while preserving other capabilities. Existing evaluations measure forget efficacy (FE) under normal inference, treating low FE as evidence of successful forgetting. This assumption can fail for Mixture-of-Experts (MoE) models, where each input activates only a sparse subset of experts: an unlearning method may suppress target knowledge along the natural route - the model's default expert selection without inference-time intervention - while leaving the same knowledge recoverable through alternative experts. We expose this off-route recovery through counterfactual routing and introduce Worst-Route Forget Efficacy (WRFE), which measures the maximum forget-set recovery over an audited route family including the natural route. Statistically above-chance recovery under such rerouting provides direct evidence that supposedly forgotten knowledge remains accessible in the model. Empirically, we find substantial off-route recovery despite successful natural-route forgetting. On WMDP-Cyber, for example, GA+SEUF lowers FE by percentage points, from to , under natural routing, yet bypassing the edited expert raises FE by percentage points to . Thus, natural-route FE alone can substantially underestimate residual recoverability in MoE models. To mitigate this failure, we propose Recovery-Expert Unlearning (RExU), which iteratively identifies recovery experts contributing most to worst-route recovery and selectively unlearns them. Across three MoE models, four base unlearners, and two benchmarks, RExU substantially reduces WRFE while maintaining competitive natural-route FE and retain utility. Our results motivate jointly evaluating MoE unlearning through natural-route FE, WRFE, and retain utility, and establish RExU as a practical method for improving this trade-off.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Mixture-of-Experts

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

The paper addresses an important question in Machine Unlearning (MUL) community, how to apply MUL in mixture-of-experts model. The main motivation behind the work is accounting for routing during MUL, since some of the experts (e.g. activations) may still have information about the data to be forgotten. To account for this, the authors propose a new iterative unlearning algorithm which recursively finds highest-recovery route in MOE (with respect to the forget set), and updating the experts in this family. They provide strong empirical evidence for the iterative unlearning procedure proposed: against the SEUF baseline, RExU (the proposed algorithm by the authors) provides better forget-utility Pareto front. This is further strengthened by reasonable ablation studies. The closest prior work is SEUF, which also proposes to select experts for unlearning using an affinity-based localization. In addition, RExU investigates alternative routes in MOE and updates the weights for these that may still carry information. Theoretical novelty is modest and primarily explanatory, so most of the novelty relies on empirical gains.

Critical strengths

Strengths

  1. The empirical gains over prior work (SEUF) are promising, where RExU provides a better utility-forget trade-off. This is with respect to the range of MUL algorithms authors test, including gradient ascent and negative preference optimization. The promising empirical result is further displayed in paired analysis, where masking recovers 190 of 230 pretrained predictions that became wrong under SEUF’s route.
  2. The evaluation is bnroad with respect to unlearning objectives and models. The authors cover multiple MOE models, and use both multiple-choice and generation tasks, including low-rank adaptation to show the trade-off between MUL algorithms (in terms of forgetting efficiency and broader utility).
Critical weaknesses

Weaknesses:

  1. On the empirical side, while the experimental setup is broad (multiple MoE and tasks), the machine unlearning techniques compared against do not have strong baselines from classification literature, and NLP. For example, algorithms like SCRUB/NegGrad from Kurmanji et al., for cases where the retain set is available. This would broaden the scope of the paper, and show that, for the SOTA algorithms, the improvements still stand even if replicated in a reduced setting for MoE.
  2. The theoretical argument is mostly explanatory and does not have sufficient depth. For example, theorem 6 does not prove a convergence guarantee for the implemented algorithm. Theorem 3 has strong assumptions (Lipschitz) which may not hold true in practice, therefore highlighting a disconnect.
Questions for authors
  1. Do the improvement translate to other machine unlearning algorithms that show SOTA improvements beyond MoE? Gradient Ascent and preference optimisation are too narrow to qualify the claims, and it is well known that gradient ascent will induce overforgetting.
  2. Can the theoretical claims be adjusted to account for the strong assumptions, and to show convergence of the recursive operator proposed?
4 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer aHfX [1/3]

Thank you for your constructive feedback! We agree that stronger retain-aware unlearning baselines and careful qualification of the theory would strengthen the paper. Following your suggestions, we added a SCRUB experiment (Appendix F.4, Table 14) and clarified what Theorem 6 does-and does not-guarantee (Appendix C.8). We address both concerns below.

AuthorsOpenReview ↗

Official Author Responses to Reviewer aHfX [2/3]

W1: On the empirical side, while the experimental setup is broad (multiple MoE and tasks), the machine unlearning techniques compared against do not have strong baselines from classification literature and NLP. For example, algorithms like SCRUB/NegGrad from Kurmanji et al., for cases where the retain set is available. This would broaden the scope of the paper, and show that, for the SOTA algorithms, the improvements still stand even if replicated in a reduced setting for MoE.

Q1: Do the improvement translate to other machine unlearning algorithms that show SOTA improvements beyond MoE? Gradient Ascent and preference optimisation are too narrow to qualify the claims, and it is well known that gradient ascent will induce overforgetting.

Thank you for suggesting SCRUB and NegGrad. We agree that GA alone can overforget and that comparisons with stronger retain-aware methods are important. We respectfully disagree only that our original evaluation was limited to GA and preference optimization: as already reported in §6.1 (Lines 420–426), it also covers retain-aware GDIFF and representation-level RMU alongside GA and NPO, with MoE-specific GRIP[R1] and TRACE[R2] comparisons in Table 10. NegGrad and NegGrad+ share objective-level principles with GA and GDIFF, respectively, but we do not treat this as an evaluation of their exact implementations.

Following your suggestion, we added the retain-aware teacher–student SCRUB baseline on DeepSeek-V2-Lite/WMDP-Cyber (Appendix F.4, Table 14, Lines 3080–3091):

MethodNatural FE ↓WRFE ↓MMLU ↑
SCRUB28.18%30.35%50.68%
SCRUB+SEUF28.23%35.88%48.60%
SCRUB+RExU28.49%30.20%54.10%

Compared with SCRUB+SEUF, SCRUB+RExU reduces WRFE by 5.69 percentage points and improves MMLU by 5.50 points. Relative to standalone SCRUB, it attains comparable WRFE (30.20% vs. 30.35%) with +3.42 MMLU points. This supports RExU's benefit for an additional retain-aware objective in the tested setting, without claiming universal superiority across unlearning methods.

References

[R1] GRIP: Algorithm-Agnostic Machine Unlearning for Mixture-of-Experts via Geometric Router Constraints. arXiv:2601.16905 (2026)
[R2] Routing-Aware Expert Calibration for Machine Unlearning in Mixture-of-Experts Language Models. arXiv:2606.10338
AuthorsOpenReview ↗

Official Author Responses to Reviewer aHfX [3/3]

W2: The theoretical argument is mostly explanatory and does not have sufficient depth. For example, theorem 6 does not prove a convergence guarantee for the implemented algorithm. Theorem 3 has strong assumptions (Lipschitz) which may not hold true in practice, therefore highlighting a disconnect.

Q2: Can the theoretical claims be adjusted to account for the strong assumptions, and to show convergence of the recursive operator proposed?

Thank you for raising these concerns. We agree that Theorem 3 relies on a restrictive assumption and that Theorem 6 does not prove convergence. We would, however, distinguish the conditional results from our existing structural guarantees: as established in Appendices C.6–C.7, Lemma 5 shows exact invariance when a routed computation avoids all edited experts under fixed router/shared parameters, and Theorems 4–5 give optimization-independent lower bounds on residual recovery and necessary edit budgets. Neither requires Lipschitz continuity or convergence of RExU.

As already stated in Appendix C.5, Theorem 3 uses local Lipschitz continuity only as a sufficient condition for approximate redundancy; the existing exact-copy special case (Lines 1512–1524) preserves the target margin without a quantitative Lipschitz bound. We do not claim the assumption holds universally in practical MoEs.

Theorem 6 gives a conditional one-round descent bound, not an unconditional convergence guarantee. Progress requires the targeted recovery reduction to exceed cross-route shifts and finite-audit error. As already illustrated in Appendix I (Figure 12), WRFE rises from 27.78% at round 5 to 36.03% at round 7 as recovery shifts between routes, showing why monotonic progress cannot be assumed. Following your feedback, we clarified in Appendix C.8 (Lines 1935–1951) that Theorem 6 does not establish parameter convergence or attainment of the forgetting threshold, and that the differentiable-route argument does not automatically cover hard top-k boundary crossings. Establishing convergence for the implemented recursive operator remains open.

Dear Reviewer aHfX,

As the discussion period nears its end, we sincerely thank you for your thoughtful review and valuable suggestions. Your feedback motivated us to extend our evaluation with SCRUB, a stronger retain-aware unlearning baseline, and clarify the assumptions and guarantees of our theoretical analysis.

We have incorporated these additions into the revised manuscript. If you have a moment, we would greatly appreciate hearing whether our new results and theoretical clarifications address your concerns.

Regardless, we sincerely appreciate your time and constructive feedback, which have helped improve both the empirical and theoretical presentation of our work.

Best regards,

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

This paper highlights that MoE models might appear to forget under its natural router meanwhile leaving target knowledge concealed in other experts. Then it introduces a counterfactual routing audit and a metric WRFE to fully examine the concealed knowledge in other experts. Accordingly, they propose RExU, which aims to remove knowledge from MoE models by repeatedly identify and audit any related knowledge across experts. Extensive experiments across 3 MoE scales, 4 base unlearners and 2 benchmarks demonstrate the completeness of its unlearning.

Critical strengths
  1. The highlighted off-route recovery is important for MoE models, and the performed routing interventions provide convincing evidence that good natural-route forgetting could hide substantial recoverability.
  2. The paper is well-organized and clearly written, with intuitive demonstration of its idea and findings.
  3. The proposed RExU is comprehensively and convincingly evaluated across various benchmarks and settings.
Critical weaknesses

My main concern is that it seems that the method and evaluator are tightly coupled. WRFE is the maximum recovery over a hand-designed routing family, while RExU directly optimizes against routes from essentially the same family. The results therefore show robustness to the specified audit family, but not necessarily general route-robust deletion. Could the authors further clarify or demonstrate such adversarial robustness by holding out an entire intervention mechanism during RExU training and test whether robustness transfers out of family? I would be happy to raise my score if this is well clarified.

Questions for authors

Although two routes are held out in RExU, it seems that they still belong to the same family of router-logit transformations used in RExU. How can we distinguish genuine removal of the underlying knowledge from adaptation to this particular audit family? Could the authors provide more complete investigation from this perspective? Otherwise it might be unable to claim a generalized or truly complete removal.

6 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer MMer [1/5]

We appreciate the recognition of the significance of off-route recovery and the comprehensive scope of our evaluation. We emphasize that our primary contributions are the identification and characterization of off-route recovery as a failure mode of MoE unlearning, as well as the introduction of WRFE as a systematic evaluation framework. RExU serves as our initial recovery-guided approach to mitigate this failure mode, rather than asserting universally complete knowledge deletion.

To address your concern about training–evaluation coupling, we highlight the existing mechanism-level holdout (Table 12) and provide additional evaluations using strictly held-out routing interventions and an alternative expert-dispatch width (Appendix F). We also clarify the scope of our robustness claims. Please find our detailed responses below.

AuthorsOpenReview ↗

Official Author Responses to Reviewer MMer [2/5]

W1.1: My main concern is that it seems that the method and evaluator are tightly coupled. WRFE is the maximum recovery over a hand-designed routing family, while RExU directly optimizes against routes from essentially the same family…

We agree that overlap between training and evaluation routes can lead to evaluator-specific adaptation. The overlap is intentional in RExU's constraint-generation formulation, but this alone does not rule out such adaptation. As already described in §5.2, RExU iteratively identifies the most violated recovery constraint, edits the responsible experts, and re-audits, since suppressing one recovery route may reveal another. This interaction is intrinsic to the method's objective and is analogous to adversarial training.

Importantly, the audit family is motivated by distinct failure modes rather than being tailored to RExU. A1 bypasses edited experts, A2 perturbs gate preferences, and A3 activates rarely selected fallback experts (§3.1). As already shown in Table 1, the most effective intervention varies by unlearning method: A1 reveals the most recovery for SEUF, while A2 is most effective for NPO and RMU.

As already specified in §3.1 (Lines 194–200), §5.1 (Lines 320–346), and Appendix D (Lines 2125–2146), optimization and reporting are explicitly separated. RExU is trained using four routes, whereas the final WRFE evaluation employs six routes, incorporating stronger gate noise and rare-expert forcing not present in the training-route selection. Therefore, the suitable test of evaluator coupling is whether the observed improvement transfers to previously unseen constraints, which we examine in the responses below.

AuthorsOpenReview ↗

Official Author Responses to Reviewer MMer [3/5]

W1.2: The results therefore show robustness to the specified audit family, but not necessarily general route-robust deletion. Could the authors further clarify or demonstrate such adversarial robustness by holding out an entire intervention mechanism during RExU training and test whether robustness transfers out of family? I would be happy to raise my score if this is well clarified.

We acknowledge that no finite audit can establish universal route-robust deletion. However, our results demonstrate robustness that extends beyond the routes used during optimization.

To address your concern, we added a more stringent held-out audit in Appendix F.5 (Table 15, Lines 3095–3109), covering extended expert masking, increased gate noise, rare-expert forcing, and a composite masking plus noise intervention. These interventions are excluded from route selection, expert attribution, parameter updates, and checkpoint selection.

On DeepSeek-V2-Lite/WMDP-Cyber, GA+RExU increases from 27.43% natural-route forget accuracy to a maximum of 28.54% under these held-out interventions (+1.11 percentage points). Notably, these probes can reveal substantial recovery: GA+SEUF increases from 26.57% to 37.90% under masking and to 34.63% under the composite intervention.

Held-out evaluation results (Table 15):

MethodHeld-out evaluationNatural FE ↓Stress-test FE ↓Increase
GA+SEUFExtended expert masking26.57%37.90%+11.32 pp
GA+SEUFComposite masking + noise26.57%34.63%+8.05 pp
GA+RExUMaximum across four held-out interventions27.43%28.54%+1.11 pp

As already reported in Appendix H, broader routing-search controls provide another comparison. For GA+SEUF, the best label-informed gradient search achieves 29.99%, and the best of 500 random hard routings achieves 27.63%, both of which are below the 37.90% recovery revealed by expert masking (Appendix H). This finding suggests that the failure-mode-driven audit is more effective than generic routing searches in identifying residual recovery.

Collectively, these results support transfer beyond the specific optimization audit, while universal route robustness remains an open question (See Response for Reviewer Gws7 W1).

Reviewer_Gws7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper identifies a limitation of existing unlearning methods for Mixture-of-Experts (MoE) models, where knowledge that appears to be forgotten under natural expert routing can be recovered by modifying the routing decisions. The authors demonstrate that existing methods, particularly SEUF, can achieve low forget-set accuracy under natural routing, while masking the edited experts allows the model to recover performance comparable to the pretrained model.

To evaluate this, the authors propose Worst-Route Forget Efficacy (WRFE), which measures the maximum forget-set accuracy across a predefined family of routing interventions. These interventions include expert masking, perturbations to router logits, and forcing rarely activated experts. To address this, they further propose Recovery-Expert Unlearning (RExU), an iterative unlearning framework. At each iteration, RExU identifies the routing configuration that produces the highest forget-set accuracy. It then identifies experts contributing most to this recovery by measuring the change in correct-answer margins when individual experts are ablated. The selected experts are subsequently unlearned while the router and other model parameters remain frozen. This procedure is repeated to account for recovery through alternative experts after each round of unlearning.

The authors evaluate RExU across three MoE models, two unlearning benchmarks, and four base unlearning methods. They demonstrate that RExU generally improves worst-route forgetting while preserving retain utility compared to SEUF.

Critical strengths
  • The paper identifies an important limitation of existing MoE unlearning methods, demonstrating that knowledge that appears forgotten under natural routing can remain recoverable through alternative experts.
  • To mitigate this, the paper proposes a novel WRFE metric, which provides a useful way to evaluate the robustness of MoE unlearning against routing interventions.
  • In addition to finding the limitation, they propose a potential sound solution to the problem. Critically, their method performs better than SEUF with respect to WRFE , while also preserving utility.
  • There are extensive experiments which establish the limitation and the effectiveness of their method.
Critical weaknesses
  • The authors emphasize : "Natural-route forgetting can therefore reflect path suppression rather than knowledge removal". However, there is no evidence or mechanism which makes RExU a knowledge removal algorithm. I strongly suggest the authors to tone down these parts or provide evidence that this is truly knowledge removal. Note that removal vs suppression is an open issue for current unlearning methods. So since the base methods (NPO, RMU , GA) are classic unlearning methods, knowledge removal is not expected.

The following is not a critical weaknesses in the paper, but an open question or gap:

  • Six routing configuration are used. There are potentially much more configurations and we need to understand what is the one which gives the worst-case intervention scenario. This is potentially an open question.
Questions for authors

Please see weaknesses.

5 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer Gws7 [1/4]

We appreciate your positive assessment of the paper’s soundness and empirical contribution, and for highlighting two important scope questions - one concerning how we characterize knowledge removal, and the other, as you note, concerning an open question about the broader routing space.

To address your concerns, we have clarified the distinction between reduced recoverability and complete knowledge removal, added an independent hidden-state probing analysis (Appendix F.7), and further highlighted our findings on method-dependent worst-route interventions and broader routing searches. Please find our detailed responses below.

AuthorsOpenReview ↗

Official Author Responses to Reviewer Gws7 [2/4]

W1: The authors emphasize : "Natural-route forgetting can therefore reflect path suppression rather than knowledge removal". However, there is no evidence or mechanism which makes RExU a knowledge removal algorithm. I strongly suggest the authors to tone down these parts or provide evidence that this is truly knowledge removal. Note that removal vs suppression is an open issue for current unlearning methods. So since the base methods (NPO, RMU , GA) are classic unlearning methods, knowledge removal is not expected.

We agree with the distinction, and we would like to clarify that we do not claim RExU certifies complete knowledge removal. As already stated in the Conclusion (§7, Lines 501–503): “WRFE only measures recovery within a finite audited route family, so low WRFE does not prove total knowledge removal. RExU reduces but does not eliminate recoverability.” Following your feedback, we clarified this one-sided interpretation beside the main finding in the revised Introduction (§1, Research Gap, Lines 076–081). The finite-audit scope is also described in §3.1 (Lines 157–160) and Appendix K (Lines 3566–3581).

The statement “Natural-route forgetting can therefore reflect path suppression rather than knowledge removal” is intended as a one-sided diagnostic of natural-route evaluation. If the same fixed checkpoint and input recover the correct answer after only the route is changed, the target information remains accessible despite low natural-route accuracy. This does not imply that failure to recover constitutes evidence of erasure.

As already shown in the item-level analysis in Table 6 (Appendix F.2.1, Lines 2324–2344), among pretrained-correct answers missed under the edited model’s natural route, expert masking recovers 190/230 (82.6%) after GA+SEUF but only 35/210 (16.7%) after GA+RExU. Therefore, RExU substantially reduces routing-based recoverability, while the 35 recovered answers indicate that some residual accessibility persists.

To further address your concern, we added a frozen hidden-state probe in Appendix F.7 (Table 17, Lines 3149–3160) to examine whether this reduction is specific to routing. This probe maintains natural routing and measures answer accuracy using an independent readout of the model’s internal representations. This approach provides an orthogonal test of residual accessibility beyond the routing audit. For GA+SEUF, the probe increases WMDP-Cyber accuracy from 26.57% to 32.71%, a net increase of 122 correct predictions (+6.14 pp). For GA+RExU, accuracy increases from 27.43% to 28.94%, a net increase of 30 correct predictions (+1.51 pp). These are net accuracy gains, not counts of native-wrong items individually recovered by the probe. The result supports reduced accessibility under this readout, although residual accessibility remains after RExU.

Frozen-readout results (Table 17):

MethodNatural-output accuracyFrozen-probe accuracyChange
GA+SEUF26.57%32.71%+6.14 pp
GA+RExU27.43%28.94%+1.51 pp

The revised Introduction (§1, Research Gap) now makes the same scope explicit as the existing Conclusion (§7): RExU substantially reduces recoverability under the audited routing interventions, and our additional probe tests whether this reduction extends to an independent readout; neither result is a certificate of universal knowledge removal.

AuthorsOpenReview ↗

Official Author Responses to Reviewer Gws7 [3/4]

MW1.1: Six routing configuration are used. There are potentially much more configurations…

We agree that six routing configurations do not exhaust the routing space. As already reported in Appendix H (Figure 11), we examined broader searches to see whether greater recovery could be found. For the representative GA+SEUF/DeepSeek-V2-Lite/WMDP-Cyber checkpoint, expert masking increases forget-set accuracy from 26.57% to 37.90%. A label-informed gradient search over routing perturbations achieves 29.99% across three restarts, while the best result among 500 random hard routing configurations is 27.63% (Appendix H, Figure 11). Neither broader search surpasses the recovery achieved by A1 masking.

In this representative failure case, the structurally motivated intervention identifies a stronger recovery path than substantially broader gradient-based and combinatorial searches (Appendix H, Figure 11, Lines 3313–3352). It is not claimed that 37.90% represents the global maximum across all possible routing configurations; rather, it is the strongest recovery identified in this study, and as the reviewer notes, determining the global worst intervention remains an open question.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Official Author Global Responses to AC & All Reviewers

Dear AC and reviewers,

Thank you all for your careful reading and constructive feedback! In brief, RExU studies off-route recovery in MoE unlearning: a model can appear to forget under natural routing while retaining recoverable target answers through alternative experts. We introduce WRFE to measure the largest observed recovery within a finite routing audit, and RExU to iteratively identify and edit recovery-contributing experts.

Based on your feedback, we addressed the following points in our rebuttal and manuscript revisions:

  1. Following your feedback, we clarified our claims in §1 (Lines 076-081): reduced audited recoverability is not certified knowledge erasure, consistent with the existing limitation in §7 (Lines 501-503) and Appendix K. (Gws7, MMer.)
  2. We highlight existing item-level evidence: as already shown in Appendix F.2.1 (Table 6), masking restores 190/230 (82.6%) eligible answers after GA+SEUF, versus 35/210 (16.7%) after GA+RExU, with paired statistical analysis. (Gws7.)
  3. We added robustness tests beyond the optimization routes: four held-out interventions and a top-6 → top-8 dispatch change; the latter increases FE by +4.63 pp for GA+SEUF but only +0.91 pp for GA+RExU. (MMer; Appendix F.5-F.6, Tables 15-16.)
  4. We added a complementary frozen hidden-state readout: its accuracy increase is +6.14 pp after GA+SEUF versus +1.51 pp after GA+RExU, while still detecting residual accessibility. (Gws7, MMer; Appendix F.7, Table 17.)
  5. We added SCRUB: SCRUB+RExU achieves 30.20% WRFE↓ / 54.10% MMLU↑, compared with 35.88% / 48.60% for SCRUB+SEUF. (aHfX; Appendix F.4, Table 14.)
  6. Following your feedback, we clarified Theorem 6's scope in Appendix C.8: its conditional one-round descent differs from the existing structural lower bounds (Appendices C.6-C.7). We also refer to the existing route-search controls (Appendix H, Figure 11), which do not claim a global worst case, and the recovery dynamics in Appendix I (Figure 12). (aHfX, Gws7.)

We hope the new evidence and clarifications address your concerns. Thank you again for helping us improve the paper!

Dear reviewers,

Please check your assigned papers for author responses. If the authors have addressed any of your questions or concerns, please leave a brief comment for them indicating which points have been resolved and which, if any, remain open. Even if your assessment is unchanged, a short follow-up is helpful.

Best, AC

Dear Reviewers,

The Author-Reviewer Discussion period is coming to an end soon. Please actively participate in the rebuttal discussion with the authors after carefully reading the authors' responses.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.