acceptodds
Under review as a conference paper at ICLR 2027

Making LLMs Truly Forget: Deep Unlearning by Searching, Selecting, and Severing Knowledge Paths

Abstract

While an unlearned language model may no longer recall a fact directly, the fact often remains recoverable through multi-hop reasoning over related knowledge. Most existing unlearning techniques overlook this vulnerability, targeting facts in isolation while leaving their supporting knowledge intact. To achieve true forgetting, we propose a general deep unlearning framework compatible with existing unlearning algorithms. Our approach adaptively explores both explicit responses and latent internal representations to discover valid reasoning paths, compiles them into a confidence-aware supporting subgraph, and we apply a graph minimum cut to sever all recovery paths while preserving unrelated knowledge. To rigorously evaluate deep unlearning, we introduce a model-specific pipeline that extracts and completes knowledge graphs from raw text, filtering them by calibrated model confidence to reflect what the model genuinely retains. Comprehensive experiments demonstrate that selectively unlearning supporting knowledge yields substantially deeper forgetting than superficial methods while preserving model utility, highlighting that genuine unlearning requires breaking the relational structures that enable factual reconstruction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deep Unlearning, Machine Unlearning, Large Language Models

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

This research proposes an unlearning method that addresses a limitation of previous approaches: directly unlearning a target fact may still leave it recoverable through multi-hop reasoning over related knowledge. The proposed approach examines both explicit model responses and latent representations to identify recovery paths and selectively unlearn supporting facts that interrupt these paths. Experimental results demonstrate more effective forgetting than the baselines while limiting the loss of unrelated knowledge.

Critical strengths
  • The research is clearly motivated, addressing superficial unlearning by expanding the search to include multi-hop reasoning paths that can reconstruct forgotten knowledge.
  • The confidence-aware minimum-cut approach appears sound and practical, although its effectiveness remains limited by the coverage of the discovered supporting knowledge.
  • The experimental results support the effectiveness of the proposed method, showing substantial improvements over the baselines.
Critical weaknesses
  • One of my major concerns is the limited scope of the experimental setting. The evaluation relies on a single fictional domain with relatively bounded knowledge, which may make the relevant information easier to identify and control than in real-world settings. This could partly explain the substantial performance gains. Experiments on more diverse, real-world knowledge would help establish the method’s generalizability and practical effectiveness.
  • The paper also omits a closely related work: Yang et al., “FAITHUN: Toward Faithful Forgetting in Language Models by Investigating the Interconnectedness of Knowledge” (EMNLP 2025). The authors should discuss its relationship to their work, particularly its treatment of superficial unlearning and interconnected knowledge. FAITHUN also provides a relevant real-world evaluation setting that could strengthen the experimental validation.
Questions for authors

.

4 comments

We thank Reviewer hpEu for the constructive feedback and for recognizing the clear motivation of our work, the soundness and practicality of the confidence-aware minimum cut, and the effectiveness shown in our experiments.

[W1. Limited experimental scope] Thank you for this constructive suggestion. Our choice of Harry Potter is motivated by the need for rigorous evaluation. Access to the complete Harry Potter original texts allows us to construct a comprehensive reference from raw text as an approximate upper bound on the model’s knowledge within this domain. Our model-specific filtering then aligns this reference with the knowledge actually retained by each evaluated model, minimizing omissions of relevant supporting knowledge. The same procedure can be applied to other domains, especially when model holders have access to their training corpora. We focus on Harry Potter because, for real-world knowledge domains such as Wikipedia, the evaluated models’ complete training corpora are unavailable, making it difficult to construct a comparable upper bound. Evaluations in these domains remain useful, but missing supporting knowledge introduces uncertainty into path discovery and evaluation.

To examine generalization beyond the fictional Harry Potter domain, we additionally conducted experiments on YAGO3-10, a knowledge graph dataset containing real-world facts about entities and their relationships. Using Qwen3.5-4B with LoRA fine-tuning, we randomly selected 100 target triples for evaluation.

Search statisticResult
Target triples100
Discovered supporting paths306
Target triples covered by discovered paths61/100
Model / MethodBase Qwen3.5-4BGANPO
Accuracy ↓0.420.290.23

These results demonstrate that our framework can discover supporting paths and perform unlearning beyond Harry Potter, with target accuracy decreasing from 0.42 to 0.29 with GA and 0.23 with NPO.

[W2. Relationship to FAITHUN] Thank you for pointing out this relevant work. We will add a discussion of FAITHUN [1] and clarify its relationship to our paper. FAITHUN tests whether forgetting extends to related questions, including multi-hop questions whose reasoning requires the target fact. Our work examines whether alternative paths composed of other retained facts can reconstruct the target fact itself. We therefore develop a framework to discover and verify these paths, select facts to unlearn through a confidence-aware minimum cut, and evaluate the remaining paths after unlearning. This also differs from FAITHUN’s KLUE method, which selectively updates neurons associated with target knowledge.

[1] Yang et al. "FAITHUN: Toward Faithful Forgetting in Language Models by Investigating the Interconnectedness of Knowledge." EMNLP, 2025. arXiv:2502.19207.

We hope these responses address your concerns and are happy to discuss further.

AuthorsOpenReview ↗

We are keen to discuss further with you

Dear Reviewer hpEu,

We sincerely thank you again for your time and constructive comments.

In our response, we have: (1) added experiments on YAGO3-10, a real-world knowledge graph, to evaluate generalization beyond Harry Potter; and (2) clarified our relationship to FAITHUN: while FAITHUN examines whether forgetting extends to related questions that depend on the target fact, our work discovers and interrupts alternative supporting paths that can reconstruct the target fact from other retained knowledge.

As the discussion period is approaching its end, we would greatly appreciate it if you could kindly let us know whether our response addresses your concerns or whether there are any further questions. We would be happy to address them.

Best wishes,

Authors

AuthorsOpenReview ↗

We are keen to discuss further with you

Dear Reviewer hpEu,

We sincerely thank you again for your time and constructive comments.

In our response, we have added experiments to evaluate generalization beyond Harry Potter and clarified the distinction between our work and FAITHUN.

As the discussion period is approaching its end, we would greatly appreciate it if you could kindly let us know whether our response addresses your concerns or whether you have any further questions. We would be happy to address them.

Best wishes,
Authors

Thank you for conducting the additional experiments. While I still think the evaluation is not comprehensive, these results would be a valuable addition to the main paper. Please also clarify the experimental setup. In particular, what do GA and NPO refer to here? Are they combined with your proposed method, as in Table 1?

Reviewer_Mam8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DUMIC, a deep unlearning framework designed to prevent forgotten facts from being reconstructed through related knowledge. It combines adaptive supporting-path discovery, latent entity extraction, path verification, and confidence-aware minimum-cut selection. Experiments on two LLMs demonstrate improved unlearning effectiveness while mitigating utility degradation.

Critical strengths
  1. The paper identifies an important limitation of conventional factual unlearning: suppressing direct recall does not necessarily prevent target facts from being reconstructed through retained multi-hop supporting knowledge. This motivates a stronger and practically relevant unlearning objective.
  2. DUMIC provides a structured search, verification and selection framework that combines latent entity discovery with confidence-aware minimum-cut selection. The experiments demonstrate improved supporting-path discovery and consistently higher UES than superficial unlearning across the evaluated configurations, while generally preserving more local utility than full supporting-subgraph removal.
Critical weaknesses
  1. The evaluation does not directly verify the stated objective of deep unlearning. Section 3 requires suppressing both the target fact and its supporting paths, whereas UES (Section 6.1) measures changes in the confidence or correctness of supporting facts without directly testing whether target facts can still be recalled or reconstructed through multi-hop reasoning. Moreover, the evaluation of target forgetting is limited to the Harry Potter domain, with the main experiments focusing on 100 targets selected for having the most verified supporting paths. This limits the evidence for generalization to other domains and knowledge structures.
  2. Section 4.3 merges individually verified supporting paths into a single graph, which may create additional paths between the target entities that have not been independently verified as valid inferences. Section 4.4 uses calibrated model confidence as the cost of removing a fact, but does not establish whether this confidence reflects the actual impact on model utility. Comparisons with alternative weighting strategies would help justify this design. In addition, the experiments lack direct comparisons with existing deep unlearning approaches discussed in Section 2, making it difficult to assess the advantages of DUMIC over prior methods.
Questions for authors
  1. How sensitive are DUMIC's supporting-path coverage and unlearning effectiveness to the maximum search depth, candidate-entity limits, and confidence-filtering thresholds? Could the authors provide a sensitivity analysis?
  2. How are independently selected minimum cuts coordinated when multiple targets share supporting facts? Additionally, how does the method handle source-to-target path combinations introduced by merging separately verified paths?
  3. In Table 2, several superficial-unlearning methods have negative UES values. For example, NPOGDR improves from -0.377 to -0.051, but its UES remains negative. Could the authors clarify how the reported relative improvements are calculated and provide absolute UES gains alongside them?
4 comments

We thank Reviewer Mam8 for the constructive feedback and for recognizing the importance of moving beyond superficial unlearning and the structured search, verification, and selection design of DUMIC.

[W1(a). Not Directly Verify Deep Unlearning ] Thank you for this important point. For rigorous evaluation, we define confidence to precisely characterize the model’s knowledge of individual facts and quantify their suppression after unlearning. To complement this measure, we directly evaluated forgetting on the 100 target triples using Qwen3.5-27B with full fine-tuning, comparing all four methods under deep and superficial unlearning settings without relying on confidence.

MethodSuperficial unlearning: Forget ↓Deep unlearning: Forget ↓
GA44.538.0
NPO69.551.5
SimNPO78.554.0
GA-GDR35.522.0

Deep unlearning consistently achieves lower Forget across all four methods. These results complement our evaluation by directly showing that suppressing supporting knowledge also directly improves forgetting of the target facts themselves.

[W1(b). Generalization beyond the Harry Potter Domain] Thank you for this suggestion. Our framework is applicable beyond Harry Potter, and our choice of domain is motivated by rigorous evaluation. Access to the complete Harry Potter original texts allows us to construct a comprehensive reference from raw text as an approximate upper bound on the model’s knowledge within this domain. Our model-specific filtering then aligns this reference with the knowledge actually retained by each evaluated model, minimizing omissions of relevant supporting knowledge. The same procedure can be applied to other domains, especially when model holders have access to their training corpora. We focus on Harry Potter because, for broader domains such as Wikipedia, the evaluated models’ complete training corpora are unavailable, making it difficult to construct a comparable upper bound. Evaluations in these domains remain useful, but missing supporting knowledge introduces uncertainty into path discovery and evaluation.

To examine applicability beyond Harry Potter, we additionally conducted experiments on YAGO3-10, a knowledge graph dataset, using Qwen3.5-4B with LoRA fine-tuning. We randomly selected 100 target triples.

Search statisticResult
Target triples100
Discovered supporting paths306
Target triples covered by discovered paths61/100
Model / MethodBase Qwen3.5-4BGANPO
Accuracy ↓0.420.290.23

These results demonstrate that our framework can discover supporting paths and perform unlearning beyond Harry Potter, with target accuracy decreasing from 0.42 to 0.29 with GA and 0.23 with NPO.

[W2(a) & Q2. Paths created by merging and shared supporting facts across targets] Thank you for your careful feedback. In our implementation, any additional unverified paths created by merging are excluded from the optimization constraints. We will clarify this distinction in the paper. Since we focus on deep unlearning for individual targets, we select a minimum cut for each target independently, then merge and deduplicate the selected facts.

[W2(b). Selection strategy] Thank you for this question. Confidence serves as a proxy for removal cost, since unlearning confidently held facts disturbs more of the model's knowledge. To validate this choice, we compare DUMIC with an unweighted baseline that unlearns the same number of distinct triples (100 targets + 136 supporting triples) on Qwen3.5-27B with full-parameter fine-tuning.

MethodSelectionUnlearning Effectiveness ↑Loc ↑
GARandom0.900.73
DUMIC0.910.75
NPORandom0.900.76
DUMIC0.920.78
NPO_GDRRandom0.770.87
DUMIC0.870.89
SimNPORandom0.880.78
DUMIC0.900.78

DUMIC consistently achieves higher unlearning effectiveness with equal or higher Loc across all four algorithms. This shows the effectiveness of confidence-weighted selection.

[W2(c). Comparison with existing method] Thank you for raising this concern. Prior deep unlearning research primarily focuses on evaluating whether forgotten knowledge remains inferable [1, 2], rather than proposing directly comparable methods for selecting supporting facts. To systematically assess DUMIC, we evaluate three strategies—unlearning only target facts, removing all supporting facts, and selectively removing facts with DUMIC.

[1] Evaluating Deep Unlearning in Large Language Models.
[2] Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness.

We hope these responses address your concerns and are happy to discuss further.

[Q1. Sensitivity to search hyperparameters] Thank you for the question. We evaluate sensitivity to the maximum search depth by restricting supporting paths to at most 3 hops instead of 4. The table below reports the search result:

ModelMethodTop-100Top-200
≤ 3 hops≤ 4 hops≤ 3 hops≤ 4 hops
Qwen3.5-27BJ-Lens2384334328764067
KG3163354543185006
Gemma4-31B-BaseJ-Lens1810263222083181
KG3317365943185006

The search is effective at both depths. With a maximum depth of 3, J-Lens already discovers 1,810 to 2,876 judge-accepted supporting paths across models and target sets.

We further evaluate how the maximum search depth affects unlearning by applying DUMIC to supporting paths of at most 3 or at most 4 hops, using LoRA on Qwen3.5-27B:

MethodUES ↑Loc ↑
≤ 3 hops≤ 4 hops≤ 3 hops≤ 4 hops
GA0.6610.7890.8430.730
NPO0.5910.3370.8650.914
SimNPO0.1790.1410.9120.944
GA_GDR0.3410.5480.9480.911

With either a maximum depth of 3 or 4, DUMIC achieves effective forgetting while preserving local knowledge. DUMIC is therefore robust to the maximum search depth.

For the candidate-entity limit, we reduce the number of candidates per expansion from 50 to 30 on the targets:

ModelMethodTop-100
Limit 30Limit 50
Qwen3.5-27BJ-Lens21643343
Gemma4-31B-BaseJ-Lens16532632

The search remains effective under a smaller candidate budget. With a limit of 30, J-Lens still discovers 2,164 supporting paths for Qwen3.5-27B and 1,653 for Gemma4-31B-Base.

Concerning the choice of the threshold (u^*=1), we would like to clarify that this value is not selected arbitrarily. As explained in Appendix E and Table 6, 1 bit corresponds to an equal probability split between the correct answer and one alternative. Requiring the correct answer to be the most probable among the five options and entropy below 1 bit ensures that its probability exceeds 0.5, providing an interpretable criterion for retained knowledge.

[Q3. Negative UES] Thank you for raising this question. A negative UES indicates that unlearning is ineffective overall in this setting. Similar failures have been observed in prior work [1–3], reflecting the limited effectiveness of current unlearning methods in these settings.

[1] Wei et al. Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness. NeurIPS 2025.

[2] Borisiuk et al. Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning. Findings of ACL 2026.

[3] Feng et al. Existing Large Language Model Unlearning Evaluations Are Inconclusive. arXiv:2506.00688, 2025..

We hope these responses address your concerns and are happy to discuss further.

AuthorsOpenReview ↗

We are keen to discuss further with you

Dear Reviewer Mam8,

We sincerely thank you again for your time and constructive comments.

In our responses, we have: (1) added direct target-forgetting evaluations to complement our evaluations; (2) added experiments on YAGO3-10 to address generalization beyond Harry Potter; (3) clarified the treatment of unverified paths introduced by graph merging and supporting facts shared across targets; (4) added a comparison with random selection to assess confidence-weighted selection; and (5) provided sensitivity analyses for search depth and candidate-entity limits.

As the discussion period is approaching its end, we would greatly appreciate it if you could kindly let us know whether our responses address your concerns or whether there are any further questions. We would be happy to address them.

Best wishes,

Authors

AuthorsOpenReview ↗

We are keen to discuss further with you

Dear Reviewer Mam8,

We sincerely thank you again for your time and constructive comments.

In our responses, we have added experiments to further evaluate forgetting effectiveness, generalization beyond Harry Potter, and the design of our framework. We have also clarified how supporting paths and facts shared across targets are handled.

As the discussion period is approaching its end, we would greatly appreciate it if you could kindly let us know whether our responses address your concerns or whether you have any further questions. We would be happy to address them.

Best wishes,
Authors

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

The paper studies LLM unlearning beyond suppressing direct recall of target facts, arguing that forgotten facts may remain recoverable through inference over retained supporting knowledge. It proposes DUMIC, a framework that discovers model-specific supporting paths and uses a confidence-weighted minimum cut to select additional supporting facts for the forget set. Experiments with five unlearning algorithms on two LLMs show that DUMIC improves forgetting effectiveness over target-only unlearning while better preserving utility than a baseline that unlearns all discovered supporting facts. The paper also argues that evaluations focused on direct recall can overlook inferential recovery and introduces a model-specific knowledge-graph evaluation pipeline to assess this aspect of forgetting.

Critical strengths
  1. A concrete framework for selecting supporting facts to unlearn. DUMIC combines model-specific supporting-path discovery with confidence-weighted minimum-cut selection and can be paired with existing unlearning algorithms. It helps prevent the model from easily reconstructing forgotten facts while preserving other useful knowledge.

  2. Consistent empirical evidence across multiple unlearning algorithms. Experiments with five algorithms on two LLMs show improved forgetting over target-only unlearning and better utility preservation than removing all discovered supporting facts. These results support the practical value of selective supporting-fact removal and suggest that its benefits extend beyond a particular unlearning objective.

Critical weaknesses
  1. The limited ablations leave it unclear whether DUMIC’s gains come from its selection strategy or simply from unlearning additional supporting facts. The authors should compare against random supporting-fact selection with the same number of distinct forget-set facts and clarify whether Appendix F.1 matches distinct facts or only training examples. An unweighted minimum-cut baseline would also help establish the benefit of confidence weighting.

  2. The paper claims that target-only unlearning suppresses direct recall while leaving facts recoverable through inference, but UES alone does not establish this gap because it measures supporting-path disruption rather than actual recovery. The authors should report direct-recall accuracy and inference-attack success alongside UES to test this claim and whether DUMIC reduces recovery.

Questions for authors
  1. In Appendix F.1, how is the target-only forget set expanded to 236 examples? Are these repetitions or paraphrases of the original 100 target facts, or additional distinct facts? Please clarify whether this ablation matches the number of training examples or the number of distinct facts being unlearned.

  2. Could the authors report direct target-recall accuracy and inference-attack success, following an evaluation such as [1], alongside UES? This would distinguish suppressed direct recall with continued inferential recovery from a general failure to forget, supporting the claim that direct recall alone overestimates forgetting.

[1] Wei et al. “Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness.” NeurIPS, 2025. arXiv:2506.05735.

4 comments

We thank Reviewer cHuk for the constructive feedback and for recognizing DUMIC's concrete selection framework and its consistent gains across unlearning algorithms.

[W1. Effectiveness of selection strategy] Thank you for raising this important practical concern. To verify the effectiveness of our selection strategy, we compare DUMIC with a random baseline that unlearns the same number of distinct triples (100 targets + 136 supporting triples) on Qwen3.5-27B with full-parameter fine-tuning.

MethodSelectionUnlearning Effectiveness ↑Loc ↑
GARandom0.900.73
DUMIC0.910.75
NPORandom0.900.76
DUMIC0.920.78
NPO_GDRRandom0.770.87
DUMIC0.870.89
SimNPORandom0.880.78
DUMIC0.900.78

DUMIC consistently outperforms random selection in unlearning effectiveness, with equal or higher Loc. With the forget-set size fixed, the gains stem from DUMIC's selection strategy. We will add these results to the revision.

[W2 & Q2. Evaluation with direct accuracy] Thank you for this insightful suggestion. Following the evaluation of Wei et al. [1], we report direct target recall (Recall ↓) alongside the UES metric from [1] for superficial unlearning (targets only), DeepUnlearn (all discovered supporting facts), and DUMIC on Qwen3.5-27B with full-parameter fine-tuning.

MethodSettingUES ↑Recall ↓Loc ↑
No Unlearn–0.001.001.00
GASuperficial0.660.310.89
DeepUnlearn0.610.270.71
DUMIC0.70 (+6.96%)0.230.75
NPOSuperficial0.660.330.89
DeepUnlearn0.770.200.65
DUMIC0.84 (+26.72%)0.180.78
SimNPOSuperficial0.620.340.85
DeepUnlearn0.790.170.65
DUMIC0.75 (+21.07%)0.250.78
NPO_GDRSuperficial0.280.560.95
DeepUnlearn0.890.120.61
DUMIC0.76 (+170.21%)0.250.89

Compared with superficial unlearning, DUMIC achieves both higher UES (+6.96% to +170.21%) and lower Recall for all methods, making the targets harder to recall directly and harder to infer from retained knowledge. Compared with DeepUnlearn, DUMIC retains substantially more local knowledge (Loc +0.04 to +0.28), achieving a better trade-off between forgetting and utility.

[1] Wei et al. Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness. NeurIPS 2025.

[Q1. Construction of the forget set] Thank you for this critical observation. In Appendix F.1, the 236 examples are the original 100 target facts plus 136 repetitions of randomly selected target facts, so that the number of unlearning examples matches that of DUMIC. This ablation thus controls the number of training examples, while the comparison in W1 further controls the number of distinct facts. We will clarify this in Appendix F.1.

We hope these responses address your concerns and are happy to discuss further.

I thank the authors for the additional experiments and detailed response.

Regarding W1, the random baseline appears to perform well relative to the other two baselines in the paper (Superficial Unlearn and DeepUnLearn). DUMIC performs only marginally better than the random baseline, except in the NPO (GDR) setting. This suggests that DUMIC’s benefits may stem primarily from unlearning additional facts rather than from its selection strategy. My concern regarding the effectiveness of the proposed method therefore remains. In this regard, could the authors clarify in what respects they believe DUMIC offers an advantage over random selection?

Regarding W2, thank you for the additional results. My concern regarding recall has been largely addressed. One remaining question is why the UES and loc values reported here differ from those reported in the response to W1 and in Table 1.

We thank Reviewer cHuk for the follow-up.

[W1. Advantage over random selection] For GA, NPO, and SimNPO, random selection already achieves unlearning effectiveness scores of 0.88–0.90. The limitations of the underlying unlearning algorithms make substantial further gains difficult in this setting. For NPO_GDR, this algorithm yields weaker forgetting, need to be impreoved. DUMIC raises unlearning effectiveness from 0.77 to 0.87 and improves Loc from 0.87 to 0.89. These results illustrate the benefit of confidence-aware selection, particularly when there is more room to improve forgetting while preserving utility.

[W2. Result Difference] The W2 results follow the evaluation protocol of Wei et al. [1]. Their protocol uses the same metric names, UES and Loc, as our paper. The values differ because W1 and Table 1 use our evaluation protocol, while W2 uses theirs. We will clarify this in the revision.

We hope these clarifications help address your remaining concerns, and we would be grateful if you could take them into account in your final assessment.

[1] Wei et al. Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness. NeurIPS 2025.

Thank you for the clarification. If random selection already approaches the limits of the underlying unlearning algorithms for GA, NPO, and SimNPO, this would suggest that random selection is sufficient in these settings, leaving limited benefit from the proposed selection strategy. I acknowledge the improvement for NPO_GDR, but this demonstrates an advantage in one setting rather than a consistent benefit across the evaluated algorithms. My concern therefore remains that the reported gains over the original baselines may primarily come from unlearning additional facts, while the added value of DUMIC’s selection strategy is limited or dependent on the underlying algorithm.

Regarding W2, is there a fundamental reason for using different evaluation protocols across these experiments? Aligning the protocols would make the results easier for readers to compare and interpret

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.