acceptodds
Under review as a conference paper at ICLR 2027

Where to Perturb? Perturbation Allocation in Representation-Space Augmentation for Sparse Event Detection

Abstract

Representation-space augmentation can conflate how much an input is perturbed with where perturbation acts. We separate these factors in sparse event detection by calibrating post-normalization squared displacement and independently varying training allocation and test channel. In the frozen-representation CASIE setting under strong Gaussian stress, sentence-only and token-only training reverse their Trigger F1 ordering between the sentence and token test endpoints in 14 of 15 paired seeds. Averaged equally over five test allocations, however, sentence-only training has a -point advantage. Joint training improves the token-endpoint mean over sentence-only training while sacrificing performance at the sentence endpoint, illustrating a tradeoff rather than a uniform gain. Aggregate rankings therefore depend on the weights assigned to evaluation channels. Positive allocation–channel interactions persist across frozen-setting controls for distance, loss, encoder, and classifier, and extend to exact-span detection on PHEE. The corresponding interaction estimates are near zero under the tested end-to-end protocol. Local gradients show lower sensitivity to the trained-noisy path and higher sensitivity to the other path; permutation diagnostics provide complementary evidence consistent with a redistribution of path reliance. These results show why augmentation comparisons should report conditional robustness profiles alongside total displacement and specify the optimization and evaluation protocols under which they hold.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation-Space Augmentation, Perturbation Allocation, Robustness, Sparse Event Detection, Cybersecurity

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

The paper explore the effects of injecting perturbations in the latent space between a frozen BERT-like encoder and a trainable classifier. Various types of noise is injected at a token-level, sentence-level, and combination in order to explore and isolate various effects. One of the main contributions is that the paper isolates noise magnitude and allocation. The authors show that training on certain noise allocations may results in degradations of other allocations. The findings are primarily evaluated on CASIE, a security even-detection dataset.

Critical strengths
  • The authors diligently explore the problem of which types of perturbation allocations during training result in robustness during evaluation. By fixing the amount of perturbation budget, the authors aim to isolate the affects of token-level versus sentence-level perturbations. The authors conduct an extensive set of experiments and perform various ablation and control studies to assess the robustness of their findings and isolate the effects of different perturbation allocations.
  • The authors demonstrate an observation that the allocation of training perturbations affects robustness under differing evaluation perturbations. The performance of sentence-level and token-level training perturbations flips depending on the evaluation perturbation channel. This suggest that there is no universal best strategy.
Critical weaknesses
  • The authors lack clarity and provide insufficient motivation for their methods. There are a multitude of concepts which are ill-explained including "training arms", "cache", and "test allocations". Additionally, the metrics Trigger-I and Trigger-C are not well-defined. The PHEE dataset is also not cited properly in the main part of the paper, despite being a central point for evaluation. Finally, the modeling choices are poorly defined and not sufficiently motivated. It is unclear why SecureBERT and ModernBERT where chosen as the models for exploration, without giving additional details about their size and training procedure. The classifier which was used for training could be described with more detail.
  • There is weak significance and limited generalizability. Is not immediately clear how the methods explored here extend to modern fine-tuning approaches. Additionally, despite the focus on cyber security detection events, the paper could benefit from a stronger explanation for why this particular use-case is chosen. Tying together how the methods explored in this paper would benefit cyber security event detection more concretely would have helped piece together the paper.
Questions for authors
  • Can the authors please clarify the decision behind choosing SecureBERT as the main encoder of choice? Additionally by ModernBERT was used for additional experiments?
  • Can the authors please clarify what concrete failure modes the perturbations are approximating?
  • Can the authors please clarify what role "cache" and "finite precision storage" play in the role of their proposed methodology?
4 comments
AuthorsOpenReview ↗

Clarifications on Model Choices, Perturbations, and Caching

Thank you for the careful review and for identifying where our presentation and definitions need clarification. We address Q1–Q3 first, followed by W1–W2. Additional experimental results are marked [new].

Q1. Why SecureBERT, and why ModernBERT?

  • SecureBERT (Aghaei et al., 2023; RoBERTa-base architecture, ~125M parameters, 768 hidden dimensions) is adapted to cybersecurity text, making it a relevant representation model for CASIE’s technical terminology and documents. Freezing it provides a standardized feature space for comparing augmentation allocations while training the downstream classifier under the same architecture, budget, and optimization protocol.
  • ModernBERT-base (Warner et al., 2024; ~149M parameters, 768 hidden dimensions) provides a contemporary general-purpose encoder. It tests whether our allocation findings hold beyond a cybersecurity-specific representation model and supports the matched comparison of frozen and jointly trained encoders described in W2.

We will make these selection reasons and model specifications explicit in the experimental setup.

Q2. What concrete failure modes do the perturbations approximate?

The probes examine instability in two inputs used to classify each candidate token:

  • Token path: changes in the candidate’s contextualized representation, approximating local shifts caused by unfamiliar terminology, unseen technical entities, or local lexical variations.
  • Sentence path: changes in the additional pooled sentence summary supplied alongside that representation, approximating shifts in global sentence context (e.g., incomplete or distracting reporting).

The token representation itself already incorporates context, and natural-text changes can alter both inputs together. Our experiments independently perturb the classifier inputs under a fixed total post-normalization displacement budget to isolate their respective contributions to robustness. We will clarify this interpretation in the Methodology and consolidate the scope of natural-text validation in the Limitations.

Q3. What role do “cache” and “finite-precision storage” play?

Because the encoder is frozen in our primary controlled study, we compute and cache the representations once, allowing all augmentation conditions to reuse identical features without repeated encoder computations. The vectors are stored in float16 and cast to float32 when loaded. Noise is added to the loaded values before normalization, and displacement is measured relative to the corresponding normalized clean vectors. The calibration and realized-budget checks account for this numerical implementation.

To assess the numerical consistency of cached and freshly reconstructed representations, our ModernBERT reconstruction check rebuilt the features directly from the converted texts using the identified encoder weights, tokenizer, and projection, then evaluated them with the trained classifiers held fixed. The maximum F1 difference was 0.0086 points on clean evaluation and 0.0269 points across 270 noisy evaluations, while the positive stress interaction was preserved. These results support the numerical fidelity of the cached evaluation pipeline. We will explain the cache’s practical role briefly in §3.1 and retain the reconstruction and precision details in Appendices A.4–A.5.

AuthorsOpenReview ↗

Addressing Clarity, Significance, and Generalizability

W1. Clarity and definitions

We will bring the following definitions into the main text:

  • Training condition and allocation. We will replace “training arm” with “training condition.” At a fixed training budget, α∈[0,1]\alpha \in [0, 1] denotes the fraction of the total post-normalization squared-displacement budget assigned to the token path. The test allocation β∈[0,1]\beta \in [0, 1] independently specifies the corresponding fraction during evaluation. A value of 00 denotes sentence-only perturbation and 11 denotes token-only perturbation. Both representation inputs remain available in every condition.
  • CASIE metrics. Trigger F1 evaluates token-level event identification by merging event types into one positive class. Typed F1 additionally requires the correct event type.
  • PHEE and its metrics. PHEE (Sun et al., 2022) is a pharmacovigilance event-extraction dataset. Our PHEE experiments evaluate trigger detection: Trigger-I requires exact trigger boundaries, and Trigger-C additionally requires the correct event type. We will cite and define PHEE when first introducing it.

We will also move the classifier description from Appendix C to §3.1. A fixed orthonormal projection maps 768-dimensional encoder states to 128 dimensions, followed by L2L_2 normalization. The sentence input is the L2L_2-normalized mean of these token vectors. The primary CASIE classifier forms separate 64-dimensional tanh projections of the two inputs, concatenates their elementwise product with the original inputs, and applies a linear layer, GELU, LayerNorm, and an output layer (five event classes plus background for CASIE). The allocation–channel interaction also persists with a simpler concatenation head that removes the multiplicative feature: at test budget T=8E0T=8E_0, the interaction is 12.94 Trigger F1 points (95% CI [8.41, 17.46]) (Appendix B.4, Table 23).

W2. Significance and generalizability

  • Encoder updating. [new] We conducted 20 matched ModernBERT training runs across frozen/trainable encoders, two endpoint training allocations, and five paired seeds under a common protocol. At test budget T=8E0T=8E_0, the Trigger interaction—the change in the relative advantage of the two training allocations between the endpoint test channels—decreases from 13.50 to 1.03 F1 points with encoder updates. The paired reduction is 12.47 points (95% Student-tt CI [9.52, 15.43]). This identifies encoder updating as a strong moderator of allocation-dependent robustness.

  • Evidence beyond CASIE. [new] PHEE now covers all five TextEE splits, with eight paired seeds per split. At test budget T=8E0T=8E_0, both Trigger-I and Trigger-C exhibit positive interactions of approximately one F1 point, with all ten split-specific 95% Student-tt confidence intervals entirely above zero. The smaller PHEE interaction is accompanied by weaker redistribution of path sensitivity. These results consistently reproduce the conditional robustness finding across data partitions and extend it to pharmacovigilance exact-span detection.

  • Cybersecurity motivation and practical significance. Sparse cybersecurity event detection (such as CASIE) is a particularly fitting testbed because threat intelligence reports combine rapidly evolving technical terminology (challenging local token representations) with complex, noisy sentence structures (challenging global sentence summaries) under limited annotation budgets. More broadly, our findings show that equal perturbation budgets can produce different conditional robustness profiles, with aggregate rankings depending on evaluation-channel weights. This motivates reporting robustness across test allocations and making aggregation weights explicit. We will highlight these findings and their implications in the revised Results and Discussion.

Thank you for adding answers to my questions/weaknesses.

I checked the paper manuscript again and did not find the updates you mentioned (i.e. defining PHEE when it is first introduced). Are you planning on adding these before the deadline?

Thank you for following up and for taking the time to check the manuscript again. We have submitted the revised manuscript and incorporated the changes discussed in our response.

We clarified the perturbation settings and evaluation metrics, explained our modeling choices and the motivation for studying cybersecurity event detection, and defined and cited PHEE in the abstract (p. 1, line 025). We also report the matched comparison of frozen and fine-tuned ModernBERT and the results across all five PHEE splits (p. 8).

We hope these revisions address your concerns and make the study’s motivation and the scope of its findings clearer.

Reviewer_XKZ6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses a methodologically motivated research question and proposes a budget calibration framework supported by multiple controlled experiments.

The authors identify an interaction between perturbation allocation and test-time perturbation channels under frozen-representation settings. However, this interaction is not evident under end-to-end ModernBERT training, and the cross-dataset evidence remains limited.

Overall, the paper has research merit, but the significance and generalizability of its central findings require further validation.

Critical strengths
  1. Reasonable motivation and a meaningful research direction. The authors point out that changing the perturbation path may also alter the total perturbation magnitude, potentially compromising the fairness of comparisons between augmentation strategies. This is a reasonable methodological concern.

  2. A well-designed experimental framework with targeted controls. The main experiments are supported by multiple control experiments that examine the observed findings under different experimental settings.

  3. Insightful findings with practical implications. Under the frozen-representation setting on CASIE, different training allocation schemes exhibit different relative advantages across test-time perturbation channels. The reversal of endpoint rankings is also observed across multiple random seeds, suggesting that a single aggregate metric may obscure trade-offs in conditional robustness.

Critical weaknesses
  1. The generalizability of the main findings is limited by their sensitivity to experimental settings. Under end-to-end ModernBERT training, the estimated interaction effect is close to zero. Although the authors discuss possible explanations, the underlying reasons are not sufficiently investigated, limiting our understanding of the conditions under which the observed interaction holds.

  2. The novelty relative to existing work on representation-space perturbation and robustness remains insufficiently established. The paper investigates perturbation allocation through budget calibration and crossed training–test evaluation. However, the authors have not sufficiently clarified what new insights their framework and empirical findings offer beyond the most closely related prior work.

  3. The cross-dataset evidence remains limited. The interaction effect observed on PHEE is smaller than that in the main experiments, and the evaluation is based on a single data split. Although these results provide some cross-dataset evidence, they are insufficient to establish that the phenomenon generalizes to a broader range of tasks.

Questions for authors
  1. Why is a pronounced interaction between perturbation allocation and test-time perturbation channels observed under frozen-representation settings, whereas the interaction effect is close to zero under end-to-end ModernBERT training? Can the current experiments disentangle the respective effects of encoder adaptation, optimization dynamics, and other differences in experimental settings?

  2. Compared with existing studies on representation-space perturbation and robustness, what new insights does the proposed budget calibration framework and the analysis of perturbation allocation–test-channel interactions provide beyond what has already been established in prior work?

  3. Given that the interaction effect observed on PHEE is smaller than that in the main experiments, what range of tasks and model settings is supported by the current evidence? What additional validation would be needed to establish broader generalizability?

4 comments
AuthorsOpenReview ↗

Matched ModernBERT Comparisons

Thank you for your constructive feedback. To directly address your questions, we conducted matched ModernBERT comparisons (20 additional training runs), expanded PHEE evaluations across five splits (64 additional training runs), and pathwise sensitivity and permutation diagnostics.

Q1. Why the Interaction Is Attenuated Under End-to-End Training & Disentangling Experimental Factors

1. Controlled Comparison of Encoder Updating (20 Additional Training Runs). In our initial submission, the frozen (SecureBERT) and end-to-end (ModernBERT) experiments differed in both backbone architecture and training protocol. To isolate the effect of enabling encoder updates under a common architecture and protocol, including the associated changes in optimization dynamics, we trained 20 matched ModernBERT models (Frozen vs. Trainable×Sentence-only (α=0) vs. Token-only (α=1)×5 paired seeds\text{Frozen vs. Trainable} \times \text{Sentence-only } (\alpha=0) \text{ vs. Token-only } (\alpha=1) \times 5 \text{ paired seeds}). Within each seed, all four conditions share matched parameter initialization, representation construction, fixed orthonormal projections, and checkpoint selection based on clean development-set performance. Perturbations are calibrated using post-normalization squared displacement against a common reference budget E0E_0 established before task-specific training.

At training budget E=E0E=E_0 and test budget T=8E0T=8E_0, the endpoint interaction is I=(F00−F10)−(F01−F11)I = (F_{00} - F_{10}) - (F_{01} - F_{11}), where subscripts denote training and test token-budget shares (00: sentence-only; 11: token-only). For CASIE token-level F1 metrics, we first average the three evaluation noise draws within each seed and condition, then compute seed-level interactions and report means and 95% Student-tt intervals across the five paired seeds (ΔI=Ifrozen−Itrainable\Delta I = I_{\mathrm{frozen}} - I_{\mathrm{trainable}} is computed from within-seed differences):

MetricFrozen ModernBERT IITrainable ModernBERT IIPaired Difference ΔI\Delta I
Trigger F113.50 [11.64, 15.36]1.03 [−0.94, 2.99]12.47 [9.52, 15.43]
Typed F115.19 [13.62, 16.77]1.01 [−0.96, 2.98]14.18 [11.28, 17.08]

2. Pathwise Diagnostics & Mechanistic Interpretation.

  • Effect of Encoder Updating: Frozen ModernBERT exhibits a substantial interaction (13.5013.50 Trigger and 15.1915.19 Typed F1 points), showing that the phenomenon also occurs strongly with frozen ModernBERT and is not specific to SecureBERT. Enabling encoder updates under the same protocol substantially attenuates II.
  • Path-Reliance Diagnostics: Clean-input diagnostics across all 20 matched models (evaluated on all 3,054 positive CASIE test event tokens for gradient sensitivity and all 87,317 test candidate rows for permutation drops) identify two corresponding changes that explain this attenuation:
    1. Reduced Reliance on the Explicit Sentence Branch: Under trainable encoders, sentence-branch permutation reduces Typed F1 relative to the clean baseline by only 2.99 (α=0\alpha=0) and 3.19 (α=1\alpha=1) F1 points, compared to 14.72 and 22.09 F1 points under frozen encoders (with the sentence share of squared tangent sensitivity dropping from 32.87%32.87\% and 60.71%60.71\% to 16.78%16.78\% and 24.96%24.96\%, respectively).
    2. Weaker Allocation-Induced Redistribution: Shifting from sentence-only (α=0\alpha=0) to token-only (α=1\alpha=1) training increases the sentence share of squared tangent sensitivity by 27.85 [26.98,28.71][26.98, 28.71] percentage points under frozen training, versus 8.19 [5.31,11.06][5.31, 11.06] percentage points under trainable encoders. Correspondingly, the allocation-induced increase in Typed F1 sentence-path permutation drop shrinks from +7.37 [5.96,8.77][5.96, 8.77] F1 points (frozen) to +0.21 [−1.54,1.95][−1.54, 1.95] F1 points (trainable).
  • Interpretation & Practical Scope: These findings support the interpretation that encoder adaptation allows task-relevant contextual information to be represented more effectively within the token path, reducing classifier reliance on the explicit sentence branch and muting allocation-induced sensitivity shifts. Budget calibration applies to both regimes, and frozen representations remain practically relevant for precomputed feature caches, multi-task deployment sharing a frozen encoder, and settings with limited training resources.
AuthorsOpenReview ↗

Novelty Beyond Prior Work

Q2. New Insights Beyond Existing Representation-Space Perturbation Literature

Prior studies such as FreeLB (Zhu et al., 2020), R3F (Aghajanyan et al., 2021), and HyPe (Yuan et al., 2023) add adversarial or random perturbations to input embeddings or intermediate hidden states—with HyPe also comparing layer locations—to improve generalization and robustness. Our framework addresses a complementary question and provides three distinct insights:

  1. Separating Path Allocation from Magnitude Confounding: In architectures combining token and pooled-sentence representations, altering where perturbations are injected or activating a second path at fixed nominal scales can also alter total post-normalization displacement. By matching the total squared displacement after normalization (E0E_0-calibrated), our framework isolates how distributing perturbations across token and sentence paths changes the resulting robustness profile.
  2. Aggregate Metrics Can Obscure Conditional Robustness Reversals: By independently crossing training allocations with test perturbation channels, we show that an advantage averaged over test conditions can coexist with reversals in the relative ranking of training allocations across individual channels. In the primary CASIE comparison with frozen SecureBERT, at training budget E=E0E=E_0 and test budget T=8E0T=8E_0, sentence-only training has a 3.53-point Trigger F1 advantage when averaged equally across the five test channels, yet sentence-only and token-only training reverse their endpoint ranking in 14 of 15 paired seeds (and 15 of 15 for Typed F1).
  3. Encoder Adaptation and Path Reliance Shape Interaction Strength: Combining our budget-matched comparisons with pathwise diagnostics, we identify encoder adaptation as an empirical boundary condition (Q1) and show how interaction magnitude relates to the distribution and allocation-induced redistribution of sensitivity across representation paths (Q3).
AuthorsOpenReview ↗

5-Split PHEE Validation

Q3. Consistency Across Five PHEE Splits, Mechanism of the Smaller Effect Size, and Supported Scope

1. Consistency Across Five PHEE Splits (64 Additional Training Runs). To test whether the cross-dataset evidence extends beyond a single partition, we evaluated four additional TextEE PHEE splits with frozen SecureBERT, scaling all splits to 8 paired seeds per allocation endpoint (64 additional training runs in total) and retaining split 1 (8 seeds). Trigger-I measures exact-span identification; Trigger-C additionally requires correct event typing. At training budget E=E0E=E_0 and test budget T=8E0T=8E_0, the endpoint interactions II (in F1 points, with 95% seed-level Student-tt intervals) are:

SplitSeedsTrigger-I II [95% CI]Trigger-C II [95% CI]
1 (Original)81.07 [0.52, 1.61]1.18 [0.62, 1.75]
280.99 [0.37, 1.60]1.05 [0.43, 1.67]
380.97 [0.61, 1.33]1.01 [0.65, 1.38]
480.80 [0.58, 1.02]0.80 [0.52, 1.09]
581.08 [0.37, 1.80]1.09 [0.37, 1.81]
5-Split Mean8/split0.981.03

All five splits exhibit positive mean interactions with 95% CIs strictly excluding zero on both metrics, yielding equal-weight cross-split means of 0.98 (Trigger-I) and 1.03 (Trigger-C) F1 points. Meanwhile, the five-channel average difference (sentence-only minus token-only) remains near zero across splits (approximately −0.13-0.13 F1 points for both Trigger-I and Trigger-C), indicating little aggregate preference despite consistent positive endpoint interactions.

2. Mechanistic Explanation: Why the Interaction Is Smaller on PHEE Than on CASIE. We tested whether the smaller interaction on PHEE corresponds to weaker allocation-induced redistribution of path reliance on the development sets (comparing frozen SecureBERT models on PHEE split 1 against a matched CASIE development control):

  • Tangent-Sensitivity Redistribution: Across all positive development-set event tokens (PHEE split 1: n=1,054n=1,054; CASIE development control: n=1,598n=1,598), the token share of squared tangent sensitivity under sentence-only training (α=0\alpha=0) is 91.09% on PHEE and 67.71% on CASIE. Moving from sentence-only (α=0\alpha=0) to token-only (α=1\alpha=1) training reduces this token sensitivity share by only 2.51 percentage points on PHEE (91.09%→88.58%91.09\% \rightarrow 88.58\%; 95% CI [2.06,2.96][2.06, 2.96], 8 seeds), compared to 27.51 percentage points on the CASIE development control (67.71%→40.20%67.71\% \rightarrow 40.20\%; 95% CI [26.59,28.43][26.59, 28.43], 5 seeds). The relative increase in sentence sensitivity share is likewise smaller on PHEE (1.28×1.28\times vs. 1.85×1.85\times on CASIE).
  • Complementary Permutation Diagnostic: Across the full development splits, we additionally measured the allocation-induced difference (α=1\alpha=1 minus α=0\alpha=0) in sentence-path permutation drop, providing a functional assessment of sentence-path reliance that complements normalized sensitivity shares. For trigger identification, the allocation-induced difference in sentence-path permutation drop is 0.05 F1 points [−0.19,0.28][−0.19, 0.28] on PHEE (exact-span Trigger-I), compared with 6.35 F1 points [5.02,7.67][5.02, 7.67] on CASIE (token-level Trigger F1).

3. Supported Scope and Criteria for Broader Validation.

  • Supported Regime: Our current evidence supports allocation–channel interactions in frozen-representation event detection on CASIE and PHEE, covering cybersecurity and biomedical events with token-level and exact-span evaluation, respectively, using task heads that combine token and pooled-sentence representations. Within this regime, PHEE exhibits a much more token-dominated sensitivity profile and substantially weaker allocation-induced redistribution of sentence-path reliance than CASIE, providing a coherent explanation for its smaller interaction magnitude.
  • Additional Validation for Broader Generalizability: Establishing broader generalizability beyond sparse event detection will require replicating this budget-matched crossed evaluation across additional multi-path tasks (e.g., relation extraction or argument role labeling) and parameter-efficient adaptation regimes (e.g., LoRA/adapters), testing whether larger allocation-induced shifts in path sensitivity and permutation cost consistently accompany larger channel interactions. We will incorporate all new experiments and clarifications into the revised manuscript.

Thank you again for your thoughtful review. We have submitted the revised manuscript and incorporated the experiments and clarifications discussed in our response.

On CASIE, our matched ModernBERT comparison shows that fine-tuning reduces the interaction between training allocation and test channel from 13.50 to 1.03 Trigger F1 points under strong perturbations. On PHEE, the interaction was smaller but positive on both metrics across all five splits, with eight paired seeds per split (p. 8).

We also clarified how our evaluation extends prior work. Holding the total perturbation budget fixed reveals when the ranking of training strategies reverses across test channels, even when their average scores favor one strategy. We hope these revisions address your concerns and clarify both the contribution and the scope of our findings.

Reviewer_nD4pOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper investigates how allocating a fixed perturbation budget between token-level and sentence-level representations affects the robustness of neural event detection models. The authors introduce a controlled evaluation framework that separates perturbation magnitude from perturbation allocation and examine how different training allocations influence performance under different test-time perturbation conditions. The experiments primarily focus on sparse cybersecurity event detection, with additional analyses across model architectures, training settings, and datasets.

Critical strengths

The paper addresses an interesting question regarding how perturbation allocation affects model robustness and presents a systematic experimental investigation. However, as I am not an expert in NLP event detection and representation-space augmentation, I do not have sufficient domain expertise to reliably assess the novelty and significance of its contributions.

Critical weaknesses

I am unable to provide a sufficiently reliable assessment of the paper's critical weaknesses. In particular, evaluating the methodological novelty, the appropriateness of the experimental setup, and the significance of the findings within the event detection and representation-learning literature requires expertise beyond my primary research area. I therefore do not feel qualified to make a well-informed judgment about the overall technical contribution of this submission.

Questions for authors

NA

1 comment
AuthorsOpenReview ↗

Clarifications and Presentation Improvements

Thank you for recognizing the research question and the systematic experimental investigation. We appreciate your feedback regarding the presentation.

We will revise the Introduction and Methodology to clarify the motivation, the two representation paths, and the distinction between perturbation magnitude and allocation. We will also reorganize the Results to highlight the central finding: even under a fixed perturbation budget, different training allocations can exhibit opposing advantages across evaluation channels, making aggregate robustness rankings dependent on the evaluation conditions.

These revisions aim to make the paper's main contributions and their implications more accessible to readers outside this specific research area.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.