acceptodds
Under review as a conference paper at ICLR 2027

LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs

Abstract

Large language models exhibit misbehaviors that hinder deployment in high-stakes settings. One approach to monitor such failures is to use an external LLM as a monitor, but this is prohibitively expensive. A well-known and efficient class of monitors can detect such failures from internal representations (such as activation probes), but are tied to the model they were trained on, requiring costly re-annotation as new LLMs appear. A natural way to avoid this repeated annotation cost is to develop internal-representation based monitors that transfer zero-shot across LLMs. However, such transfer is challenging because internal representations are not naturally aligned across models; in particular, attention maps, known to contain signals of misbehavior, may differ in head function, ordering, and model depth. In this work, we characterize those cross-model attention mismatches as algebraic symmetries and use geometric deep learning principles to build **LEMON** (**L**anguage-model **E**quivariant **MO**nitoring **N**etwork) an *online* graph neural network-based monitor, whose predictions are invariant under these symmetries. LEMON processes attention maps token-by-token, as they are produced and outputs a score after every token (or every window of tokens) at negligible cost. Across 16 LLMs and six datasets, in the zero-shot cross-model setting, LEMON outperforms all baselines on both hallucination and jailbreak monitoring tasks, improving over the strongest by up to 14 ROC-AUC points.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Symmetries, Large Language Models, Misbehaviors, Graph Neural Networks, Geometric Deep Learning, hallucinations, jailbreak

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yFk8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LEMON, a symmetry-aware neural architecture for zero-shot misbehavior monitoring across LLMs. Motivated by the observation that misbehavior-related attention patterns may appear at different layer-head coordinates across models, the authors characterize head permutation symmetries and introduce a group-equivariant graph neural network for processing structured attention features. LEMON is evaluated on hallucination and jailbreak detection, demonstrating improvements over existing baselines in cross-model transfer settings with relatively low inference overhead.

Critical strengths
  1. Interesting motivation. The paper offers a valuable perspective on coordinate mismatch in attention representations across LLMs. The connection to algebraic symmetries provides a principled motivation for architecture design.

  2. Promising cross-model generalization. LEMON consistently improves over prior baselines in leave-one-model-out evaluations. The additional experiments on unseen datasets and models further support its transferability.

Critical weaknesses

The necessity of the full architectural design is insufficiently justified.

Table 4 shows that the simpler LEMON-per-cell variant outperforms the full architecture on average (69.4 vs. 67.8 AUROC), while removing the cross-token Transformer has negligible impact. These results suggest that the main gains may come from the simpler equivariant, weight-sharing design rather than the more elaborate cross-layer/head interactions and sequence modeling components. The paper would benefit from a clearer justification of these additional architectural choices.

Questions for authors

Given the strong performance of LEMON-per-cell, what additional benefit does the full wreath-product equivariant architecture provide? Could the authors clarify whether the gains primarily arise from simple parameter sharing and permutation equivariance, rather than the more elaborate cross-layer/head interactions? A parameter-matched comparison would help disentangle these factors.

3 comments
AuthorsOpenReview ↗

Rebuttal Response to yFk8

We thank the reviewer for the thoughtful comments and questions. We are pleased that the reviewer finds our perspective on coordinate mismatches across LLMs valuable and recognizes both the principled design of our approach and its promising cross-model generalization results. Below, we address each of the points raised in turn.


W1

We would like to emphasize that both LEMON-W and LEMON-W-Per-Cell naturally arise from the structure we identify in the tensor of attention maps across layers and heads. In particular, arriving at LEMON-W-Per-Cell as an architectural choice relies on the following reasoning process, which we consider part of our contribution.

First, we identify that the heads can be permuted, without changing the underlying function (please refer to section 3.2). As for the layers, motivated by prior work [1] we treat translations as an approximate symmetry along a one-dimensional signal axis. On a discrete axis (such as the layer axis in our setting) translations correspond naturally to shifts. For a finite sequence, these shifts can be represented cyclically, providing a convenient boundary convention without requiring the underlying sequence itself to be periodic. Finally, we understand that this combined symmetry of the layers and heads are formalized as the wreath product of the two groups, W\mathcal{W}. This thought process, which results in the identification of the group combining layers' and heads' mismatches W\mathcal{W}, is what allows us to develop both LEMON-W, and LEMON-W-Per-Cell, which is a special case of LEMON-W.

Furthermore, and more broadly, we believe there is value in first deriving the general architecture that respects the full symmetry (LEMON-W) and then considering meaningful special cases (LEMON-W-Per-Cell). Different misbehavior-monitoring tasks may benefit from different inductive biases: in our experiments, LEMON-W-Per-Cell performs best for hallucination detection, while LEMON-W performs best for jailbreak detection. Therefore, rather than considering either variant universally preferable, our results indicate that the appropriate degree of information sharing may depend on the specific monitoring task.

We also note that studying restricted special cases of more general equivariant architectures is common in the literature. For instance, [2] considers permutation-equivariant set architectures with both a main architecture with global information sharing and a restricted element-wise variant, paralleling the relationship between LEMON-W and LEMON-W-Per-Cell, although in a different symmetry setting. Moreover, more constrained variants actually can sometimes perform better in practice: in a different setting and task, [3] finds that preserving information locally, rather than combining information across the entire representation, consistently leads to better performance in their experiments. Our hallucination experiments with LEMON-W-Per-Cell exhibit a similar pattern.


Q1

First, we want to note that, we believe that the observed differences go beyond simply a parameter-count effect: both LEMON variants are comparable in size and are lightweight, on the order of 100K parameters. Thus, the distinction between them primarily concerns how information is shared and mixed across the layer-head structure, rather than a substantial difference in model capacity.

That being said, we believe the reviewer’s question opens up an important direction for future work: leveraging structure-aware architectures over model internals for interpretability. In particular, architectures with different inductive biases (such as LEMON-W and LEMON-W-Per-Cell studied in this work) could help reveal where signals associated with a given misbehavior emerge, given suitable analysis and experimental design. For instance, if LEMON-W substantially outperforms LEMON-W-Per-Cell on a given task, this could indicate that the relevant signal is distributed across multiple heads and layers, making information sharing across them beneficial. Conversely, if LEMON-W-Per-Cell performs substantially better, this could suggest that the relevant signal is more localized and can be effectively exploited without sharing information across heads and layers. We therefore see a systematic interpretability analysis of these architectural differences as a promising avenue for future work.


References

[1] Understanding Deep Convolutional Networks; Mallat, 2016, Philosophical Transactions of the Royal Society A

[2] On Universal Equivariant Set Networks; Segol et. al. ICLR 2020

[3] Set Norm and Equivariant Skip Connections; Zhang et. al. ICML 2022

Thank you for your response. The rebuttal has adequately addressed most of my concerns. Therefore, I will raise my score accordingly.

AuthorsOpenReview ↗

Response to Reviewer yFk8

Thank you for your positive and thoughtful feedback. We’re happy that we were able to address your concerns, and we truly appreciate the increase in score.

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

This paper proposes LEMON, an attention-based framework for zero-shot misbehavior monitoring across different LLMs. By modeling cross-model attention mismatches through algebraic symmetries, LEMON uses an equivariant graph neural network to extract transferable features from attention maps. Experiments across 16 LLMs and six datasets demonstrate improvements over existing baselines in hallucination and jailbreak monitoring, with relatively low inference overhead.

Critical strengths
  1. The paper addresses an important practical problem: transferring misbehavior monitors to unseen LLMs without requiring additional labeled data.

  2. The symmetry-aware architecture is well motivated and provides an interesting approach to handling differences in attention heads and layers across models.

  3. The experiments are extensive, covering multiple model families and datasets. The ablation studies and online monitoring experiments further support the effectiveness of the proposed method.

Critical weaknesses
  1. The theoretical motivation relies partly on approximate symmetries. In particular, cyclic layer permutations are not exact symmetries of Transformer architectures, and the extent to which this assumption holds across different models remains unclear.

  2. The jailbreak evaluation is limited to a single universal GCG suffix. It is unclear whether the observed improvements generalize to more diverse or adaptive jailbreak attacks.

  3. Some hallucination annotations are generated automatically. Although the authors provide a manual audit, the reported annotation errors raise concerns about label reliability and their potential influence on training.

  4. The efficiency evaluation is relatively limited, focusing on a single model and sample. More comprehensive latency and memory measurements across different models and sequence lengths would strengthen the practical claims.

Questions for authors

Please refer to Weaknesses.

4 comments
AuthorsOpenReview ↗

Rebuttal Response to bYmx [part 1/2]

We thank the reviewer for the positive feedback. We are pleased that the reviewer recognizes the importance of the problem addressed in our paper, finds our symmetry-aware architecture well motivated, and appreciates the extensive experimental evaluation. Below, we address each of the points raised by the reviewer.


W1

We thank the reviewer for raising this important point. In our response, we first explain the motivation for using cyclic shifts from a structural perspective and then present empirical evidence supporting this modeling choice.

Treating translations as an approximate symmetry along a one-dimensional signal axis is well established [1]. For a finite sequence, such shifts can be represented cyclically, offering a convenient boundary convention without assuming that the underlying sequence itself is periodic. This viewpoint is particularly useful for designing structure-aware neural architectures, as in our work, where we use cyclic shifts to explicitly capture the sequential structure of transformer layers, consistent with related prior work [2].

We next provide empirical evidence supporting this choice. In particular, we refer the reviewer to the “LB+Align” baseline in Table 1. LB+Align uses the same LB features (shown in Figure 2), while additionally applying our alignment procedure described in Appendix E. This procedure follows directly from our structural model of attention maps: layers are modeled as a sequence, while heads are modeled as a set; see section 3.2 in our paper. Notably, LB+Align consistently outperforms LB, yielding an average AUC improvement of approximately 6 points on unseen language models. These results provide direct empirical support for the way we model the structure of layers and heads. We would be happy to emphasize this result more prominently and expand the corresponding discussion in the revised paper if the reviewer finds this helpful.


W2

We thank the reviewer for raising this interesting point. To demonstrate that our method learns general patterns associated with successful jailbreaks, rather than simply identifying requests that are more likely to succeed under a particular attack, we conducted the following experiment.

We introduced five additional suffixes that are different from the suffix used during training and evaluated our model directly on detecting jailbreaks induced by these new suffixes, without any training on them. Specifically, we randomly selected 100 test examples and attempted to jailbreak the model by appending each of the five new suffixes to every example.

The table below summarizes the results. The leftmost column reports the in-distribution (in-dist.) performance already presented in the paper, while the remaining columns report the average transfer detection performance (across the six LLMs considered) for each new suffix, none of which was seen during training.

Paper (in-dist.)sfx1 (transfer)sfx2 (transfer)sfx3 (transfer)sfx4 (transfer)sfx5 (transfer)Avg (transfer)
Attn→Sys58.550.657.259.352.965.058.0
CHARM50.158.053.453.459.446.654.6
HS ℓ2447.241.138.254.644.054.245.8
LEMON-𝒲65.058.454.060.365.867.561.3
LEMON-𝒲-per-cell63.360.458.763.561.767.063.3

For every unseen suffix, the best-performing method is one of our variants, and overall (rightmost column), our variants achieve the best and second-best average transfer performance. These results provide clear evidence that our LEMON model can be useful and go beyond one single suffix.

Finally, we thank the reviewer for this valuable suggestion, which we believe further strengthens our jailbreak experiments. Accordingly, we have revised the manuscript, with all changes relative to the original version highlighted in blue, and added a new appendix (Appendix E) providing further details on this experiment and its results.


W3

We understand the reviewer’s concern. However, we would like to draw the reviewer’s attention to Table 2, where we report results across 10 different LLMs and two datasets with examples annotated manually by humans rather than automatically generated. These results show that our methods perform well when evaluated on human annotations. Moreover, we expect that using cleaner, human-annotated training data would further improve the performance of our methods.


We continue with the remaining points in a separate response below.

AuthorsOpenReview ↗

Rebuttal Response to bYmx [part 2/2]

W4

We thank the reviewer for this important suggestion. We extended the run-time analysis from one model and one sample to all six LLMs in our pool, with a 1024-token prompt and response lengths $R \in { 256,512,1024,2048 } $. For this per-token analysis, we implement our monitor by caching the intermediate states of preceding nodes. Much like an LLM’s KV cache, and owing to the causal nature of both its inputs and its computation, our monitor only needs to compute the representation(s) of the latest generated tokens at each step (please refer to the last paragraph of section 3.3 for additional details). This optimization is purely at the implementation level: the monitor’s outputs are identical to those obtained without caching, yet, like the graph construction, the monitor can be updated incrementally at each step. We detail this implementation in Appendix I of the revised paper (in blue), along with the new run-time analysis. Below we report the main results.

For each response length RR, we report the unmonitored decoding step per token averaged over the six LLMs we consider (leftmost column) and then, for each LEMON variant, the added time broken down into attention extraction, graph update and critic, followed by the overhead (relative to the unmonitored step) and the extra peak memory. Both the overhead and the memory rise slightly as RR grows, as expected, but remain small: on average, the overhead stays below 10% for both variants, even at the longest response length, R=2048R=2048.

RRUnmonitored step (ms)VariantAttn. (ms)Graph (ms)Critic (ms)Overhead (%)Mem. (GB)
25627.51LEMON-W0.810.400.638.00.15
LEMON-W-per-cell0.810.400.156.1
51227.63LEMON-W0.810.400.627.90.18
LEMON-W-per-cell0.810.400.146.0
102427.86LEMON-W0.870.420.628.20.29
LEMON-W-per-cell0.870.420.146.3
204828.36LEMON-W1.020.440.699.00.48
LEMON-W-per-cell1.020.440.166.9

References

[1] Understanding Deep Convolutional Networks; Mallat, 2016, Philosophical Transactions of the Royal Society A

[2] Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT; Bar-Shalom el. al. NeurIPS 2025

Thanks for you response. Most of my concerns have been addressed, so I maintain my original positive rating.

AuthorsOpenReview ↗

Response to Reviewer bYmx

We are thankful for your constructive feedback and positive recommendation.

Reviewer_1AxBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce LEMON, a GNN-based framework for zero-shot misbehavior monitoring across LLMs. The authors address the problem related to transferring internal representation-based monitors to unseen LLMs where attention patterns associated with misbehaviour appear across different layer-head coordinates. The authors formulate the cross-modal dissimilarities using algebraic symmetries through attention-head permutations and cyclic layer shift. Based on this characterization, the proposed approach develops attention-based graphs with structured layer-head features and deploys equivariant message passing followed by an invariant prediction head. The proposed framework supports online monitoring during generation and does not require labelled data or retraining target LLMs. The proposed method has been evaluated on hallucination detection and jailbreak across 16 LLMs ( 6 leave-one-out, 10 additional zero-shot unseen LLMs). The results show improvements over baselines, along with relatively low monitoring overhead.

Critical strengths
  1. The paper addresses an important problem in LLM misbehaviour monitoring. The proposed method transfers detectors to unseen LLMs based on internal model representations and it does not require labelled data or retraining the detector model.

  2. The authors propose an interesting methodological approach that targets mismatches in attention-head ordering and layer indexing across LLMs. It incorporates equivariant message passing into attention-based graph representations for improving cross-model generalization.

  3. The paper provides extensive experiments across 16 LLMs , 6 datasets, covering jailbreak monitoring and hallucination detection. LEMON outperforms the baselines considered in this paper. The ablation studies further investigate the contributions of different components, including equivariance, graph structure, and token probabilities.

  4. LEMON supports online monitoring during generation and showcases low computational overhead, make it suitable for easy deployment

Critical weaknesses
  1. In the paper the authors model the mismatches using attention-head permutations and cyclic layer shifts. However, this assumption isn't sufficiently justified. Permuting heads preserves underlying computation, which is justified; however, the same can't be stated for layers; as input of one layer goes to the next and has a direct impact on successive layers. The authors acknowledge that the cyclic layer shift is only an approximation, and that the assumption reflects how independently trained LLMs actually differ. In Figure 2, the authors show that the attention patterns associated with hallucinations appear at different layer-head positions in Llama-2-7B and Mistral-7B. However, this observation alone does not establish that these differences can be explained by the proposed head permutations and cyclic layer shifts. Further evidence supporting this assumption would strengthen the theoretical motivation.

  2. The advantage of the proposed symmetry structure is not sufficiently established. The ablation in Table 4 shows that invariance is important for cross-model generalization but it does not establish the advantage of the specific symmetry structure proposed in the paper.. However, LEMON-W-per-cell performs better than LEMON-W on hallucination detection (69.4 vs. 67.8), even though it processes each layer-head coordinate independently and then averages the features. This former is invariant to any reordering of layer-head coordinates. So, it is not clear whether the improvements come from the proposed symmetry structure or simply from shared feature processing and averaging. Although LEMON-W performs slightly better on jailbreak detection (65.0 vs. 63.3), the results are mixed.

  3. The jailbreak evaluation is limited to a single attack. The authors use the same universal GCG suffix for all harmful requests across the six LLMs. As the monitor only observes the prompt; and these prompts contain the same suffix. It is not clear whether the method learns general patterns of successful jailbreaks or rather it identifies requests that are more likely to succeed with this particular attack. The authors should evaluate the method on different jailbreak attacks; especially attacks unseen during training. It would also be useful to compare against a simple text-based classifier that predicts attack success from the original harmful request. This would help establish whether the proposed method captures additional information beyond the request content.

Questions for authors
  1. Can the authors provide empirical evidence that attention patterns across different LLMs are related through the proposed head permutations and cyclic layer shifts? And how well do these transformations align the patterns shown in Fig. 2?

  2. As shown in ablation LEMON-W-per-cell performs better than LEMON-W on hallucination detection; can the authors clarify what advantages the specific symmetry structure provides over shared per-cell processing and invariant averaging?

  3. How well does LEMON generalize to jailbreak attacks that were not seen during training?

  4. How does LEMON compare with a simple text-based classifier trained to predict attack success using only the original harmful request? Does the attention information provide a significant improvement over this baseline?

4 comments
AuthorsOpenReview ↗

Rebuttal Response to 1AxB [part 1/2]

We thank the reviewer for the positive feedback. We are pleased that the reviewer recognizes the importance of the problem addressed in our paper, the value of our methodological approach, the extensive experimental evaluation, and the ability of our method to support online monitoring with low computational overhead. Below, we address each of the points raised by the reviewer.


W1:

We thank the reviewer for raising this important point. In this response, we will first motivate our use of cyclic shifts from a structural perspective, and then provide empirical evidence supporting this modeling choice.

Treating translations as an approximate symmetry along a one-dimensional signal axis is well established [1], and, for a finite sequence, these shifts can be represented cyclically, providing a convenient boundary convention without requiring the underlying sequence itself to be periodic. Importantly, this perspective is useful for designing structure-aware neural architectures, as we did in this work, where we considered cyclic shifts to specifically model the sequentiality in transformer layers as also done in prior work [2].

We now provide empirical evidence supporting our modeling choice. We refer the reviewer to the “LB+Align” baseline in Table 1. LB+Align uses the same LB features visualized in Figure 2, but applies our alignment procedure from Appendix E. This alignment procedure follows directly from our structural model of attention maps: layers are treated as a sequence, whereas heads are treated as a set. Importantly, LB+Align consistently outperforms LB, improving AUC by approximately 6 points on average on unseen language models. This provides direct empirical evidence that explicitly accounting for the structural relationships among layers and heads improves cross-model generalization, supporting the modeling choices underlying our approach. We would be happy to highlight this result more prominently and expand the corresponding discussion in the revised paper, if the reviewer finds it helpful.


W2:

First, we would like to emphasize that both variants, LEMON-W and LEMON-W-Per-Cell, arise from the structure we identify in the tensor of attention maps across layers and heads (section 3.2 in the paper). In particular, we believe there is value in deriving the general architecture that respects the full symmetry (LEMON-W) and then studying meaningful special cases of it (LEMON-W-Per-Cell). Different misbehavior-monitoring tasks may favor different inductive biases. As noted by the reviewer, in our experiments, LEMON-W-Per-Cell performs best for hallucination detection, whereas LEMON-W performs best for jailbreak detection. Thus, rather than viewing one variant as universally preferable, our results suggest that the appropriate degree of information sharing may depend on the monitoring task.

Additionally, we note that considering restricted special cases of general equivariant architectures is common in the literature. For example, [3] studies permutation-equivariant set architectures using both a full architecture with global information sharing and a restricted element-wise variant, analogous to the relationship between LEMON-W and LEMON-W-Per-Cell (despite the different symmetry setting). Moreover, simpler and more constrained variants can perform better in practice. For example, in a different setting and task, [4] finds that preserving information locally, rather than combining information across the entire representation, consistently leads to better performance in their experiments. Our hallucination experiments with LEMON-W-Per-Cell demonstrate a similar phenomenon.


We continue with the remaining points in a separate response below.

AuthorsOpenReview ↗

Rebuttal Response to 1AxB [part 2/2]

W3

We thank the reviewer for pointing out this interesting point. To show that our method indeed learns the general patterns of successful jailbreaks rather than identifies requests that are more likely to succeed with a single particular attack, we designed the following experiment.

We introduced five additional suffixes that are different from the suffix used during training and evaluated our model directly on detecting jailbreaks induced by these new suffixes, without any training on them. Specifically, we randomly selected 100 test examples and attempted to jailbreak the model by appending each of the five new suffixes to every example.

The table below summarizes the results. The leftmost column reports the in-distribution (in-dist.) performance already presented in the paper, while the remaining columns report the average transfer detection performance (across the six LLMs we consider) for each new suffix, none of which was seen during training.

Paper (in-dist.)sfx1 (transfer)sfx2 (transfer)sfx3 (transfer)sfx4 (transfer)sfx5 (transfer)Avg (transfer)
Attn→Sys58.550.657.259.352.965.058.0
CHARM50.158.053.453.459.446.654.6
HS47.241.138.254.644.054.245.8
LEMON-𝒲65.058.454.060.365.867.561.3
LEMON-𝒲-per-cell63.360.458.763.561.767.063.3

For every unseen suffix, the best-performing method is one of our variants, and overall (rightmost column), our variants achieve both the best and second-best performance. These results provide clear evidence that our LEMON model can be useful and go beyond one single suffix.

Finally, we thank the reviewer for this valuable suggestion, which we believe further strengthens the jailbreak experiments in our paper. Accordingly, we have revised the manuscript (with all changes relative to the original version highlighted in blue) and added a new appendix, Appendix E, providing further details on this experiment and its results. We also refer the reviewer to our response to Q4, where we address his comment regarding the text-based classifier.


Q1

We kindly refer the reviewer to the final paragraph our answer to W1 from above.


Q2

This is an interesting question, and we believe it opens up an important direction for future work: using structure aware internal-based architectures for interpretability. In particular, structure-aware architectures with different inductive biases - such as LEMON-W and LEMON-W-Per-Cell considered in this work- may help identify where the signal associated with hallucination originates, given appropriate analysis and experimental design. For example, if LEMON-W substantially outperforms LEMON-W-Per-Cell on a particular task, this may suggest that the relevant signal is distributed across multiple heads and layers, such that sharing information across them is important for effectively exploiting it. Conversely, if LEMON-W-Per-Cell substantially outperforms LEMON-W, this may suggest that the relevant signal can be exploited locally, without requiring information sharing across heads and layers. We therefore view a systematic interpretability analysis of these architectural differences as a promising direction for future work.


Q3

We kindly refer the reviewer to our answer to W3 from above, as we showed that LEMON generalizes well to jailbreak attacks that were not seen during training.


Q4

We thank the reviewer for the question, but we believe a text-based classifier is not a suitable baseline for our setup; let us clarify. Consider a prompt (PP) that successfully jailbreaks LLM A but not LLM B: a text-based encoder receives exactly the same input (PP) for both LLMs and therefore cannot possibly distinguish between these two outcomes. In contrast, our approach relies on model internals, which differ across LLMs even for the same prompt (PP).

More concretely, the performance of a text-based classifier in our cross-LLM setting would, thus, tend to reflect the extent to which different LLMs are vulnerable to the same prompts, rather than its ability to capture model-specific signals associated with jailbreak success.


References

[1] Understanding Deep Convolutional Networks; Mallat, 2016, Philosophical Transactions of the Royal Society A

[2] Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT; Bar-Shalom el. al. NeurIPS 2025

[3] On Universal Equivariant Set Networks; Segol et. al. ICLR 2020

[4] Set Norm and Equivariant Skip Connections; Zhang et. al. ICML 2022

Reviewer_1AxBOpenReview ↗

Reviewer Comment

Overall, I appreciate the authors' efforts and the additional results. While the rebuttal partially addresses my concerns, the central methodological questions remain. I therefore maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer 1AxB

We sincerely thank the reviewer for the positive feedback and constructive comments. We are glad that our revisions have addressed some of your concerns. For any points that could benefit from further clarification, we would greatly appreciate it if you could let us know, and we would be happy to address them further.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their positive and constructive feedback. We are pleased that all reviewers expressed a positive assessment of our work and supported the paper’s acceptance, with each assigning a score of 3 by the end of this rebuttal process. The reviewers acknowledged that our work addresses a well-motivated and important problem:

  • “The paper addresses an important problem…” (1AxB)
  • “The paper addresses an important practical problem…” (bYmx)
  • “The symmetry-aware architecture is well motivated..” (bYmx)
  • “Interesting motivation. The paper offers a valuable perspective.. The connection to algebraic symmetries provides a principled motivation” (yFk8)

The reviewers also highlighted our extensive experiments, promising results, low computational overhead, and interesting methodological approach:

  • “The paper provides extensive experiments across 16 LLMs , 6 datasets…” (1AxB)
  • “The experiments are extensive, covering multiple model families and datasets..“(bYmx)
  • “The authors propose an interesting methodological approach…”  (1AxB)
  • “Promising cross-model generalization. LEMON consistently improves over prior baselines… “ (yFk8)
  • “LEMON supports online monitoring during generation and showcases low computational overhead..” (1AxB)

Posterior to our rebuttal, reviewer yFk8 raised their score from 2 to 3, and we believe that we addressed the reviewers’ main concerns while clarifying some key aspects of our work.

We sincerely thank all reviewers for their thoughtful feedback and fruitful discussion, which helped us clarify and strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.