acceptodds
Under review as a conference paper at ICLR 2027

Risk Exposure Does Not Mean Unsafe Action: Evidence-Dynamics Alignment for Agent Safety

Abstract

Large language model (LLM)-based agents increasingly operate through multi-step interactions with external environments and tools, creating new challenges for trajectory-level safety monitoring. A key difficulty is that risk exposure does not necessarily imply unsafe action: an agent may encounter risky contexts while ultimately behaving safely, yet existing LLM-based monitors can conflate the two and produce false alarms. We find that this failure does not simply arise from missing safety information. Frozen monitor representations retain distinguishable evidence of risk exposure and unsafe enactment, but such evidence is not fully reflected in native decisions. Based on this observation, we propose Evidence-Dynamics Alignment (EDAlign), a lightweight post-hoc framework that tracks Exposure and Enactment evidence across model depth and aligns final safety decisions with their evolving evidence state. Experiments across agent safety benchmarks and LLM backbones show that EDAlign reduces exposure-induced false alarms while maintaining strong detection of unsafe behavior. Comparisons with direct unsafe probes and layer-range ablations further support the value of structured Exposure–Enactment evidence and its evolution across model depth.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-based agent safety monitoring, safety guard, LLM representation analysis, safety alignment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_3BYSOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the distinction between risk exposure and unsafe enactment in LLM-based agent safety monitoring. The authors propose Evidence-Dynamics Alignment (EDAlign), a post-hoc framework that extracts Exposure and Enactment evidence from intermediate representations of frozen safety monitors. These signals are aggregated across model layers into a structured feature representation and leveraged by a lightweight alignment head to refine trajectory-level safety predictions. Experiments on ATBench and ASSEBench across four LLM backbones show improvements in safety classification accuracy and F1, particularly by reducing false positives on risk-exposed but behaviorally safe trajectories. Additional analyses examine evidence recoverability, prediction errors, and feature ablations.

Critical strengths
  1. The paper highlights an important distinction between encountering risky information and actually performing unsafe actions. This perspective is meaningful for trajectory-level safety monitoring.
Critical weaknesses
  1. Limited methodological novelty and insufficiently controlled comparisons. As described in Sections 3, EDAlign primarily combines supervised linear probes, handcrafted statistical features, and a lightweight post-hoc classification head. Although the Exposure-Enactment decomposition is intuitively reasonable, the approach does not directly modify or causally align the internal decision process. Furthermore, the direct unsafe-probe baseline in Appendix B.8 adopts a 10-dimensional feature representation, compared with the 24-dimensional representation of EDAlign. Without stronger capacity-matched comparisons, it remains unclear whether the improvements arise from the proposed evidence factorization or simply from richer supervised features.
  2. Model-dependent effectiveness and insufficient justification of the feature design. The results in Table 1 show that EDAlign does not consistently outperform the baselines across different backbones and datasets. In particular, for Llama3.1-8B on ATBench, EDAlign achieves an F1 of 67.75%, below 72.93% of AgentDoG. More importantly, Table 2 shows that removing any individual feature group improves ATBench F1 for Llama3.1-8B, with a substantial 7.02% improvement when removing evidence endpoints. Although Section 5.4 acknowledges model-dependent feature contributions, it provides limited explanation for this behavior. These findings raise concerns about the necessity and robustness of the proposed structured evidence representation across different models and datasets.
Questions for authors
  1. Could the authors provide a capacity-matched comparison between EDAlign and direct unsafe probing, including comparable feature dimensionality and classification-head complexity? This would help establish the independent contribution of Exposure-Enactment factorization.
  2. Table 1 shows that EDAlign underperforms AgentDoG for Llama3.1-8B on ATBench, while Table 2 indicates that removing any individual feature group improves its F1, particularly evidence endpoints (+7.02 points). How do the authors explain these model-specific results, and what do they imply about the necessity and generalizability of the proposed feature design?
3 comments

W1&Q1: Capacity-matched comparison between EDAlign and direct unsafe probing

We thank the reviewer for this important suggestion. U-Probe trains an independent binary unsafe probe at each layer. Since it models a single unsafe evidence stream, applying our structured feature construction naturally yields a 10D representation. EDAlign instead learns two probe families: Exposure (D00 vs. D10) and Enactment (D10 vs. D11), yielding 24D features that capture both streams, their discrepancies, interactions with the native decision, and cross-layer dynamics. Thus, the dimensionality difference reflects the evidence structures modeled, rather than an arbitrary restriction of U-Probe.

To address this concern, we introduce U-Probe-Dynamics (24D). It retains a single unsafe probe per layer and aggregates its layer-wise scores SlS_l into 24 features: native-decision features, three anchor-point scores, segment-wise changes, global/segmented path statistics, and native-decision/evidence interactions. It uses the same middle-to-final layer range and classification head as EDAlign, but has no separate Exposure/Enactment channels.

Table X. Capacity-matched comparison on ATBench (%, ↑ higher is better except FPR):

ModelMethodACCRecallFPRF1
Qwen3-4BU-Probe-Dynamics82.0071.724.6181.84
EDAlign85.1484.8514.4786.60
Qwen3-8BU-Probe-Dynamics79.7172.2210.5380.11
EDAlign84.5780.309.8785.48
Qwen3.5-4BU-Probe-Dynamics86.5784.3410.5387.66
EDAlign87.4389.3915.1388.94
Llama3.1-8BU-Probe-Dynamics69.1445.450.0062.50
EDAlign71.7152.533.2967.75

EDAlign consistently outperforms U-Probe-Dynamics on ATBench, with F1 gains of 1.28–5.37 points (mean: 4.17).

We further evaluate cross-benchmark transfer from ATBench to ASSEBench. Probes are trained on ATBench and frozen; 80 ASSEBench samples recalibrate the alignment head, and the remaining 682 are used for testing.

Table Y. Cross-benchmark transfer on ASSEBench (%, ↑ higher is better except FPR):

ModelMethodACCRecallFPRF1
Qwen3-4BU-Probe-Dynamics84.3192.5723.1284.82
EDAlign90.0385.766.1389.07
Qwen3-8BU-Probe-Dynamics78.0180.8024.5177.68
EDAlign89.0091.3313.0988.72
Qwen3.5-4BU-Probe-Dynamics86.3695.0521.4586.85
EDAlign90.9192.2610.3190.58
Llama3.1-8BU-Probe-Dynamics85.9289.4717.2785.76
EDAlign85.1988.2417.5584.95

EDAlign outperforms U-Probe-Dynamics on three of four backbones, with F1 gains of 3.73–11.04 points; U-Probe-Dynamics is 0.81 points better on Llama3.1-8B. Across eight settings, EDAlign wins 7/8 with a mean F1 gain of 4.36 points. These results show that the gains are not solely due to higher dimensionality and support the benefit of Exposure–Enactment factorization under limited target-domain alignment data.

W2&Q2: Model-specific results and generalizability of feature design

We thank the reviewer for raising this concern. The observed differences reflect distinct native-monitor biases and backbone-dependent internal safety evidence.

On Llama3.1-8B/ATBench, AgentDoG has higher Recall than EDAlign (82.32% vs. 52.53%) but much higher FPR (56.58% vs. 3.29%), yielding F1 of 72.93% vs. 67.75%. We acknowledge this trade-off: EDAlign favors substantially lower FPR.

More broadly, Vanilla Qwen3-4B/8B have high FPRs (79.61%/55.92%), indicating over-flagging of safe trajectories, whereas Llama3.1-8B has low Recall (38.89%), indicating missed unsafe behaviors. Relative to Vanilla, EDAlign reduces FPR while increasing Recall for Qwen, and improves both Recall and FPR for Llama. Thus, EDAlign can correct different native decision biases rather than simply shifting predictions in one direction.

Performance also depends on backbone-specific evidence distributions. For Llama3.1-8B, All-layer evidence improves F1 over Mid-to-final by 0.45 points on ATBench and 2.32 on ASSEBench, suggesting useful evidence in earlier layers. Feature ablations further show dataset-dependent effects: removing evidence endpoints improves ATBench F1 by 7.02 points, whereas removing path features reduces ASSEBench F1 by 2.79 points.

Overall, EDAlign can mitigate different native-monitor biases through post-hoc alignment of internal safety evidence, but its final performance remains dependent on backbone-specific evidence, feature interactions, and the resulting decision boundary and threshold. We therefore do not claim every feature group is universally beneficial or that EDAlign consistently outperforms every baseline.

AuthorsOpenReview ↗

Looking forward to your reply and further discussion.

Looking forward to your reply and further discussion.

Thank you for the detailed rebuttal and additional experiments.

The newly introduced U-Probe-Dynamics (24D) baseline provides stronger evidence for the effectiveness of Exposure-Enactment factorization, partially addressing my concerns. However, the explanation of the model-dependent ablation results remains insufficient, particularly why removing individual feature groups consistently improves performance for Llama3.1-8B on ATBench. The proposed framework still appears methodologically incremental, and the robustness of its feature design remains questionable.

Additionally, I noticed that the initial submission already exceeds the 9-page main-text limit, with the Conclusion extending onto page 10, and no corresponding revisions to the main text have been made during the rebuttal period. Considering these concerns, I maintain my original rating of 1/4.

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

This paper studies the conflation of risk exposure with unsafe behavior in trajectory-level safety monitoring for LLM agents. The authors show that frozen monitor representations contain separately decodable information about risk exposure and unsafe enactment, and thus propose Evidence-Dynamics Alignment (EDAlign), which learns a lightweight post-hoc classifier while keeping the underlying monitor frozen. Experiments on ATBench and ASSEBench show substantial reductions in false positives relative to native monitors and generally improved F1 scores.

Critical strengths
  1. The problem of risk exposure and enactment deserves investigated.
  2. The paper is generally well-organized to understand.
  3. The proposed method is lightweight.
Critical weaknesses
  1. The papers states that internal representations could distinguish risk exposure and enactment, but no interpertation cases and results are presented, which makes the motivation of method seem weak.

  2. For ATBench and ASSEBench,their statistics towards risk exposure and enactment are not clear, nor descirbed how the instances are divided as risk exposure and enactment.

  3. The comparision is not suffient, where in Table-1, typical agent guard like AgentDog, Safiron, are not compared and analyzed.

Questions for authors

Pleae refer to the part of Weakness.

3 comments

W1: Lack of interpretable cases and supporting results We thank the reviewer for this helpful comment. We would like to clarify that the original manuscript already includes a paired case study in Appendix A (Figure 6 and Table 4), although we agree that it was not sufficiently highlighted in the main text. The case study contrasts two trajectories with different behavioral outcomes. In Case 851, the agent detects suspicious document modifications, stops processing, and requests a fresh document (SAFE). In Case 438, the agent recognizes that the retrieved information is unrelated to its task but nevertheless invokes a bank-slip retrieval tool outside the authorized scope (UNSAFE). Notably, all four native monitors classify both trajectories as UNSAFE. EDAlign corrects the false alarm on Case 851 for Qwen3-8B and Llama3.1-8B, while retaining the correct UNSAFE prediction for Case 438 across all four backbones. We further examine the internal evidence of Qwen3-8B. At the final layer, Case 851 exhibits an Exposure score of 1.223 and an Enactment score of -0.662, whereas Case 438 exhibits scores of -0.288 and 1.910, respectively. Accordingly, EDAlign assigns unsafe probabilities of 0.2454 and 0.9775 to these cases, successfully distinguishing their behavioral outcomes. This case study illustrates how internal evidence can distinguish risk exposure from actual unsafe enactment, and how EDAlign can recover this distinction when it is not reflected in the native monitor's decision. We will move this case study into the main text and provide a more explicit discussion of its implications for trajectory-level safety monitoring.

W2:Unclear dataset statistics and cohort construction We thank the reviewer for identifying this lack of clarity. For ATBench, we leverage its original fine-grained annotations to construct three behavioral cohorts based on the combination of behavioral labels and risk-source annotations:

  • D00 (Benign-Safe): (y=0) with a benign risk source, representing trajectories without risk exposure and with safe behavior.
  • D10 (Exposed-Safe): (y=0) with a non-benign risk source, representing trajectories exposed to risky information but exhibiting safe behavior.
  • D11 (Unsafe): (y=1), representing trajectories in which the agent exhibits unsafe behavior. These cohorts are used to train two distinct probes: Exposure is trained on D00 vs. D10, while Enactment is trained on D10 vs. D11. Each cohort contains 100 trajectories in the ATBench training set. For ASSEBench, we do not construct the above three cohorts or retrain the probes. Instead, we use 80 examples (40 safe and 40 unsafe) as an alignment set to fit the alignment head after transferring the Exposure and Enactment probes trained on ATBench in a frozen manner. The remaining 682 examples are reserved for independent testing. We will clarify these dataset construction procedures and their respective roles in the revised manuscript.

W3:Comparison with agent safety guardrails We appreciate the suggestion to include representative agent safety guardrails. We would like to clarify that AgentDoG1.5 is already included in Table 1 as the fine-tuned reference for Qwen3.5-4B and Llama3.1-8B. For Qwen3-4B and Qwen3-8B, we use the corresponding Qwen3Guard models.In particular, while EDAlign outperforms AgentDoG1.5 on Qwen3.5-4B across both benchmarks, its F1 on Llama3.1-8B is lower on ATBench (67.75% vs. 72.93%). However, EDAlign substantially reduces FPR in this setting (3.29% vs. 56.58%), reflecting a different safety–false-alarm trade-off. On ASSEBench, EDAlign achieves a higher F1 (84.95% vs. 79.17%) and a lower FPR (17.55% vs. 40.56%).

AuthorsOpenReview ↗

Looking forward to your reply and further discussion.

Looking forward to your reply and further discussion.

Reviewer_GXyMOpenReview ↗

Thanks for your responses

Thanks for your responses, but my concerns have not been completely addressed or resolved. I maintain my original score.

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

This paper proposes Evidence-Dynamics Alignment (EDAlign), a lightweight post-hoc framework for distinguishing risk exposure from unsafe enactment in LLM-agent trajectories. It extracts layer-wise Exposure and Enactment evidence using frozen linear probes and integrates them through a integrated feature representation and a lightweight prediction head. Experiments on ATBench and ASSEBench across four LLM backbones show improved safety-monitoring performance, supported by baseline comparisons, statistical tests, and ablation studies.

Critical strengths
  1. The paper distinguishes risky context from actual unsafe behavior, addressing a meaningful source of false alarms in agent safety monitoring.
  2. The paper proposes a simple and straightforward approach to improve the effectiveness of judge models, which is reasonable and easy to follow.
Critical weaknesses
  1. Integrating embeddings from multiple attention layers of LLMs to identify unsafe info is a widely used and common approach IMHO. Thus, the contribution of EDAlign seems limited, though the author designs several integration tricks, as in "3.2 STRUCTURED EVIDENCE INTEGRATION REPRESENTATION".
  2. I suggest the author investigate and study more existing works from the field "INTERNAL SAFETY REPRESENTATION ANALYSIS", because i notice that the literature review is not sufficient, especially the content in "2.3 INTERNAL SAFETY REPRESENTATION ANALYSIS".
  3. I suggest the author provide more ablation studies, e.g., investigating the role of layer selection, the training data selection, etc.
Questions for authors
  1. Is the result of EDAlign sensitive to layer selection?
  2. How robust are the ASSEBench results to alignment-set selection and observable refusal cues?
3 comments

W1: Limited methodological contribution

We agree that probing multi-layer LLM representations and aggregating layer-wise safety signals are established techniques; we do not claim multi-layer feature extraction itself as novel. EDAlign's contribution lies in the task-specific decomposition and alignment of trajectory-level safety evidence. It distinguishes (i) risk exposure, where an agent encounters non-benign information but behaves safely, from (ii) unsafe enactment, where the agent performs or attempts an unsafe action. Accordingly, EDAlign trains two independently supervised probe families using D00 vs. D10 for Exposure and D10 vs. D11 for Enactment. It then models their discrepancy, interactions with the native monitor decision, and evolution across model depth. Thus, the contribution is the factorized supervision and structured alignment of two distinct forms of evidence, rather than layer-wise aggregation alone.

W2: Insufficient literature review

We thank the reviewer for this suggestion and agree that Section 2.3 needs a broader discussion. We will expand it to cover: (1) refusal-related representations and safety directions (Arditi et al., 2024; Pan et al., 2025; Wollschläger et al., 2025); (2) representations of harmfulness, refusal, and jailbreak-related concepts (Zheng et al., 2024; Zhang et al., 2025; Zhao et al., 2025); and (3) multi-layer internal safety detection (Jiao et al., 2026). We will compare these works in terms of research objectives, supervision signals, and representation-analysis methods, and clarify that EDAlign focuses on distinguishing risk exposure from unsafe enactment in agent trajectories, rather than solely detecting harmful prompts, refusal mechanisms, or jailbreak-related concepts.

W3/Q1: Sensitivity to layer selection

We have examined layer sensitivity in two ways. First, Figure 3 reports layer-wise AUROC for Exposure and Enactment probes: Exposure remains decodable across layers, whereas Enactment varies more in early and middle layers and becomes more stable in deeper layers. Second, Table 3 compares Final-only, Late-to-final, Mid-to-final, and All configurations. Mid-to-final outperforms Final-only across all four backbones and both benchmarks (all eight settings), while the relative ranking among Late-to-final, Mid-to-final, and All varies by backbone and dataset. These results indicate that layer selection matters, but incorporating intermediate-layer evidence consistently benefits EDAlign over using only the final layer. We use Mid-to-final as a practical trade-off between predictive performance and layer-readout cost, not as a universally optimal choice.

W3/Q2: Alignment-set selection and observable refusal cues

To examine sensitivity to alignment-set selection on ASSEBench, we conduct 40 repeated runs. In each run, 80 trajectories are randomly selected for alignment and the remaining 682 are used for testing. EDAlign achieves the highest mean F1 across all four backbones.

Mean F1 (%) over 40 runs (higher is better):

\text{Backbone} & \text{Vanilla} & \text{ScoreCal.} & \text{Guard} & \text{U-Probe} & \text{SIREN} & \text{EDAlign} \\\\ \hline \text{Qwen3-4B} & 70.53 & 70.28 & 44.70 & 79.83 & 81.82 & \mathbf{84.53} \\\\ \text{Qwen3-8B} & 76.44 & 76.15 & 45.49 & 77.92 & 84.26 & \mathbf{85.63} \\\\ \text{Qwen3.5-4B} & 69.36 & 80.57 & 85.62 & 82.71 & 75.50 & \mathbf{89.30} \\\\ \text{Llama3.1-8B} & 72.13 & 73.95 & 79.84 & 81.34 & 78.55 & \mathbf{83.36} \end{array}$$ EDAlign outperforms the strongest baseline by 1.37–3.68 F1 points and U-Probe by 2.02–7.71 points across the four backbones. Note that Guard uses Qwen3Guard for the first two backbones and AgentDoG v1.5 for the latter two. For observable refusal cues, Appendix C.5 (Table 14) analyzes 87 ATBench trajectories with explicit refusal expressions (73 behavior-safe and 14 behavior-unsafe). The Enactment probes achieve AUROC values of 0.878–0.943 across four backbones. EDAlign makes only 8–9 prediction errors, compared with 14 for a refusal-only baseline, suggesting that it captures behavioral evidence beyond the mere presence of explicit refusal expressions. **Additional references:** Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction," NeurIPS 2024; Zheng et al., "On Prompt-Driven Safeguarding for Large Language Models," ICML 2024; Zhang et al., "JBShield," USENIX Security 2025; Pan et al., "The Hidden Dimensions of LLM Alignment," ICML 2025; Wollschläger et al., "The Geometry of Refusal in Large Language Models," ICML 2025; Zhao et al., "LLMs Encode Harmfulness and Refusal Separately," NeurIPS 2025; Jiao et al., "LLM Safety From Within," ACL 2026.
AuthorsOpenReview ↗

Looking forward to your reply and further discussion.

Looking forward to your reply and further discussion.

AuthorsOpenReview ↗

Looking forward to your reply and further discussion

Looking forward to your reply and further discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Discussion, Additional Evidence, and Remaining Issues

Summary of the Discussion, Additional Evidence, and Remaining Issues

We thank the reviewers for their feedback. During rebuttal, we clarified EDAlign's methodology and dataset construction and conducted additional experiments on feature dimensionality, robustness, and generalizability. We also clarify where reviews overlooked analyses or baselines in the original submission.

Reviewer uGLs — Layer Sensitivity, Literature, and Robustness

Layer selection. Figure 3 reports layer-wise AUROC, and Table 3 compares Final-only, Late-to-final, Mid-to-final, and All. Mid-to-final outperforms Final-only across all four backbones and both benchmarks (8/8 settings), showing consistent benefits from intermediate-layer evidence.

Alignment-set robustness. We conducted 40 ASSEBench runs, each randomly selecting 80 trajectories for alignment and using 682 for testing. EDAlign achieves the highest mean F1 across all four backbones, exceeding the strongest baseline by 1.37–3.68 points and U-Probe by 2.02–7.71 points.

Observable refusal cues. Appendix C.5 (Table 14) analyzes 87 ATBench trajectories with explicit refusal expressions (73 behavior-safe, 14 behavior-unsafe). EDAlign makes 8–9 errors versus 14 for a refusal-only baseline; Enactment probes achieve AUROC of 0.878–0.943 across four backbones, indicating evidence beyond explicit refusal cues.

Literature positioning. We acknowledged the need for broader coverage of refusal representations, harmfulness/jailbreak concepts, and multi-layer probing. We clarified that EDAlign targets the distinction between risk exposure and unsafe enactment in agent trajectories.

Reviewer GXyM — Interpretability, Dataset Construction, and Baselines

Interpretability. The claim that no cases were presented overlooks Appendix A (Figure 6 and Table 4), which already illustrates risky-but-safe behavior versus actual unsafe actions. We clarified these cases and acknowledged that their appendix placement made them less visible.

Dataset construction. ATBench comprises D00D_{00} (benign/safe), D10D_{10} (risk-exposed/safe), and D11D_{11} (unsafe) trajectories. For ASSEBench, ATBench-trained probes remain frozen; 80 samples recalibrate the alignment head and 682 are reserved for testing.

Baseline comparisons. The claim that agent-guard baselines were absent does not reflect Table 1. AgentDoG v1.5 was included for Qwen3.5-4B and Llama3.1-8B, while Qwen3Guard was used for Qwen3-4B and Qwen3-8B. AgentDoG covers the requested agent-guard baseline category; thus, the claim that this category was entirely absent is inaccurate.

Reviewer 3BYS — Capacity Matching, Feature Robustness, and Methodological Contribution

Capacity-matched comparison. Original U-Probe yields 10D features because it models one unsafe evidence stream, whereas EDAlign separately models Exposure and Enactment. We introduced U-Probe-Dynamics (24D), retaining one unsafe probe with a 24D structured representation and the same classification-head architecture as EDAlign.

EDAlign achieves higher F1 in 7/8 backbone–dataset settings (mean gain: 4.36 points). In ATBench-to-ASSEBench transfer, it also outperforms U-Probe-Dynamics on three of four backbones. The reviewer acknowledged that this experiment provides stronger evidence and partially addresses the concern. These results show that EDAlign's gains are not solely due to higher feature dimensionality.

Model-dependent feature contributions. Vanilla Qwen3 monitors tend to over-flag safe trajectories, whereas Llama3.1-8B tends to miss unsafe behaviors. EDAlign can mitigate both biases, although performance remains backbone-dependent. For Llama3.1-8B, All-layer improves F1 over Mid-to-final by 0.45 points on ATBench and 2.32 on ASSEBench. However, removing evidence endpoints improves ATBench F1 by 7.02 points, while removing path features reduces ASSEBench F1 by 2.79 points. The precise mechanism remains unverified and warrants further study.

Methodological contribution. We do not claim multi-layer probing or feature aggregation as novel. EDAlign contributes task-specific Exposure–Enactment decomposition, distinct supervision signals, and structured alignment of complementary trajectory-level evidence. The reviewer still questions its novelty, while the capacity-matched results provide additional empirical support for the decomposition.

Page-limit compliance. The initial submission exceeded the 9-page main-text limit. We acknowledge this and have revised the manuscript draft to comply.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.