acceptodds
Under review as a conference paper at ICLR 2027

Maps of Failures: Observed LLM Vulnerabilities in Behavioral Descriptor Spaces

Abstract

Safety evaluation often reduces a red-teaming run to attack success rate or a set of successful prompts. We instead study the distribution of discovered failures over an explicitly chosen behavioral descriptor space. Our framework uses MAP-Elites to construct a finite-budget archive in which each occupied cell records the highest Alignment Deviation observed for prompts assigned to that cell. The resulting object is a protocol-dependent conditional vulnerability map. Across three full-budget and five reduced-budget model evaluations, we observe up to 72.32% behavioral coverage and up to 370 vulnerabilities. The primary models exhibit different archive-level signatures: broadly distributed observed failures for Llama-3-8B, fragmented high-deviation regions for GPT-OSS-20B, and no observed cell above the selected threshold for GPT-5-Mini. We also measure how the discovered high-deviation cells change under three input-side defenses on Llama-3-8B. The maps, framework, metrics, and anonymized supplementary artifacts are released for evaluation use.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
benchmark, evaluation, maps-elite, safety, red-teaming, illumination

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_6wnROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a MAP-Elites-based framework for mapping where red-teaming failures occur in a behavioral descriptor space, with each occupied cell retaining the highest observed Alignment Deviation. Across three full-budget and five reduced-budget model evaluations, it reports up to 72.32% coverage and 370 vulnerable cells, with different failure patterns for Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini.

Critical strengths
  1. The paper addresses a timely and practically relevant problem in LLM safety evaluation.

  2. The proposed framework offers an intuitive way to organize and visualize failure patterns beyond aggregate scores.

  3. The evaluation across multiple models and defense settings provides a useful starting point for further study.

Critical weaknesses
  1. The paper is short and does not reach nine pages, which makes it difficult for me to understand the work fully.

  2. There are obvious formatting errors, such as “Appendix ??” on line 93, and the authors should check the manuscript carefully.

  3. The paper defines a vulnerable cell as having AD strictly greater than 0.5, and Appendix A.11.2 states that Figure 13 reports the mean AD of these cells. However, the figure reports 0.49 for “information”, 0.47 for “process”, and 0.38 for “create”, which is impossible if all the cells included in each mean have AD greater than 0.5.

Questions for authors

see weaknesses

Reviewer_8uwLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a framework for mapping LLM vulnerabilities into a two-dimensional behavioral descriptor space using MAP-Elites. Rather than reporting only aggregate attack success rates, the authors aim to characterize how observed failures are distributed across different prompt characteristics, namely Query Indirection and Authority Framing. The framework organizes generated adversarial prompts into a 25×25 grid, retaining the highest-scoring prompt-response pair in each cell according to an automated Alignment Deviation metric. The authors evaluate the approach across eight LLMs, compare it with existing red-teaming methods, including PAIR, TAP, and GCG, and investigate the persistence of observed vulnerabilities under three input-side defenses. The main contribution is an evaluation and visualization framework for characterizing the distribution, coverage, and connectedness of observed vulnerabilities rather than a fundamentally new attack-generation technique.

Critical strengths
  1. An interesting perspective on vulnerability evaluation and visualization The paper proposes a useful way to organize red-teaming results beyond conventional aggregate metrics such as attack success rate. Representing observed failures in a behavioral descriptor space can provide an additional perspective on their distribution and potentially help identify areas requiring further investigation.
  2. Evaluation across multiple models and defensive configurations The authors evaluate the framework across several LLM families and compare it with established red-teaming methods. The additional experiments examining how vulnerability maps change under different input-side defenses provide potentially useful information for security assessment.
Critical weaknesses
  1. Limited methodological novelty My primary concern is the limited novelty of the proposed approach. The framework is largely based on existing Quality-Diversity optimization techniques, particularly MAP-Elites, which have already been applied to adversarial prompt generation in prior work such as Rainbow Teaming (Samvelyan et al., 2024). The main contribution appears to be the organization, visualization, and analysis of existing red-teaming results within a predefined behavioral descriptor space. However, a similar vulnerability map could be constructed through a relatively straightforward preprocessing or postprocessing procedure, in which generated adversarial prompts are classified according to selected behavioral descriptors, assigned to corresponding grid cells, and associated with their highest observed vulnerability scores. In fact, the authors already apply a similar projection procedure to the outputs of existing red-teaming methods in their baseline comparison. While the additional analysis of connected vulnerable regions may have practical value, it is unclear what fundamentally new methodological or scientific contribution is introduced compared to existing quality-diversity-based red-teaming approaches. Moreover, although Rainbow Teaming is acknowledged as the closest related work, a direct experimental comparison is missing. Consequently, I find the methodological contribution relatively incremental and insufficiently differentiated from prior work.
  2. Limited experimental significance and insufficient evaluation validation Although the experimental results are reasonably good, their scientific significance is not sufficiently demonstrated. In particular, the comparison with PAIR does not establish a consistent advantage of the proposed approach. On Llama-3-8B, MAP-Elites achieves only a marginal improvement in behavioral coverage (63.04% compared to 61.44% for PAIR). On GPT-OSS-20B, its coverage is substantially lower (36.32% compared to 63.20%). On GPT-5-Mini, MAP-Elites achieves better coverage, but neither method identifies any vulnerable cells. While the proposed approach discovers more vulnerable cells on some models, the Diversity metric simply counts the number of occupied cells whose Alignment Deviation exceeds the selected threshold. It does not necessarily represent distinct vulnerabilities or independent failure mechanisms. Therefore, it remains unclear whether the reported improvements in Diversity correspond to meaningful security findings or simply reflect a broader distribution of similar attacks across the predefined descriptor space. More importantly, the experimental evaluation does not clearly demonstrate what additional scientific insights or practical advantages the proposed approach provides over existing red-teaming methods. There are also several concerns regarding the reliability of the evaluation: The behavioral descriptors are assigned using an LLM, but their consistency and semantic validity are not independently established. Consequently, the identified connected vulnerability regions may depend substantially on the choice of descriptors, grid resolution, and classification procedure. Alignment Deviation relies on automated judges without independent human validation. Appendix A.10 reports notable disagreement between automated scoring configurations, raising concerns about the reliability of the vulnerability classifications. The primary full-budget experiments do not provide sufficient evidence of reproducibility across independent runs. Furthermore, retaining the maximum score from stochastic responses may overestimate the persistence of discovered vulnerabilities. Table 2 contains an apparent numerical inconsistency: GCG achieves 7.20% coverage on Llama-3-8B, corresponding to approximately 45 occupied cells in a 625-cell grid, while its reported Diversity is 173 vulnerable cells. Under the stated metric definitions, Diversity cannot exceed the number of occupied cells. Taken together, these limitations make it difficult to assess whether the reported vulnerability maps represent stable and meaningful properties of the evaluated models. Given the limited methodological novelty, I would expect stronger experimental evidence demonstrating the significance and practical value of the proposed framework.
Questions for authors
  1. Novelty relative to existing approaches Could the authors clarify the key methodological differences between the proposed framework and Rainbow Teaming, beyond the additional visualization and analysis of the MAP-Elites archive? In particular, what prevents the proposed vulnerability maps from being constructed through a straightforward classification and postprocessing procedure applied to adversarial prompts generated by existing red-teaming methods? A direct comparison with Rainbow Teaming or a simpler postprocessing-based baseline would help establish the novelty and necessity of the proposed framework.
  2. Scientific significance of the experimental results Given that the proposed approach does not consistently outperform PAIR in behavioral coverage, what is the main empirical evidence supporting its practical advantage over existing red-teaming approaches? Furthermore, can the authors demonstrate that the additional vulnerable cells identified by MAP-Elites correspond to genuinely distinct vulnerability mechanisms rather than variations of similar attacks assigned to different cells? This distinction is particularly important because the reported Diversity metric does not directly measure the diversity of underlying vulnerabilities. Insufficient experimental repetitions. The primary full-budget experiments appear to rely on a single run per model. Given the stochastic nature of both MAP-Elites and LLM response generation, I believe that multiple independent runs are necessary to establish the reliability and statistical significance of the reported results. The authors should repeat the main experiments with different random seeds and report the mean and variance of the key metrics, including coverage, diversity, and vulnerability-region statistics. This is particularly important when comparing against PAIR, where some reported performance differences are relatively small and may fall within the natural variability between runs.
  3. Reliability of the behavioral descriptors and vulnerability scoring How sensitive are the resulting vulnerability maps and connected regions to the choice of behavioral descriptors, descriptor-scoring models, grid resolution, and Alignment Deviation threshold? Additionally, can the authors provide independent human validation of the vulnerability scores, particularly given the disagreement between automated scoring configurations reported in Appendix A.10?
  4. Consistency and reproducibility of the experimental results Could the authors clarify the apparent inconsistency in Table 2, where GCG reports 7.20% coverage but 173 vulnerable cells on Llama-3-8B? Also, how stable are the main vulnerability maps across independent experimental runs? Are the reported differences between methods statistically meaningful, particularly when improvements in coverage are relatively small?
Reviewer_24eaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Maps of Failures, a framework for organizing failures found during LLM red teaming. It uses MAP-Elites to search for adversarial prompts. Each prompt is assigned to a two-dimensional grid based on query indirection and authority framing. Each cell retains the prompt–response pair with the highest observed harmfulness score from two LLM judges. The paper compares coverage and connected vulnerable regions across three primary models and five additional models. It also evaluates three input-side defenses on Llama-3-8B. The results show different failure patterns across models and partial persistence of vulnerable cells after defense.

Critical strengths
  1. The maps provide a clear way to inspect the failures found during red teaming. They show differences in coverage and connected regions across models. These differences are not captured by peak harmfulness scores alone.
  2. The evaluation covers several target models and three input-side defenses. The paper also includes ablations and a small response-resampling study. These experiments help assess the behavior and limitations of the framework.
Critical weaknesses
  1. Understanding the distribution and diversity of LLM safety failures is not a new research problem. Prior work has already studied jailbreak effectiveness across much richer taxonomies of attack strategies. The paper offers limited new insights beyond these existing studies.
  2. The methodological contribution is limited. The framework uses existing MAP-Elites search and established prompt mutation strategies. The paper does not clearly demonstrate the practical value of these summaries beyond existing red-teaming metrics. It also does not show how the maps can guide better attacks, data generation, or defenses. The contribution therefore appears largely descriptive.
  3. Poor presentation quality. The figures are visually unappealing and convey limited information beyond the reported numerical results. In particular, the 3D visualizations add little analytical value. The overall presentation and writing also lack polish, with unresolved references such as “Appendix ??” in Section 2.
Questions for authors
  1. Why were Query Indirection and Authority Framing selected as the two behavioral dimensions?
  2. How reliable are the LLM-generated behavioral descriptors? Have the authors evaluated their agreement with human annotations or alternative descriptor models?
  3. Can the proposed maps provide actionable benefits beyond existing red-teaming metrics? For example, can they guide more effective attacks, targeted data generation, or defense improvements?
  4. How robust are the vulnerability maps to different grid resolutions, AD thresholds, and independent search runs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.