acceptodds
Under review as a conference paper at ICLR 2027

Direction-Indexed Evolutionary Red-Teaming of Large Language Models

Abstract

We introduce Direction-Indexed Evolutionary Red-Teaming (DIER), a black-box jailbreak method that learns its search coordinate from a target language model's own responses rather than from the text of the attack. A search that accumulates reusable framings, such as roles, pretexts and personas, has to decide which are close enough to recombine, and text does not carry that decision: across 21 target models, how a framing is worded says little about what it does, and a whole response is dominated by what was asked rather than by how it was asked. After each attempt DIER takes the displacement of the response from the request in a fixed sentence-embedding space and projects out the direction every response to that request shares; what survives is a signature of the tactic alone. That signature is cheap to measure and not predictable from a framing's text, so it can be read only after an attempt and used only to rank the library the search has already built. That ranking is the coordinate's only causal path into an attack: the direction never enters a prompt, is never decoded into words, and never touches the target's activations. On the 400-behavior HarmBench text benchmark at 32 target queries per behavior, DIER attacks at the state of the art: a mean attack success rate of 86.4 across the 16 targets with published baselines against 55.2 for the strongest attack in the HarmBench suite, level on the mean with the strongest prior strategy-library agent and ahead of it on all five targets where that agent stays below 90 percent. Replacing the ranking with a uniform random draw from the same library, with the attacker model, the library and the budget untouched, costs 13.5 to 39.0 points, placing the effect in the measured coordinate rather than in the language model that writes the attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
red teaming, jailbreak, AI safety, evolutionary search

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JD5MOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes DIER, a black-box jailbreak method that uses task-residualized response embeddings to guide attack strategy retrieval. Specifically, the method computes the embedding difference between the target model's response and the original task, estimates a task-specific common direction from previously observed difference vectors for the same task, and removes this direction through orthogonal projection. The resulting residual vector is used as a retrieval representation for attack strategies within an LLM-based evolutionary search framework. The authors evaluate the method on 400 HarmBench behaviors across 21 target models and conduct representation analyses, ablation studies, and cross-model transfer experiments.

Critical strengths

This paper proposes a representation for attack strategy retrieval based on the target model's responses. By removing task-relevant directions to guide evolutionary search, it offers a novel approach for optimizing the retrieval and selection of attack strategies.

Critical weaknesses
  1. Limited innovation: DIER essentially follows the jailbreak search paradigm established by AutoDAN-Turbo, which relies on a strategy library and LLM-based generation. The primary distinction lies in the use of "task-residualized response representation" as the new embedding for strategy retrieval. Consequently, the substantive methodological innovation of this paper relative to existing work is rather limited.
  2. Insufficient experimental validation: First, the baseline results are largely cited from prior literature rather than being reproduced under a unified experimental setup; thus, the fairness of comparisons across different methods warrants further clarification. Second, the paper’s coverage of more robust closed-source models is limited, and it lacks direct comparative experiments against representative existing jailbreak methods on the closed-source models that were tested. Furthermore, given that the core innovation is the "task-residualized response representation," the absence of an ablation study directly comparing performance with and without this "task-residualization" makes it difficult to fully substantiate the independent contribution of this design. Finally, since the authors posit that this representation improves strategy retrieval, could it be extended to other similar automated jailbreak optimization algorithms to further validate the method's generalizability, practical effectiveness, and impact on average query consumption?
Questions for authors
  1. The clarity of the paper’s writing could be improved. Certain parts are currently difficult to follow; I suggest refining the text to make the description of the proposed method clearer and more accessible. For instance, even after reading the abstract, it remains challenging to intuitively grasp the specific method being proposed and how it fundamentally differs from existing research.

  2. Regarding the use of "Judges" in the experiments: The experimental setup involves two distinct types of Judges. Please clarify the specific roles each of these Judges plays within the experiments.

  3. Regarding performance gains and search efficiency: In the main experiments, DIER achieves an average ASR of 86.4%, compared to 86.1% for AutoDAN-Turbo—an improvement of only 0.3 percentage points. However, the paper highlights a performance gain of approximately 31 percentage points; I suggest explicitly stating the baseline against which this figure is calculated. Given that the overall performance improvement of DIER over AutoDAN-Turbo is quite marginal, does the proposed method offer any advantages in terms of search efficiency or computational cost? While the paper reports the query index at which the first successful attack occurs, it appears to lack a direct comparison of efficiency against existing methods like AutoDAN-Turbo.

  4. Regarding the experimental categories in Table 3: Table 3 presents results across four categories—"All," "Standard," "Contextual," and "Copyright"—yet the paper does not clearly define what these categories entail. Please provide a detailed explanation of the definitions for each category.

Reviewer_1NnNOpenReview ↗
Rating
2(+1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

DIER uses behavior-residualized response directions to index a growing framing library. For each deployment, it embeds the target's reply relative to the behavior, projects out that behavior's running mean displacement, and retrieves two framings for an operator model to recombine. The direction stays outside the prompt and the target's activations. Evaluation covers 21 targets at 32 target queries per behavior, representation analyses, two search ablations on three targets, and frozen-library transfer. The main contribution is the retrieval coordinate; the evolutionary loop follows existing methods.

Critical strengths

(1) Residualization improves tactic identification on 19 of 21 targets; the two low-compliance models are the exceptions. Tables 9 and 10 also report near-null realization fidelity and weak parent-to-child coordinate transmission. The authors acknowledge these limits and describe the direction as a retrieval index, without claiming they can steer a new response toward it.

(2) The ablation gains hold on standard behaviors, where the contextual and copyright concerns in W1 do not apply. Gemma 4 scores 59.0% with the full method, 51.0% with random retrieval, and 19.0% without query variation; both deletions also hurt GPT-5-nano and GPT-OSS-20B (Table 8). This is useful evidence that informed retrieval and variation help in these runs.

Critical weaknesses

(1) The main ASR table mixes evaluation protocols. Section 4.1 scores all 400 behaviors from bare behavior text with the standard classifier prompt. HarmBench [1] supplies the context when judging contextual behaviors and uses a hashing-based classifier for copyright behaviors. The official code implements both branches. These categories make up 200 of the 400 items, so using one evaluator for every DIER arm does not make the results comparable to the imported HarmBench columns. A positive Llama-2 classification also does not establish verbatim reproduction of copyrighted text. The claimed 31-point gain needs to be re-scored before I can judge how much survives. The AutoDAN-Turbo comparison uses historical results from a different attacker and library-building schedule [2]; it also needs a matched experiment.

(2) A1 shows that the ranking beats random retrieval, but the paper never directly tests what residualization adds. A1 removes informed ranking, and A2 removes query variation. Neither substitutes the raw response embedding, raw displacement, or a semantic embedding of the framing in the same loop. P1 measures the correlation between n-gram text similarity and coordinate similarity; it does not test whether a predictor can infer the coordinate from text. P4 measures tactic identity, which does not establish an advantage in attack success. On GPT-5-nano and GPT-OSS-20B, residualization gives no improvement in tactic identification, yet removing the ranking causes the largest A1 losses. The ranking clearly helps there; its dependence on projection remains untested. Appendix H also needs an ASR comparison with utility ranking: low overlap only tells us the two rankings choose different framings.

References
[1] HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal, ICML 2024.
[2] AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs, ICLR 2025.

Questions for authors

Q1. Does the code follow Section 4.1 literally? If so, please re-score the already selected responses with HarmBench's contextual and copyright evaluators [1], report each of the three categories, and update the overall comparison. If the code already uses the official branches, please give the evaluator configuration and correct the description. What context did the operator receive for contextual attacks, and how were responses truncated for final scoring?

Q2. Can you compare residual retrieval with raw-response, raw-displacement, and semantic-framing retrieval using the same operator, initialization, budget, and search loop? Results on a few shared targets, repeated across seeds and behavior orders, would help establish what the residual contributes. Please change only the retrieval representation in these comparisons.

Q3. Appendix J says a behavior may appear in more than one run and counts queries within the run that first solved it. Do Tables 2, 3, and 8 use one fixed run per target and arm, or combine successes across runs? Please give the number of runs, aggregation rule, behavior order, and actual target calls over all 400 behaviors, including failures. Each behavior consumes 16 or 32 calls in two waves, so Table 13's first-success position on a selected subset does not give the total cost. How were B and N selected?

References
[1] HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal, ICML 2024.

Reviewer_AR1MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Direction-Indexed Evolutionary Red-Teaming (DIER), a black-box jailbreak framework that organizes and retrieves adversarial strategies using response-derived directions rather than textual similarity or predefined strategy categories. The key idea is to construct a tactic representation by subtracting the target model's response embedding from the corresponding behavior embedding and projecting out a behavior-specific mean direction. The resulting residual representation is used as a retrieval coordinate for selecting and recombining attack framings within an evolutionary search framework. The authors evaluate DIER on 400 HarmBench behaviors across 21 target models. On the 16 models with published baselines, DIER achieves an average attack success rate (ASR) of 86.4%, compared with 86.1% for AutoDAN-Turbo. Ablations on three models show that replacing direction-based retrieval with random selection decreases ASR by 13.5–39.0 percentage points. The paper presents an interesting perspective on black-box red teaming: learning a strategy index from a model's observed behavior rather than relying solely on the attack's linguistic surface form. However, I have several concerns regarding whether the experiments sufficiently establish the necessity of the proposed residualized representation, the interpretation of directional evolution, and the robustness of the reported improvements.

Critical strengths

S1. The paper identifies a meaningful problem in strategy-library-based red teaming: how to organize and retrieve reusable attack strategies according to their behavioral effects rather than their textual descriptions. S2. Broad experimental evaluation and substantial performance improvements over random retrieval. The evaluation covers 21 target models, including different model families and safety capabilities. The authors report both historical baseline comparisons and additional results on five more recent models. The ablation results are particularly encouraging. On Gemma 4 26B-A4B-it, GPT-5-nano, and GPT-OSS-20B, direction-based retrieval substantially outperforms uniform random selection. These results suggest that informed retrieval contributes meaningfully to the overall attack framework.

Critical weaknesses

W1. The central representation-level contribution is not adequately isolated by the ablation experiments. This is my primary concern. The paper's central claim is not merely that retrieval improves evolutionary red teaming, but that behavior-residualized response directions provide a better retrieval coordinate than existing alternatives. However, the main A1 ablation (Section 4.4, Table 5) compares the proposed cosine-based retrieval against uniform random retrieval. This establishes that the retrieval mechanism is useful, but does not demonstrate that residualization itself is necessary or superior to simpler informed retrieval strategies. Retrieval using the raw response embedding or raw response displacement, without behavior-specific projection. Retrieval using behavioral feedback such as refusal/compliance labels or success-score information. The representation pre-checks in Section 4.3 do not fully resolve this issue. For example, a low correlation between n-gram textual similarity and residual-direction similarity does not establish that stronger semantic text representations cannot predict function. Likewise, improved tactic-identity classification does not directly establish improved retrieval quality for jailbreak generation. Furthermore, the two online ablation arms generate different offspring and consequently develop different libraries over time. Thus, A1 measures the total effect of changing the selection policy, including its downstream consequences, rather than retrieval performance from an identical frozen library. I consider direct, controlled comparisons of alternative retrieval coordinates essential for substantiating the main novelty claim. W2. The interpretation of directional evolution is insufficiently supported by the observed transmission and fidelity results. Equation (3) introduces directional fusion and mutation, and A2 is intended to establish the importance of varying the query direction. However, Appendix F reports nearly zero realization fidelity between the intended query direction and the direction actually produced by an offspring. Table 9 gives mean fidelity values close to zero, sometimes negative, and without consistent improvement over A1. More importantly, Table 10 reports parent-to-child coordinate regression R² values of at most 0.04 across the 21 target models. These findings suggest that arithmetic operations in the proposed directional space have very limited predictive control over the behavioral directions realized by offspring. The authors acknowledge that the method is based on retrieval rather than direct steering, which is a reasonable interpretation. Nevertheless, the strong performance reduction caused by A2 does not establish that the particular vector arithmetic is necessary. A2 changes the query construction process, but does not compare the proposed directional variation with simpler alternatives such as randomly sampled query directions or alternative parent-selection strategies. Consequently, the results convincingly support the usefulness of a retrieval-driven search process, but provide weaker evidence that the proposed directional evolutionary operations contribute their claimed mechanism. W3. The comparison with the strongest previous method and the statistical robustness of the results require further clarification. The main results are competitive, but the state-of-the-art claim should be interpreted carefully. In Table 2, DIER achieves 86.4% average ASR versus 86.1% for AutoDAN-Turbo. The difference is only 0.3 percentage points, and DIER outperforms AutoDAN-Turbo on 8 of the 16 models while underperforming on the other 8. Although DIER improves performance on all five models where AutoDAN-Turbo achieves below 90% ASR, this subgroup analysis does not by itself establish overall superiority. The improvement on Llama-2-7b is substantial, but several other per-model differences are much smaller. The paper also uses previously published baseline results rather than controlled reruns under a common operator model, judge, and query-accounting protocol. It is therefore difficult to determine whether these comparisons are directly equivalent in experimental cost and implementation conditions. In addition, the central A1/A2 ablations are reported on only three of the 21 target models. The manuscript does not clearly establish stability across multiple independent random seeds. Statistical tests over 400 behaviors quantify uncertainty in the observed outcomes, but cannot substitute for measuring variation across independent evolutionary runs. Finally, the attack framework carries an evolving library across benchmark behaviors. Its sensitivity to behavior ordering, random initialization, and the composition of previously encountered behaviors remains unclear. These issues make it difficult to assess the robustness and generality of the claimed performance advantages. W4. The residual direction may encode coarse response states rather than tactic-specific functional mechanisms. Section 4.3 argues that the residual direction captures tactic-specific behavior after removing request-specific information. However, Figure 4(a) shows that the response space is strongly organized around a refusal-versus-compliance axis, with the harm direction nearly aligned with that axis. This observation raises an alternative explanation: much of the useful signal in the residual representation may come from distinguishing broad response states, such as refusal, partial compliance, and successful completion, rather than from identifying distinct reusable attack mechanisms. The tactic-identity experiment in Figure 3(a) provides some evidence for tactic-related information, but does not exclude this explanation, particularly because tactics may differ systematically in their compliance rates. A stronger analysis would compare the proposed representation against compliance-matched controls, or measure tactic identity and retrieval effectiveness within subsets sharing the same response-status category. Relatedly, the reported correlation of −0.996 between residualized behavior variance and log compliance is striking, but the proposed 10% compliance threshold is not sufficiently established as a general scope condition. Appendix K notes that there are no observed targets with compliance rates between 10% and 40%. The threshold is therefore supported by a substantial gap in the sampled model distribution rather than a direct investigation of the transition region. I would encourage the authors to characterize this observation as an empirical relationship under the evaluated models, rather than as a general law, unless additional evidence is provided.

Questions for authors

Q1. Can the authors provide a controlled comparison of alternative retrieval representations? In particular, I would like to see the proposed residual direction compared against raw response embeddings, raw response displacements, semantic framing embeddings, and simple compliance- or utility-informed retrieval under the same evolutionary framework. This is the most important question for my recommendation. If residualization consistently improves ASR over these alternatives, it would substantially strengthen the central contribution. Q2. What specifically makes directional variation useful despite the near-zero parent-to-child fidelity? Given the results in Tables 9 and 10, could the authors compare Equation (3) against random query directions, direct parent retrieval, or alternative methods of constructing retrieval queries? It would be particularly helpful to distinguish benefits from directional arithmetic itself from benefits arising simply from selecting different subsets of the framing library. Q3. How robust are the improvements across independent runs and benchmark orderings? Could the authors report multi-seed results for the three ablation targets, and ideally extend the most important representation ablations to additional model families? Since the archive is carried across behaviors, how much do ASR and library diversity change when the benchmark behavior order is randomized? Q4. To what extent does the residual representation capture tactic identity beyond refusal/compliance? Can the authors evaluate tactic classification or retrieval quality within compliance-matched groups, or after controlling explicitly for response-status labels? This would help establish whether the representation captures distinct mechanisms instead of primarily exploiting a low-dimensional compliance signal. Q5. Can the authors clarify the experimental budget equivalence with AutoDAN-Turbo? How do target queries, auxiliary model calls, and the total amount of strategy exploration compare between DIER and the strongest baseline? Also, since Appendix J reports the position of the first successful query rather than the total number of executed queries, reporting actual average consumed target queries would make efficiency comparisons more informative.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.