acceptodds
Under review as a conference paper at ICLR 2027

Wide Exploration, Localized Escape: High-Dimensional Branching Dynamics of Multi-Turn Jailbreaks

Abstract

Safety alignment mechanisms for large language models (LLMs) are commonly optimized against isolated or single-turn malicious requests, while adversaries can distribute harmful intent across multiple adaptive interactions. We study multi-turn jailbreaks from a spatial-dynamical perspective and formulate their evolution as a high-dimensional branching random walk with adversarial drift. In this framework, adaptive prompting steers semantic trajectories toward a proxy safety boundary, generation uncertainty introduces stochastic perturbations, and parallel conversational paths form an expanding attack tree. The model predicts that branching amplifies the finite-sample semantic frontier and accelerates first boundary crossing, while successful crossings concentrate around a limited set of localized exits. We evaluate these predictions using two multi-turn jailbreak methods across multiple mainstream LLMs and semantic encoders. Increasing the number of parallel paths consistently expands the reachable semantic frontier, advances the first crossing, and raises the probability of observing at least one crossing within a fixed interaction budget. These gains arise primarily from repeated path exploration rather than greater semantic diversity among branches. Moreover, successful crossing points remain closer to localized exit regions than matched non-crossing controls, and their terminal transitions exhibit stronger directional alignment with the corresponding exits. Although exact exit locations depend on the semantic representation, these localization and alignment patterns remain stable across encoders. Finally, ablations show that combining exit directions, local semantics, successful examples, and regional context yields more reliable attack-reconstruction gains than any isolated signal. Our findings reveal a characteristic structure of multi-turn jailbreaks—wide exploration inside the safety region followed by localized boundary escape—and provide a mechanistic account of their structural risk.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-Turn Jailbreaks, LLM Safety, High-Dimensional Branching Dynamics

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

This paper analyzes multi-turn jailbreak trajectories using BGE-M3 embeddings and an SVM safety proxy. Across two datasets and five models, it examines candidate-path coverage, localization of safety-crossing points, and the utility of exit-related information for attack reconstruction.

Critical strengths
  1. Relevant problem and accessible analysis framework. Studying how jailbreak behavior changes across dialogue turns is worthwhile. Using observable responses makes the framework applicable to close-sourced models without access to internal activations.

  2. Complementary behavioral evaluation. The independent model-based audit provides evidence beyond the embedding-based safety proxy. Crossing-region context shows the most consistent benefit among the evaluated information sources.

  3. Explicit acknowledgment of limitations. The paper distinguishes the proxy boundary from the model’s internal safety boundary and acknowledges the retrospective nature of reconstruction and the limited interpretation of candidate-set monotonicity.

Critical weaknesses
  1. Limited validation across embedding models. The main geometric findings rely on BGE-M3, which is used for safety classification, crossing-event selection, and geometric analysis. Varying truncation and aggregation within the same encoder does not establish cross-representation robustness. Replicating RQ2 with an alternative encoder, such as Qwen3-Embedding, and independently assigned safety labels would strengthen the conclusions.

  2. The detector does not explicitly incorporate dialogue context. The safety proxy takes individual assistant responses as input, whereas the paper studies multi-turn safety dynamics. A response’s safety can depend on the preceding requests and conversation history. The authors should compare response-only and context-aware safety judgments and examine whether the detected crossing events remain consistent.

  3. The breadth–depth comparison may disadvantage depth. Attack_600 contains pregenerated paths whose later prompts may depend on successful progression in earlier turns. Continuing an unchanged path after an early failure may consume queries without effectively advancing the attack. The authors should clarify failure handling, report later success conditional on earlier failure, and include a failure-aware restart baseline under matched query budgets.

  4. RQ3 evaluates single-turn reconstruction rather than multi-turn attack improvements. Although the information is extracted from multi-turn trajectories, the reconstructed attacks are single-turn queries.

Questions for authors
  1. Do the localization and alignment findings persist with an alternative encoder, such as Qwen3-Embedding, particularly when crossing events are determined using independent safety labels?

  2. How consistent are response-only safety judgments with judgments that incorporate dialogue history? Does context-aware evaluation change the detected crossing turns or the RQ2 conclusions?

  3. In the breadth–depth comparison, how are early refusals or failed steps handled? What proportion of paths succeed after an earlier failure, and does the breadth advantage remain against a failure-aware restart baseline?

5 comments

Response to Q1: With independently labeled behavioral crossings, the first unsafe response showed a positive localization gap in both BGE-M3 and Doubao Embedding. The result depended on how we defined the crossing location. When we used the segment midpoint, the interval included zero in both spaces. The original small positive alignment effect also did not replicate across all 18 events.

Encoding sensitivity in the original manuscript

Appendix D.2, Table 9 examines five ways of encoding responses with BGE-M3. For Doubao 1.6 responses on Attack_600, the main configuration yielded 242 SVM-defined crossing points and eight non-noise clusters. Encoding only the first 512 tokens yielded 339 crossing points and six clusters. Averaging embeddings across 512-token chunks yielded 347 crossing points and two clusters. Table 9 thus shows how response truncation and aggregation affect crossing detection and clustering within BGE-M3. The supplementary analysis below addresses the cross-encoder question directly.

Localization and alignment across all 18 behavioral crossings

We used Doubao Embedding (doubao-embedding-vision-251215) as the alternative encoder. An independent judge with access to the dialogue history identified 18 first behavioral crossings, 16 on Crescendo and two on Attack_600. We evaluated the same events in both embedding spaces, matched them to Safe-to-Safe controls, and constructed a single reference center in each space using other intents. A positive localization gap means that the observed location was closer to the reference center than its matched control.

Measure across all 18 eventsBGE-M3, mean [95% CI]Doubao Embedding, mean [95% CI]
Localization gap at the first unsafe response+0.0494 [0.0084, 0.0894]+0.0626 [0.0153, 0.1091]
Localization gap at the segment midpoint+0.0122 [−0.0262, 0.0507]+0.0424 [−0.0002, 0.0877]
Alignment difference relative to matched controls−0.0099 [−0.0686, 0.0486]−0.0119 [−0.0636, 0.0389]

The localization result in this pilot applies to the first unsafe response in a judge-identified transition, measured relative to a single reference center. The original analysis located each crossing by interpolation at the SVM boundary and measured its distance to the nearest cluster centroid.

Event-level agreement and the selected nine-event subset

Across all 18 events, the alignment differences were positively correlated between encoders (Pearson’s r = 0.73), and the two encoders agreed on the sign for 13 of 18 events. This correlation describes how alignment differences varied from event to event across the two spaces. The overall mean difference was close to zero in each space.

In a descriptive subset analysis, we selected the nine events with a positive alignment difference under BGE-M3. Holding those events and their matched controls fixed, the mean difference was +0.0972 under BGE-M3 and +0.0560 under Doubao Embedding.

Response to Q2: Including earlier dialogue turns reduced the number of dialogues with a detected behavioral crossing from 31 to 18, although the two evaluations agreed on 463 of 500 turn-level safety judgments (92.6%). Both evaluations used Doubao 2.0 lite and the same rubric. In both, the judge saw the current user request and the current assistant response. The only difference in the inputs was whether the judge could also see earlier dialogue turns.

Turn-level judgments and dialogue-level crossings

Among the 400 responses with earlier dialogue turns, the two evaluations agreed on 365 of 400 judgments (91.25%). With history included, 33 judgments changed from unsafe to safe and two changed from safe to unsafe. On the 100 first turns, where the inputs were identical, two judgments differed. Repeated scoring therefore contributed some variability to the comparison.

Sixteen dialogues had a crossing under both protocols. With history included, the first crossing occurred on the same turn in 11 of these dialogues and later in five. Fifteen dialogues had a crossing only under current-turn evaluation, and two had a crossing only under history-aware evaluation.

Localization and alignment across all first crossings

We evaluated all first crossings identified under each protocol: 31 under current-turn evaluation and 18 under history-aware evaluation. The table reports the mean localization gap at the first unsafe response. A positive gap means that the response was closer to the reference center than its matched safe control.

EncoderCurrent-turn evaluation (31 events)History-aware evaluation (18 events)
BGE-M3+0.0078+0.0494
Doubao Embedding+0.0270+0.0626

The mean localization gap was positive for both encoders under both protocols. For each encoder, the 95% interval excluded zero only under history-aware evaluation. Mean alignment differences were close to zero under both protocols, and all four alignment intervals included zero.

Exploratory comparison of the same crossing steps

Of the 31 first crossings under current-turn evaluation, 15 had a positive BGE-M3 alignment difference. When history was included, eight of these 15 retained the same first-crossing turn, two shifted to a later turn, and five no longer had a detected crossing. We compared the eight events with the same first-crossing turn under both protocols. Localization was not a selection criterion.

Measure for the eight selected eventsEncoderCurrent-turn evaluationHistory-aware evaluation
Mean alignment differenceBGE-M3+0.0911+0.0688
Mean alignment differenceDoubao Embedding+0.0138+0.0366
Mean post-crossing localization gapBGE-M3+0.0378+0.0777
Mean post-crossing localization gapDoubao Embedding+0.0583+0.0937

The positive alignment means apply to a subset selected for positive BGE-M3 alignment under current-turn evaluation. For localization, the same eight post-crossing responses were, on average, closer to the reference center than their matched safe controls under both protocols and in both embedding spaces. Because we re-estimated the controls and reference centers under each label set, this comparison supports consistency in the direction of the localization gap for these eight events.

Response to Q3: Thank you for pointing out that the handling of early refusals may affect the breadth–depth comparison. The original experiment replayed pregenerated dialogue paths. In the supplementary analysis, we retained the Table 8 hit criterion: a query counts as a hit if the SVM classifies the response as unsafe. We examined how often continuing after a first-turn miss yielded a later hit and compared breadth-first allocation with a failure-aware restart policy under a six-query budget.

1. How did the original experiment handle early refusals or failures?

In Table 8, we replayed the paths in a predetermined order. If a turn produced a valid response that triggered the refusal rule or failed to meet the SVM hit criterion, the depth-first policy continued to the next recorded turn on the same path. An early failure did not trigger an automatic restart. The original implementation skipped empty responses. In the supplementary budget comparison, we preserved the original turn positions and counted each empty or invalid response as one query.

We distinguished a miss from a refusal. A miss means that the turn yielded no response classified as unsafe by the SVM. We identified refusals separately using a text-based rule.

2. What proportion of paths yielded a later hit after an early failure?

We defined early failure as a first-turn miss and examined subsequent recorded responses on the same path. Each percentage below is the proportion of paths in that row with at least one SVM hit during the indicated window.

Paths with a first-turn missNumber of pathsHit in turns 2–3Hit in turns 2–5
Attack_6002,072757 (36.5%)1,093 (52.8%)
Attack_600 subset whose first turn also triggered the refusal rule279 (33.3%)12 (44.4%)
Both datasets combined3,44239.6%55.1%

The turns 2–3 window corresponds directly to the original depth allocation of up to three turns per path. Across both datasets, 39.6% of paths with a first-turn miss yielded a hit within that allocation.

3. Does breadth-first allocation outperform failure-aware restart under a six-query budget?

We compared two policies on the same three recorded paths. The breadth-first policy allocated two queries to each path. The restart policy continued along the current path until it obtained a hit. If no hit had occurred, it switched to the next untried path when a response triggered the refusal rule, a response was invalid, or the current path had used three queries. Once all three paths had been started, the policy spent any remaining budget on the last path. We averaged the paired comparison over all six path orderings and gave equal weight to each of the ten model–dataset settings.

Table 8 reported hit rates of 76.26% for breadth and 75.25% for depth. In the supplementary replay, the hit rates were 76.35% for breadth-first allocation and 75.50% for failure-aware restart. The paired breadth-minus-restart difference was +0.85 percentage points, with the interval obtained by resampling source intents.

In a sensitivity analysis, we changed the restart rule so that any miss triggered a switch to an untried path. Under that rule, the breadth-minus-restart difference was +1.53 percentage points. In both comparisons, restart means switching to another recorded path for the same intent.

Scope of RQ3

The region information in Section 4.6 comes from multi-turn trajectories. The generator uses the supplied region information to produce a new single-turn query, and we evaluate the target model’s response to that query. This experiment measures retrospective reconstruction given an observed crossing region.

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

This paper analyzes the phenomena that occur when a model becomes unsafe, focusing on the point at which LLM responses become unsafe during multi-turn interactions. For this analysis, the authors encode LLM responses using a separate text embedding model, BGE-M3, and train an SVM to classify safe and unsafe responses in the BGE-M3 embedding space. In this space, they analyze (1) effective approaches to achieving attack success, i.e., safe-to-unsafe transitions, and (2) points at which the decision boundary is crossed, i.e., crossing points. Based on these analyses, the authors present the discovery of exit regions, where departures from safe responses are concentrated, and their potential use as the paper’s contributions.

Critical strengths
  1. A novel framing of multi-turn jailbreaks through branching processes and a proxy semantic space

The authors view multi-turn jailbreaks as the exploration of multiple branches toward unsafe regions and analyze trajectories that cross the decision boundary. I also find it interesting and novel to use the semantic space of an external text embedding model as a proxy space for observing these trajectories.

  1. Detailed and rigorous experimental analysis

The authors examine and report their experimental results through a variety of detailed analyses. I appreciate that the paper provides extensive statistical information, including confidence intervals, p-values, bootstrap analyses, matching, regression, and permutation tests, which helps quantify the strength and uncertainty of the reported observations.

Critical weaknesses
  1. The paper is difficult to follow even after careful reading because statements restricting the interpretation of the findings repeatedly interrupt the discussion. For example, after defining the BGE-M3 embedding space as a proxy space, the authors continue to emphasize that its boundary is “not the LLM’s internal safety boundary” (§3.1) and that “external response embeddings also do not reveal internal model states” (§3.2). Similarly, the discussion of multi-path exploration in RQ1 includes the qualification that the observed patterns “do not by themselves establish synergy, a branching mechanism, or gains beyond independent retries” (§3.2), with a similar distinction repeated in the experimental setup (§4.4).

This repetition disrupts the flow of reading and makes it difficult to identify what the authors have found and why those findings matter. As a result, I found the paper’s core message obscured by the many defensive statements.

  1. The RQ3 reconstruction procedure is underspecified, and this affects how I read Tables 6 and 13. Table 5 lists what each condition provides, but not how that information becomes input to the attack generator (Doubao 1.6). In C1, the centroid is a 1,024-dimensional BGE-M3 vector. How is it rendered as text? The introduction mentions “representative keywords,” but neither 4.6 nor the appendix describes how they are extracted from a cluster. For E1 and P1, how many examples or responses are supplied, how are they selected (nearest to the centroid, random, or by SVM margin), and are they restricted to training-fold intents? Most importantly, what exactly is the “context” in P1 and NP1 (full responses, pre/post crossing pairs, or excerpts), and on which covariates (topic, turn, length) is the NP1 context matched? P1 vs. NP1 is the most consistent contrast in the paper, and without the matching details I cannot tell whether it isolates crossing-region information or reflects differences in context length or specificity. Please also provide the prompt templates for all eight conditions, since Table 14 shows only the generated outputs.
Questions for authors
  1. Please explain what semantic characteristics are shared by the crossing points belonging to each exit region, or by the response pairs used to construct those crossing points, and what characteristics distinguish different exit regions.

  2. Please explain the specific procedure used in RQ3 to provide exit-region information to the attack-generation model.

4 comments

Readability and repeated qualifications

We agree that repeated qualifications obscure the main findings. Feedback on an earlier submission prompted us to narrow the scope of our claims. In this version, we repeated those limits in several places. In a revision, we would define the external proxy space once in Section 3.1, and explain the scope of offline replay in Section 4.4.

Response to Q1: Each crossing point is obtained by interpolating between the BGE-M3 embeddings of two consecutive responses, so we interpret its semantic characteristics by examining the corresponding response pair. In the original clustering for Attack_600 × Doubao 1.6, representative pairs within a region often concern the same entities, events, or technical objects, although the level of detail or perspective can shift between turns. For example, a pair in one region moves from Interpol’s efforts to combat human trafficking to the activities of criminal organizations. A pair in another region moves from the regulation of chemical substances to the hazards posed by a particular substance.

Some regions differ in ways that broad topic labels do not capture. For example, three regions all concern cybersecurity. Representative responses in one focus on groups such as Anonymous and how they operate. Those in another focus on databases, application inputs, and software vulnerabilities. Those in the third describe particular security researchers, devices, and historical cases. Other regions encompass a mix of public events or narrative themes and are harder to characterize with a single label. We offer these descriptions as qualitative interpretations of representative response pairs in this setting.

Exit regions in the Attack_600 × Doubao 1.6 clustering

Counts are crossing points assigned to each cluster. Descriptions summarize exemplar pairs in this original fit and do not imply a stable taxonomy across models or folds.

ClusterCrossing pointsCharacteristics of representative response pairs
028Narratives and audience perception, including fictional characters and media scandals.
114Chemical and biological agents, their regulation, and the properties of specific substances.
211Interpretation of public events and conflicts, including security incidents and geopolitical reporting.
326Historical organizations, social groups, and ideology.
429Enforcement and criminal organizations, including shifts from the enforcement perspective to the activities being investigated.
524Online activist groups such as Anonymous and their activities.
614Software inputs, databases, and vulnerabilities in applications.
715Named security researchers, devices, and historical cases.

Response to Q2: Thank you for pointing out that Table 5 does not fully explain how exit-region information reaches the generator. RQ3 uses retrospective reconstruction. For a target sample with an observed crossing, we take the cluster kk containing that crossing as the reference region. The generator, Doubao 1.6, receives the original objective, the actor and topic category, and the region-specific material supplied by the experimental condition. It uses this text prompt to generate one standalone English question. We then submit that question to the target model and evaluate its response. The reference region is supplied in advance, so the experiment measures reconstruction performance given the observed region.

How the centroid is represented in the prompt

We compute the centroid ckc_k by averaging the BGE-M3 vectors of the crossing points in cluster kk and L2-normalizing the result. In C1, the generator receives a textual description of the centroid’s neighborhood. We select up to five crossing events nearest to ckc_k by cosine similarity. For each, the prompt includes its similarity score, the user question at the crossing turn, and an excerpt from the adjacent assistant responses.

We also rank crossing-turn user questions by the cosine similarity of their own BGE-M3 embeddings to ckc_k and include up to six of them. When available, the prompt reports the similarity between the original objective and ckc_k, together with reference similarity values for crossing-turn questions. C1 therefore measures the contribution of centroid-guided textual references. The vector determines which material the generator sees, and the generator reads that material and its similarity scores.

How examples and keywords are selected

In the leave-one-query-out version of E1 (E1_LOO), we draw examples from crossing events in the same cluster after excluding events with the target query_id. We give priority to HDBSCAN exemplars and fill any remaining slots with other eligible events from the cluster. For each selected event, we concatenate the assistant responses immediately before and after the crossing, then include an excerpt of up to 700 characters. The prompt contains at most three such excerpts and up to 12 TF-IDF terms extracted from the selected transition texts.

The exclusion in this procedure operates at the target-query level. RQ2’s five-fold split applies separately to the geometric analysis.

What “context” means in P1 and NP1

P1 supplies excerpts from representative crossing-response pairs, keywords, available rhetorical patterns, and the centroid-guided textual references described above.

For NP1, we select adjacent response pairs from other queries for which the SVM labels both responses safe. We require the same topic category and prioritize pairs whose starting turn is closest to the starting turn of the target crossing. After deduplication, we include up to three excerpts of at most 700 characters each and extract keywords from the selected texts.

P1 and NP1 use the same excerpt-length cap, but the selection procedure does not match individual excerpts on actual length or semantic similarity. P1 also includes centroid-guided references that NP1 does not receive. The P1−NP1 contrast therefore estimates the difference between these two sets of inputs.

The direct Full−E1_LOO result uses the query-excluded exemplar condition. P1 likewise excludes the target query ID when constructing its regional excerpts and centroid anchors. These query-specific exclusions do not restrict all donors to an RQ2 training fold. C1 as implemented supplies textual anchors as well as cosine scores. NP1 matches category and nearby turn but does not explicitly match semantic distance. These details qualify the “direction alone” and “semantically similar context” descriptions in Table 5.

Baseline and combined conditions

  • G1 provides only the original objective, actor, and category.
  • W1 uses a distant cluster that does not match the target’s observed crossing region.
  • Full combines centroid-guided references, keywords, representative response excerpts, rhetorical patterns, and examples of crossing-turn user questions.

Across conditions, the multi-turn trajectories supply the reference material, while the generator’s output is a single-turn question.

Table 14 in the appendix provides the reconstructed system and user prompt templates for all eight conditions.

Thank you for the clarifications. I have remaining concerns about these:

  1. Interpretation of exit regions

It remains unclear whether crossing-point clustering reveals structure beyond endpoint similarity, particularly given the comparable localization gaps with cluster, global, and topic centroids. Could the authors compare the clustering of pre-crossing responses, post-crossing responses, and crossing points to clarify this?

  1. Experimental controls in RQ3

The response clarifies several details that materially affect the interpretation of Table 5:

  • C1 provides centroid-guided textual references, including crossing-turn questions, response excerpts, and cosine similarity scores. It therefore does not isolate exit-direction information.
  • P1 includes centroid-guided references and rhetorical patterns. NP1 does not receive the centroid-guided references.
  • NP1 matches topic category and nearby turn, but does not explicitly match semantic similarity.

These details should be reflected in Table 5 and Section 4.6. More importantly, the P1 vs. NP1 comparison does not isolate the contribution of crossing-region context: the auxiliary inputs differ, and the excerpts are not matched on actual length or semantic similarity.

Your response to Reviewer S9T8 also states that the 20 evaluation blocks per setting in Table 13 were selected based on observed differences between conditions. As acknowledged in that response, the reported p-values do not account for this selection.

Could the authors provide a comparison that controls the auxiliary inputs, matches actual excerpt length and semantic similarity, and uses evaluation blocks selected independently of the observed condition outcomes? Without such controls, the conclusions should be limited to exploratory differences between these composite input conditions.

Thank you for following up on the interpretation of the exit regions and the controls needed to assess the contribution of crossing-region excerpts in RQ3.

  1. Interpretation of exit regions

Crossing-point clusters agreed most closely with clusters formed from ordinary segment midpoints. They also overlapped with clusters formed from either endpoint. The comparison shows that the geometry of the response pairs captures part of the observed grouping.

We compared four representations of the same 2,561 first valid crossing events across ten settings: the pre-crossing response, the post-crossing response, the exact SVM crossing point, and the spherical midpoint. Within each intent-grouped training fold, we held the events and clustering parameters fixed and repeated the fits with three seeds. The mean adjusted Rand index (ARI) was 0.446 for crossing points versus midpoints, 0.334 versus post-crossing responses, and 0.295 versus pre-crossing responses. The pre- and post-crossing partitions had an ARI of 0.196. For comparison, agreement across seeds was 0.630 for crossing points and 0.651 for midpoints.

In the supplementary analysis of 1,998 first crossings with interpolation-matched Safe→Safe controls, the localization gaps were +0.0164 for the cluster reference, +0.0167 for the global centroid, and +0.0157 for the same-topic centroid. A single centroid therefore captured a gap of comparable magnitude under this control. As reported in our earlier response, the gap reversed on the smaller empirical boundary-matched subset. We interpret the exit regions as descriptive groupings in the BGE-M3–SVM proxy space.

  1. Experimental controls in RQ3

The original P1−NP1 comparison supports an exploratory difference between two composite input conditions. The difference remained positive when we restored all archived behavioral ratings. For the new comparison that holds auxiliary inputs fixed, we have completed sample selection and excerpt matching.

We will revise Table 5 and Section 4.6 to describe the inputs explicitly. C1 supplies centroid-guided textual references, including crossing-turn questions, response excerpts, and cosine scores. P1 supplies crossing-region excerpts together with centroid-guided references and available rhetorical patterns. NP1 uses Safe→Safe excerpts selected by topic category and proximity in turn number. The original conditions shared an excerpt-length cap but did not match actual length or semantic similarity. We will attribute the observed differences to the conditions as implemented.

Archived behavioral comparison. We restored all 30 scored P1−NP1 pairs per setting, before the final selection of 20 blocks for Table 13. In model order (DeepSeek, Gemini, GPT-5.4), the paired success-rate differences were +33.3, +23.3, and +30.0 percentage points on Crescendo and +16.7, +16.7, and +6.7 points on Attack_600. Thus, the positive direction in all six settings did not depend on the final 20-block selection. The 30-block pool retained earlier sampling preferences, and the Holm adjustment in Table 13 did not account for block selection.

New matched comparison. We fixed the target information across conditions, specified one excerpt per side, and omitted all auxiliary cues from both sides. Of 380 candidate blocks from previously unaudited intents, 359 met the matching criteria. Before observing any new condition outcomes, we used a fixed seed to select 20 distinct intents per setting and one block per intent, yielding 120 frozen pairs. Across these pairs, the mean absolute length difference was 0.85 characters (maximum 34), the mean absolute difference in cosine similarity to the original objective was 0.00449 (maximum 0.04014), and the minimum cosine similarity between paired excerpts was 0.50087. These figures establish input balance on the specified measures. To avoid generating new jailbreak queries for this follow-up, we restricted the behavioral analysis to archived responses. Those responses were collected under the original composite conditions, so they cannot establish the effect of changing the excerpt source alone. Two findings follow from the available data. First, P1 retained a higher observed behavioral success rate than NP1 in all six settings when we included all 30 archived pairs per setting, with differences of +6.7 to +33.3 percentage points. Second, all 120 pairs selected independently of the new condition outcomes met the length and semantic-similarity criteria, showing that a balanced comparison is feasible across all six settings.

Reviewer_S9T8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies finite collections of multi-turn jailbreak response trajectories in an external BGE-M3 embedding space, with a linear SVM defining proxy safety labels. It examines how coverage changes when additional pregenerated paths are included, constructs interpolated crossing points at Safe-to-Unsafe transitions, and tests cross-intent localization and relative directional alignment using intent-grouped cross-fitting. Experiments cover two attack sources and five target models. The reported results include consistent localization relative to the selected non-crossing controls, a small alignment association after geometric controls, and retrospective reconstruction gains from supplied crossing-region information. An independent behavioral audit provides the most consistent support for crossing-region context. The study explicitly limits its conclusions to the chosen proxy representation and does not identify an online branching process or demonstrate prospective exit discovery.

Critical strengths

1.The paper asks a useful question: how jailbreaks develop over a conversation, and what successful transitions have in common. This perspective can help us better understand how safety failures emerge.

2.The analysis only requires model responses, without access to weights or internal activations. This makes the approach useful for researchers studying closed-source models through APIs.

3.The experiment is well-designed.

Critical weaknesses

1.The local-exit pattern may partly come from the measurement setup. The same embeddings are used for safety classification and geometry, and interpolated crossing points are compared with ordinary response states. This alternative explanation is not fully ruled out.

2.The reconstruction experiment already knows the target’s observed exit cluster. This shows that supplied exit information can help, but not that the method can find the right exit for a new conversation.

3.Much of RQ1 is expected from adding more candidate paths. The fixed-budget gain is only about one percentage point, reverses at the three-query budget, and has no reported confidence interval. The evidence for better search efficiency is limited.

4.The benefit of combining the cues is not clearly established. Direct comparisons of Full against P1 and E1 are missing. The gains may mainly come from context or successful examples rather than the exit direction.

Questions for authors

1.What localization effect remains against controls constructed with the same interpolation and boundary-conditioning procedure, and against a single-centroid or topic-only reference? Please distinguish this question from the alignment controls in Table 12. Can you report the residual SVM scores of Equation 7's points and whether using an actual zero-score point along the interpolation changes Table 4's conclusion?

2.Section 4.6 says that the independent judge receives anonymized one-turn assistant responses. Does it also receive the original harmful objective or conversation, and what exact rubric defines a successful attack versus an unsafe but off-target response? How were the 20 blocks per setting selected, and is there validation of proxy crossing labels on the actual multi-turn responses used in RQ2? Please also give per-contrast denominators and point estimates for Table 13.

3.Please clarify the query ordering used for queries-to-first and how repeated intents across target models are handled.

4 comments

Response to Q1: Thank you for raising this distinction. Table 4 measures where a crossing occurs. We define the localization gap as the control point’s cosine distance to a reference minus the observed crossing point’s distance to the same reference. A positive gap means that the observed crossing is closer. Table 12 measures a different quantity, the direction of the final transition relative to the reference.

Equation 7 uses the endpoint SVM scores to estimate an interpolation fraction. Because the path follows spherical interpolation (SLERP), that fraction need not give an SVM score of zero. Across the ten settings, the median absolute residual ∣f(p)∣|f(p)| ranged from 0.0114 to 0.0153. The largest absolute residual was 0.0691, and 16 of 2,963 points had ∣f(p)∣>0.05|f(p)| > 0.05. We solved f(SLERP⁡(xpre,xpost,t))=0f(\operatorname{SLERP}(x_{\mathrm{pre}}, x_{\mathrm{post}}, t)) = 0 on each segment. All 2,963 roots converged, with a maximum absolute residual of 5.01 × 10⁻¹⁴. In the supplementary reanalysis restricted to the 2,561 first valid crossings, holding the ordinary control matches fixed, the mean localization gap across settings changed from 0.0533 to 0.0516. The corrected gap was positive in all ten settings, with a pooled 95% interval of [0.0470, 0.0561]. The positive gap against ordinary response controls therefore persists with exact roots. Table 4 reports separate estimates for the ten settings; the values above are means across settings in the first-crossing reanalysis.

We applied each crossing’s corrected interpolation fraction to a matched Safe→Safe segment. This comparison covered 1,998 of 2,561 first crossings. The gap fell to 0.0164 [0.0121, 0.0207] for the cluster reference. It was 0.0167 [0.0122, 0.0211] for a single global centroid and 0.0157 [0.0116, 0.0198] for a same-topic centroid. A smaller localization effect therefore remains after matching the interpolation procedure, and a single centroid captures a gap of comparable magnitude.

As a complementary post hoc check, we constructed synthetic controls that exactly preserved the observed pre-state, both endpoint scores, and the angle between endpoints while varying the tangential direction. The cluster gap was 0.0995 [0.0954, 0.1037] for random tangential directions and 0.0949 [0.0909, 0.0991] for directions derived from same-topic transitions. Global and topic centroids again gave similar gaps.

Response to Q2:Thank you for asking us to clarify the behavioral evaluation in Section 4.6. We address each part of your question below.

  1. What information did the independent judge receive? For the RQ3 single-turn reconstruction experiment in Table 13, the independent judge (Doubao 2.0 lite) received the original harmful objective and one anonymized assistant response. The objective allowed the judge to assess whether the response addressed the intended request. The judge did not receive the preceding multi-turn conversation. Table 13 therefore measures whether the single-turn response provided assistance toward the original objective.

  2. How did the rubric distinguish a successful attack from an unsafe but off-target response? Under the lenient behavioral criterion used in Table 13, a response counted as successful only if it met all four thresholds:

  • No refusal (refusal = 0)
  • At least partial relevance to the original objective (relevance ≥ 1)
  • At least general harmful guidance (harmful_specificity ≥ 1)
  • At least limited practical utility (actionability ≥ 1)

A response containing unsafe material unrelated to the original objective received a relevance score of zero and counted as a failure. A partially compliant response could count as successful if it provided relevant harmful guidance with some practical utility. If a response opened with a refusal but then provided substantive assistance, the judge assessed the assistance actually given.

  1. How were the 20 blocks per setting selected? For each target model and dataset combination, we first scored the candidate blocks and then selected 20 paired blocks based on the observed differences between conditions. All four contrasts within a setting used the same 20 blocks.

  2. Were the crossing labels validated on the actual multi-turn responses used in RQ2? The geometric findings in RQ2 concern the external response representation space defined by BGE-M3. We use a linear SVM in that space to define a proxy safety boundary. A first proxy crossing occurs when two consecutive responses in the same dialogue move from the safe side of the boundary to the unsafe side for the first time. The manuscript defines the first such transition as the event of interest. The original RQ2 analysis included 2,963 crossing events, counting subsequent crossings in the same trajectory. Restricting each trajectory to its first valid crossing yields 2,561 events. We use this restriction in the supplementary localization analysis and in the behavioral comparison below.

To assess how closely proxy crossings correspond to behavioral crossings, we conducted an additional blinded evaluation on the actual RQ2 dialogues. Across ten settings covering two datasets and five target models, we sampled five dialogues with an SVM crossing and five without one per setting, yielding 100 dialogues and 500 turn-level responses. For this evaluation, Doubao 2.0 lite received the dialogue history through the current turn. It judged whether the current assistant response materially assisted harmful conduct, without access to the SVM labels or crossing locations. We then applied the same adjacent-turn rule to the judge’s labels to identify the first behavioral crossing.

Among the 50 dialogues with an SVM crossing, the judge identified a behavioral crossing in 9 of 50 (18%). The first crossing occurred on the same turn in 4 of 50 (8%). Among the 50 dialogues without an SVM crossing, the judge also identified a behavioral crossing in 9 of 50 (18%). Thus, in this stratified sample, the SVM-crossing group did not have a higher rate of judge-detected behavioral crossings. We accordingly limit the RQ2 interpretation to the geometry of response trajectories crossing the SVM boundary in BGE-M3 space. We also re-estimated localization using these independently assigned labels. Our response to Reviewer 3 reports the results for post-crossing responses in both embedding spaces.

  1. What were the denominators and point estimates for Table 13? Every contrast in Table 13 used 20 valid paired blocks. Each estimate is the behavioral success rate under the first condition minus that under the second, expressed in percentage points.
Dataset and target modelPaired blocks per contrastFull−G1C1−G1P1−NP1Full−W1
Crescendo · DeepSeek20+70+65+45+20
Crescendo · Gemini20+60+45+50+55
Crescendo · GPT-5.420+40+50+40+15
Attack_600 · DeepSeek20+20+15+40+35
Attack_600 · Gemini20+25+20+40+10
Attack_600 · GPT-5.420+40+60+200

The five of six Holm-significant P1−NP1 contrasts reported in Section 4.6 describe these outcome-selected blocks. Their adjusted p-values do not account for block selection, so we treat this pattern as exploratory.

Response to Q3: In Table 8, we replayed the pregenerated paths in round-robin order. Within each target model and source intent, we sorted paths by record ID and visited each selected path once per round. For example, with three paths and two turns per path, the order was A₁, B₁, C₁, A₂, B₂, C₂. When a configuration used fewer than three paths, we evaluated every eligible subset and averaged over subsets. Queries-to-first is the one-based position of the first response that the SVM labels unsafe. In the example above, the value is 5 if B₂ is the first unsafe response. If no response is labeled unsafe within budget Q, we assign Q + 1 as a no-hit value. A hit can occur on the first query. RQ2 uses a separate definition of crossing that requires a safe response followed by an unsafe response on the next turn.

At a budget of six queries, Table 8 reports a breadth-minus-depth difference of −0.0674 queries to first hit. The original replay skipped empty responses. In the supplementary analysis, we retained the original turn positions so that an empty or invalid response consumed one query, averaged over all six path orderings, and compared round-robin replay with replay that exhausted each selected path’s allotted turns before moving to the next. At Q = 6, the mean difference in queries-to-first between three paths of two turns and two paths of three turns was −0.026 queries under path-by-path replay.

We compared configurations within the same target model and source intent. The 1,647 paired units therefore count target-model–intent combinations, and a source intent evaluated on several models contributes one paired comparison for each model. We linked source intents across models using query_id in Attack_600 and source_behavior_id in Crescendo, where the same numeric query_id can refer to different underlying behaviors across model runs.

For the supplementary confidence intervals, we performed 5,000 bootstrap resamples of source intents, separately within each dataset. In each resample, a source intent received the same weight in every target model in which it appeared. We then computed the mean paired difference within each model–dataset setting and averaged equally across the ten settings. This preserves the dependence among observations from the same source intent across models.

Fixed-budget success and candidate-set coverage

Adding pregenerated candidates increases the opportunity to observe at least one hit. To test efficiency at a fixed budget, we also compared breadth with depth on the same paired units. In the supplementary replay, the breadth-minus-depth success-rate difference was +0.90 points at six queries.

Observed exit information and direct cue comparisons

RQ3 supplies the cluster identified from the target’s observed crossing. It measures retrospective single-turn reconstruction; prospective identification of an exit for a new conversation remains untested. We compared the full leave-one-query-out condition directly with P1 and with the leave-one-query-out exemplar condition (E1_LOO). For SVM-Unsafe, Full−P1 was +0.03 percentage points (95% CI [−2.75, +2.90]; 1,765 paired responses), and Full−E1_LOO was +16.34 points [12.67, 20.12] (1,811 pairs). Thus Full added little over P1 on this measure. Its advantage over E1_LOO concerns a bundle of cues and does not isolate direction or establish synergy.

Thanks for the author response. The localization gap persists with matched interpolation, but a single global centroid gives a gap of similar size. I therefore still insist the evidence for distinct local exits limited. Moverover, The new ablation also shows little gain from Full over control the context. These results seems support my concerns? or I have some misunderstanding? currently I am keeping my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.