acceptodds
Under review as a conference paper at ICLR 2027

Multiple Turns, Short Horizons: Measuring Effective-Horizon Compression in Jailbreaks

Abstract

Multi-turn jailbreaks use feedback from a victim large language model to adapt their queries over several turns. However, once an effective trajectory has been found, it is unclear how much of the earlier query history is still needed. In this work, we study this by removing early attacker queries and running the remaining queries again in a new conversation while regenerating the victim's responses. We define the effective horizon as the shortest remaining query sequence that recovers the same attack progress. Empirically, we find that roughly half of the original query sequence is sufficient on average to recover the same attack progress. We call this phenomenon effective-horizon compression. We show that this shorter history can recover the best point reached during an unsuccessful attack and provide a better starting point for continuing the attack. It also helps identify which earlier queries should receive less learning signal during attacker training, and downweighting those queries improves performance. We further find that short reward-propagation windows preserve calibrated defense performance close to full-trajectory returns. These results show that multi-turn attacks can contain more queries than are needed to recover the progress reached later in the interaction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-turn Jailbreak, LLM Safety, Temporal Credit Assignment, Effective Horizon

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mdW1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper shows that not all turns in a multi-turn jailbreak contribute equally to attack success and examines the role of earlier turns. It defines the Effective Horizon as the shortest sequence of recent turns needed to reproduce the original attack result. Based on this idea, the authors propose Effective-Horizon Context Recovery (EH-CR) and Effective-Horizon Credit Assignment (EH-CA). Experiments show that both methods improve attack performance across multiple benchmarks and LLMs.

Critical strengths

The paper presents a clear research question and a simple way to measure how much conversation history is needed to reproduce an attack. The experiments using Effective Horizon suggest ways to improve existing multi-turn jailbreak methods. EH-CR performs better than both Full Context and Peak Context, showing that removing unnecessary earlier turns can improve attack continuation beyond simply returning to the highest-scoring turn. EH-CA further improves ASR@3 by 6.9 percentage points on average across nine evaluation settings.

Critical weaknesses

(1) The main finding has limited novelty. The paper shows that earlier turns in multi-turn jailbreaks can often be removed without losing attack progress. However, M2S [1] has already shown that multi-turn attacks can be compressed into single-turn prompts, while Yang et al. [2] questioned how much multi-turn attacks benefit from conversation history. Effective Horizon measures this more directly, but the main observation is closely related to these earlier findings. (2) EH-CR, one of the two main applications, is a relatively simple method. It returns to the highest-scoring turn and removes earlier turns while preserving the attack score. Although EH-CR improves ASR over Full Context and Peak Context, the paper does not compare it with simpler methods, such as keeping a fixed number of recent turns under the same query budget. Appendix C reports query cost, but only for Full, Peak, and EH-CR. (3) The findings may not apply to all types of multi-turn jailbreaks. The effective horizon is measured only for X-Teaming and TROJail. ActorBreaker is used for EH-CR, but its EHR is not reported, and attacks such as Crescendo or FITD, where earlier turns build up context before the harmful request, are not evaluated. It is unclear whether the "roughly half" result holds for these attacks.

[1] Ha, Junwoo, et al. "M2S: Multi-turn to Single-turn jailbreak in Red Teaming for LLMs." ACL. 2025. [2] Yang, Xiaoxue, et al. "Multi-turn jailbreaks are simpler than they seem." arXiv preprint (2025).

Questions for authors

(1) How does EH-CR compare with simply keeping the last few turns under the same total query budget? (2) EH-CR already computes the effective horizon for ActorBreaker trajectories. Could you report the EHR for ActorBreaker as in Figure 2? Does Effective Horizon compression also work for attacks like Crescendo [3] or FITD [4], where earlier turns build context for the final harmful request?

[3] M. Russinovich, A. Salem, and R. Eldan, “Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack,” USENIX Security, 2025. [4] Z. Weng et al., “Foot-In-The-Door: A Multi-turn Jailbreak for LLMs,” EMNLP, 2025.

4 comments
AuthorsOpenReview ↗

Response to Reviewer mdW1

We would like to thank the reviewer for the constructive and detailed feedback.

  1. Novelty Relative to Prior Work

The key distinction is between whether multi-turn interaction improves attack effectiveness and how much preceding history a particular attack actually requires. M2S and Yang et al. examine the former from different perspectives, while Effective Horizon directly measures the latter.

Specifically, M2S consolidates the sequence of user prompts into a single structured request, retaining earlier user prompts while omitting intermediate model responses. Its results therefore do not establish which earlier user turns can be removed. Yang et al. compare multi-turn attacks with single-turn attacks given comparable retry opportunities, showing that additional attempts can largely explain the observed multi-turn advantage in their evaluated settings. This comparison does not determine how much history a given attack trajectory needs to preserve its attained progress.

Effective Horizon addresses this question through direct history truncation, measuring how much preceding context can be removed while preserving attack progress at a given turn. Although the broader motivation is related, our contribution is to characterize the extent of history dependence within attack trajectories. We will include both studies in the Related Work section and explicitly clarify how Effective Horizon’s measurement of history dependence differs from prompt consolidation and retry-based comparisons.

  1. Comparison with Fixed-Window Baselines under Matched Query Budgets

We add two simple fixed-window baselines in the X-Teaming/Qwen3-30B-A3B continuation setting: retaining the last two turns of the original trajectory, and retaining the two-turn context ending at the highest-scoring turn. For each trajectory, both baselines are capped at the victim-query budget used by EH-CR, including its replay costs.

As shown in the table below, EH-CR achieves larger continuation gains than either fixed-window baseline. This comparison isolates the benefit of adaptive suffix selection beyond returning to the peak and retaining a fixed amount of recent history. EH-CR exceeds the peak-aligned two-turn baseline by 8.80 percentage points, despite accounting for its replay costs.

MethodAdditional SuccessesΔASR from Continuation
Last two turns11/159+6.92
Last two turns ending at the peak22/159+13.84
EH-CR36/159+22.64

To additionally account for the additional replay cost, we compare methods under matched victim-query budgets in the X-Teaming/Qwen3-30B-A3B continuation setting. The table below reports the resulting ΔASR from continuation.

EH-CR exceeds both baselines by 8.18 percentage points at 10 queries and 6.28 points at 15 queries, showing that its continuation benefit persists when replay costs are included in the matched query budget. We will extend this budget-matched evaluation to additional attack methods and victim models and report the results in the next revision.

Query budgetFullPeakEH-CR
10 queries5.66%5.66%13.84%
15 queries8.81%8.81%15.09%
  1. Generalizability across Multi-Turn Attack Methods

We extend the effective-horizon analysis to Crescendo and ActorBreaker, following the protocol in Fig. 2(a). This adds an attack that progressively builds context and fills the gap in EHR measurements for ActorBreaker.

As shown in the table below, both attacks exhibit a pattern broadly consistent with Fig. 2(a). In particular, five-turn trajectories have mean EHRs of 0.50 for Crescendo and 0.42 for ActorBreaker, compared with 0.51 for X-Teaming. These results provide additional evidence that substantial effective-horizon compression extends across different multi-turn attack strategies. We will include these results in the next revision.

Original LengthCrescendo: NCrescendo: Mean EHRActorBreaker: NActorBreaker: Mean EHR
2470.787270.7857
3300.6222100.7333
4220.6250200.6625
580.5000220.4182

We would like to thank the reviewer again for their time and constructive feedback. We believe our responses and revised manuscript address the main concerns. We would be happy to clarify any remaining questions during the discussion period.

I appreciate authors’ clarification and extra experiments. Although the fixed window is set to 2 turns and fixed-window baseline is not included in the extra experiment, the authors response reasonably addressed my concerns. Regarding generalizability, the idea of “roughly half of the history is sufficient” is still unclear. In the additional X-Teaming experiments, history reduction can be confirmed but a larger fraction of the history needs to be retained (0.64 to 0.7). Thus, it is still unclear whether these quantities come from the attack difficulty or longer trajectory and “roughly half” idea need more clarification. Additionally, authors explained the difference of the paper compared with the previous approaches but the technical contribution remains limited. Therefore, I will maintain the original score.

We thank the reviewer for their thoughtful follow-up and for acknowledging our additional experiments.

Regarding the fixed-window baseline, we clarify that the window size was set to 2 turns to match the average number of turns retained by EH-CR in this setting, enabling a comparison under comparable context lengths.

Regarding generalizability, we would like to clarify that "roughly half" is intended as an approximate empirical observation rather than a universal threshold. The exact EHR may vary across attack methods, trajectory lengths, and attack difficulty. More importantly, our central finding is that substantial portions of the interaction history can be removed without losing attack progress. This phenomenon is consistently observed across multiple attack strategies, including X-Teaming, TROJail, Crescendo, and ActorBreaker, supporting the generalizability of history redundancy rather than any particular numerical ratio.

Regarding technical contribution, we would like to emphasize that Effective Horizon introduces a systematic framework for measuring history dependence within multi-turn attack trajectories, rather than merely compressing prompts. Beyond revealing this phenomenon, the Effective Horizon concept also provides actionable insights for both attack and defense, leading to improvements in attack continuation and defense training. We believe these findings and applications demonstrate the broader significance of understanding history dependence in multi-turn jailbreaks.

We thank the reviewer again for their constructive feedback and would be happy to discuss any remaining concerns.

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

This paper examines how much of a multi-turn jailbreak's history remains necessary after effective attack queries have been found. It defines the effective horizon through suffix replay with unchanged attacker queries and regenerated victim responses. The authors report that roughly half the original turns often suffice, then use this observation for attack continuation and credit assignment.

Critical strengths
  1. The paper introduces an interesting perspective by questioning the necessity of full query history in multi-turn jailbreaks and shifting the focus toward effective-horizon compression.
  2. The proposed context recovery method offers a practical solution to improve the continuation of unsuccessful attacks by trimming redundant historical context.
  3. EH-CR is evaluated against both full context continuation and peak context recovery across six victim models and two attack frameworks, providing meaningful evidence that removing redundant history can improve attack continuation.
Critical weaknesses
  1. The effective horizon is estimated from individual suffix replays, but the paper does not report replay variability or the frequency of undefined horizons, leaving the reliability of this measurement unclear.
  2. The compression analysis is limited to trajectories of at most five turns, and TROJail has no successful five-turn trajectories, making it difficult to assess whether the findings generalize to longer and more complex interactions.
  3. Could there be cases where an early turn is important to the final jailbreak while some intermediate turns are irrelevant, which the current approach (contiguous suffixes) may fail to capture?
  4. The defense experiments use only one TurnGate configuration with a single training seed, and the fixed reward windows are not derived from measured effective horizons, limiting both the generalizability.
Questions for authors

How many replay trials were conducted per candidate suffix to account for generation stochasticity? Can the authors provide empirical evidence (even on a smaller subset of behaviors) that this compression phenomenon holds for longer, more complex interactions (e.g., 8 to 10 turns)? In multi-turn dialogues, it is plausible that an early turn establishes a crucial persona or constraint, while intermediate turns act as irrelevant filler. Did the authors observe any qualitative cases where an early turn was strictly necessary for the final jailbreak, but an intermediate turn was not? Could the authors report the variance or results across multiple random training seeds to ensure that the observed "little effect" of long-range propagation (e.g., the 0.16% standard deviation in F1$_{\phi2}$) is not an artifact of this specific seed initialization?

3 comments
AuthorsOpenReview ↗

Response to Reviewer uUdc (Part I)

We would like to thank the reviewer for the constructive and detailed feedback.

  1. Reliability of Effective-Horizon Estimation

We repeat the suffix-replay procedure twice on the trajectories used in Fig. 2(a), yielding three runs in total. The table below reports the mean EHR across runs for each trajectory-length group. Standard deviations range from 0.008 to 0.016, supporting the stability of the aggregate EHR measurements.

Original Length# TrajectoriesOriginal RunRerun 1Rerun 2Mean ± SD
2550.72730.74550.73640.7364 ± 0.0091
3670.55720.56220.53230.5506 ± 0.0160
4370.58110.56760.58110.5766 ± 0.0078
590.51110.53330.53330.5259 ± 0.0128

We apologize for not making this explicit in the paper. Trajectories with undefined effective horizons are included in all aggregate statistics with EHR = 1, rather than excluded. Among the 168 trajectories in Fig. 2(a), approximately 25% have EHR = 1. We will clarify this convention in the next revision.

  1. Generalization to Longer Trajectories

TROJail uses a trained attacker whose rollout lengths are shaped by its training setup and learned policy. Its observed trajectory-length distribution therefore differs from that of X-Teaming, and the absence of successful five-turn trajectories in our TROJail sample should not be interpreted as a general limit on multi-turn attacks.

To directly examine longer interactions, we additionally run X-Teaming with a 15-turn budget on 30 selected difficult behaviors. We focus on difficult behaviors because attacks that succeed early do not provide long successful trajectories for compression analysis. The table below reports results grouped by actual trajectory length.

For these selected difficult behaviors, EHRs are higher than the roughly 0.5 observed in our original experiments. This may reflect the need for more sustained contextual buildup, requiring a larger fraction of the original history to be retained. Nevertheless, the average history reductions of 30–36% provide evidence that effective-horizon compression extends to longer interactions, even in these challenging settings.

Actual Trajectory LengthThe number of TrajectoriesMean EHR
L ≥ 5170.7014
L ≥ 880.6368
L ≥ 1060.6639
  1. Non-Contiguous Turn Dependencies

We implement the Subsequence Pruning, which fixes the highest-scoring final turn and searches over order-preserving subsequences of the preceding history. Candidates are evaluated from shortest to longest, with ties broken by a fixed lexicographic order. This approach encompasses all suffix candidates while additionally allowing the removal of intermediate turns.

On the same 51 failed X-Teaming trajectories against Qwen3-30B-A3B, this Subsequence Pruning selects non-contiguous contexts in 7 cases. In 2 cases, its continuation succeeds while EH-CR does not, illustrating the potential benefit of retaining early turns while removing intermediate ones. Inspection of these cases suggests that some omitted queries repeat requests or framing already present in retained turns, allowing non-adjacent queries to remain semantically coherent. However, pruning also changes the response history and may remove refusals, so redundancy is a plausible explanation. These results show that non-contiguous contexts can be useful in individual cases, but their additional flexibility does not improve aggregate ASR in this experiment and incurs higher query costs.

AuthorsOpenReview ↗

Response to Reviewer uUdc (Part II)

  1. Defense Robustness

We expand the defense evaluation along two axes: two additional training seeds (43 and 44) under the original configuration, and an additional configuration with the early-penalty multiplier reduced from 1 to 0.5. The results are shown in the table below. Each entry reports detection performance before → after validation calibration.

Across these additional runs, the large differences between reward windows before calibration largely disappear under the validation-calibrated setting. This supports the robustness of our observation beyond the original seed and configuration.

Early-Penalty MultiplierSeedWindow 0Window 3Full
1 (original)4340.41 → 66.1255.84 → 66.6265.13 → 66.58
1 (original)4442.07 → 66.2957.40 → 66.0065.17 → 66.23
0.5 (additional)4250.40 → 66.1460.23 → 66.0965.98 → 65.19
  1. Rationale for Using Fixed Reward Windows Instead of Measured Effective Horizons

Computing an effective horizon for every training turn would require extensive replay and incur substantial computational cost. We therefore use fixed reward-propagation windows as a coarse sensitivity analysis of how the scope of historical credit affects detection. Across the tested seeds and configurations, detection performance after correction varies little across windows. This suggests that changing the propagation window alone is insufficient under the current defense formulation. The experiment thus highlights a limitation in exploiting temporal credit, rather than directly validating EH-guided defense. We will clarify this purpose and limitation in the next revision.

  1. Connection between the defense experiment and effective horizon

Our central contribution is effective-horizon compression, and the attack and defense experiments are initial explorations of its implications. While simple EH-guided interventions improve attack performance, the defense experiment examines whether the current training method benefits from distinguishing the temporal scope of historical influence. Computing an effective horizon for every turn in the training data would require extensive replay and is costly. We therefore use fixed reward-propagation windows as a coarse sensitivity analysis. Across the tested window lengths, we observe little change in detection performance, suggesting that adjusting historical credit propagation alone is insufficient under the current formulation.

This complements our effective-horizon analysis by highlighting a gap: while attack success can depend selectively on history, varying the scope of historical credit does not translate into better detection under the current defense formulation. We do not claim that measured effective horizons would provide no benefit. Rather, this exploratory result motivates defense methods that can better exploit trajectory-specific historical dependencies. We will clarify this purpose and limitation in the next revision.

We would like to thank the reviewer again for their time and constructive feedback. We believe our responses and revised manuscript address the main concerns. We would be happy to clarify any remaining questions during the discussion period.

Reviewer_3enHOpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies a phenomenon called effective-horizon compression, in which part of the multi-turn adversarial prompt could also retain the attack capability. The experiments show that effective-horizon compression could benefit search-based attacks via context recovery and multi-turn attack learning on credit assignment. The authors also study whether the same distinction appears in defense training.

Critical strengths
  • The concept of effective-horizon compression is interesting and valuable. The average effective-horizon ratio (EHR) of 50% for X-Teaming and TROJail is somewhat surprising.
  • The authors explore different directions for applying the EHC concept.
Critical weaknesses
  • The method invoves repeatedly replaying candidate query sequences, introducing additional query costs. A comparison under matched total query budgets would make the efficiency gains more convincing. In addition, the method only removes queries from the beginning of a trajectory. Intuitively, any turn could be potentially removable with maintained attack capability. In that sense, a more general approach is to pruning trajectory arbitrarily. The author probably could show that the proposed method is better than other forms of pruning.
  • EH-CR is evaluated on X-Teaming and ActorBreaker, and EH-CA is tested on TROJail. . Evaluating EH-CR on more multi-turn attackers, e.g. SEMA (https://arxiv.org/abs/2602.06854), Crescendo (https://arxiv.org/abs/2404.01833), CoA (https://arxiv.org/abs/2405.05610), etc, and comparing EH-CA with TRACE (https://arxiv.org/abs/2605.08778) would help demonstrate the generalizability and outperformance of the proposed methods.
  • The defense experiment only studies fixed reward-propagation windows, without explicitly measuring or using the effective horizon. It is therefore unclear how this experiment directly supports the paper's central idea of effective-horizon compression.
Questions for authors

None.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3enH (Part I)

We would like to thank the reviewer for the constructive and detailed feedback.

  1. A comparison under matched total query budgets

To account for the additional replay cost, we compare methods under matched victim-query budgets in the X-Teaming/Qwen3-30B-A3B continuation setting. The table below reports the resulting ΔASR from continuation.

EH-CR exceeds both baselines by 8.18 percentage points at 10 queries and 6.28 points at 15 queries, showing that its continuation benefit persists when replay costs are included in the matched query budget. We will extend this budget-matched evaluation to additional attack methods and victim models and report the results in the next revision.

Query budgetFullPeakEH-CR
10 queries5.66%5.66%13.84%
15 queries8.81%8.81%15.09%
  1. A Comparison with more general pruning

To test whether removing intermediate turns provides additional benefits, we implement Subsequence Pruning, which fixes the highest-scoring final turn and searches over order-preserving subsequences of the preceding history. Candidates are evaluated from shortest to longest, with ties broken by a fixed lexicographic order. This approach encompasses all suffix candidates while additionally allowing the removal of intermediate turns.

We evaluate both methods on the same set of X-Teaming trajectories using Qwen3-30B-A3B, with identical continuation plans and a fixed budget of five continuation turns. Both methods achieve the same ASR, but Subsequence Pruning requires 2.3× more replay-validation queries (321 vs. 139). These results suggest that the additional flexibility of intermediate-turn pruning does not improve attack success in this setting, despite its substantially higher validation cost. Thus, EH-CR replay offers a favorable trade-off between history compression, validation cost, and attack success.

Moreover, On the same 51 failed X-Teaming trajectories against Qwen3-30B-A3B, Subsequence Pruning selects non-contiguous contexts in seven cases. In two cases, its continuation succeeds while EH-CR does not, illustrating the potential benefit of retaining early turns while removing intermediate ones. Inspection of these cases suggests that some omitted queries repeat requests or framing already present in retained turns, allowing non-adjacent queries to remain semantically coherent. However, pruning also changes the response history and may remove refusals, so redundancy remains a plausible rather than definitive explanation. These results show that non-contiguous contexts can be useful in individual cases, but their additional flexibility does not improve aggregate ASR in this experiment and incurs higher query costs.

  1. Generalization to Crescendo

We select X-Teaming and ActorBreaker for their strong attack performance. To further assess generalizability within the rebuttal time constraints, we additionally evaluate Crescendo, a classic multi-turn jailbreak method, on Qwen3-30B-A3B. As shown in the table below, EH-CR increases ASR from 71.1% to 88.7%, outperforming Full and Peak by 10.1 and 2.5 percentage points, respectively. In the revised version, we will expand the evaluation to additional jailbreak methods across all six victim LLMs and report the detailed experimental setup and results.

AttackVictim ModelBase ASRFullPeakEH-CR
CrescendoQwen3-30B-A3B71.1+7.5 (78.6)+15.1 (86.2)+17.6 (88.7)
AuthorsOpenReview ↗

Response to Reviewer 3enH (Part II)

  1. Comparison with TRACE

We present a comparison with TRACE on HarmBench and JailbreakBench in the table below. Our method achieves a higher ASR than TRACE on HarmBench against Llama3.1-8B-IT (95.3% vs. 92.5%). Although TRACE performs better in the remaining three settings, our method still achieves competitive performance across these benchmarks.

Our method isolates the benefit of effective-horizon-based credit assignment by changing only the outcome credit before the effective horizon, keeping the rest of TROJail unchanged. While TRACE combines several mechanisms, including leave-one-turn-out attribution, harmfulness and relevance penalties, and an optional refusal penalty, our simple intervention already substantially improves the original baseline.

We do not view this simple weighting rule as the best possible use of the effective-horizon principle. More refined credit allocation within the effective horizon, or integration with complementary turn-level signals, may further improve performance. We regard these as promising directions for future work.

BenchmarkModelTRACEOursΔ (Ours − TRACE)
HarmBenchLlama92.595.3+2.8
HarmBenchQwen98.1195.8−2.31
JailbreakBenchLlama96.490.3−6.1
JailbreakBenchQwen100.097.0−3.0
  1. Connection between the defense experiment and effective horizon

Our central contribution is effective-horizon compression, and the attack and defense experiments are initial explorations of its implications. While simple EH-guided interventions improve attack performance, the defense experiment examines whether the current training method benefits from distinguishing the temporal scope of historical influence. Computing an effective horizon for every turn in the training data would require extensive replay and is costly. We therefore use fixed reward-propagation windows as a coarse sensitivity analysis. Across the tested window lengths, we observe little change in detection performance, suggesting that adjusting historical credit propagation alone is insufficient under the current formulation.

This complements our effective-horizon analysis by highlighting a gap: while attack success can depend selectively on history, varying the scope of historical credit does not translate into better detection under the current defense formulation. We do not claim that measured effective horizons would provide no benefit. Rather, this exploratory result motivates defense methods that can better exploit trajectory-specific historical dependencies. We will clarify this purpose and limitation in the next revision.

I appreciate the additional experiments, particularly the budget-matched comparison and the evaluation against general subsequence pruning, which substantially address my concerns about the efficiency and design choices of EH-CR. I also appreciate the authors' candid reporting of the TRACE comparison, although EH-CA does not demonstrate a consistent advantage over existing methods. However, I remain unconvinced by the connection between the defense experiment and effective-horizon compression, as the experiment neither measures nor utilizes the effective horizon. Despite the additional experiments, broader validation across different attack methods and models is still needed. Overall, the authors have addressed several of my concerns, so I will increase my rating. However, given the remaining limitations, I will also lower my confidence.

AuthorsOpenReview ↗

Follow-up Response to Reviewer 3enH

We thank the reviewer for acknowledging the additional evidence and updating the rating. We would like to further emphasize two points regarding the remaining concerns.

  1. Broader Validation across Attacks and Models. Due to rebuttal time constraints, some experiments cover only selected settings; we will extend these evaluations to the remaining attacks and victim models in our study and include the full results in the revised manuscript.

  2. Connection between the Defense Experiment and Effective Horizon. Effective-horizon compression motivates us to examine whether defense training benefits from accounting for the temporal scope of historical influence. Because measuring effective horizons throughout the training data is computationally expensive, we vary fixed reward-propagation windows as a coarse sensitivity analysis. The limited differences after calibration suggest that the current defense formulation may not effectively exploit historical information. While this does not establish that history is unimportant for defense, it identifies a potential limitation in how historical dependencies are incorporated into training. The effective-horizon perspective thus helps examine existing design choices and motivate further investigation. We will clarify this motivation in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We thank all reviewers for their constructive feedback. Our main updates are summarized below.

1. Reliability and Generalizability of Effective Horizon. Repeated replay runs show stable EHR measurements. Additional analyses of Crescendo, ActorBreaker, and longer X-Teaming trajectories further support effective-horizon compression across attacks and interaction lengths.

2. EH-CR across Attacks and Query Budgets. EH-CR improves Crescendo’s ASR from 71.1% to 88.7% and retains its advantage under matched query budgets. Comparisons with fixed-window and subsequence-pruning baselines further support its effectiveness and query efficiency in the tested setting.

3. Implications for Attacker Training and Defense. Effective Horizon motivates credit assignment that accounts for which historical turns matter. A simple EH-based weighting change substantially improves TROJail and exceeds TRACE in one setting. Additional defense runs consistently show that changing fixed reward windows alone offers little benefit after calibration, highlighting the need to better capture trajectory-specific historical dependencies.

Due to rebuttal time constraints, some additional experiments cover only selected settings; we will extend these evaluations to all attack methods and victim models in our study and incorporate the full results into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.