acceptodds
Under review as a conference paper at ICLR 2027

Red-TTT: Test-Time Training for Automated Jailbreaking Large Language Models

18 views

Abstract

Large language models remain vulnerable to jailbreaks, and automated red teaming is the standard way to find jailbreaks in large language models at scale. Current methods either draw more samples at test time through search, rewriting, and tree expansion, or train a stronger attacker offline with reinforcement learning. Both share a limitation: once an attack on a specific target behavior begins, the attacker's weights are frozen. Any signal it gathers about the behavior stays in its context window and is discarded afterward. The attacker never adapts its proposal distribution mid-attack, so success depends almost entirely on the sampling budget, and under a budget affordable at scale, many behaviors remain unbroken. We propose Red-TTT, which updates the attacker's parameters during the attack on each behavior. At each round, the attacker samples a group of candidates, scores them against the victim's replies, and takes a policy-gradient step before drawing the next group, so what it discovers about the current victim is consolidated into weights rather than accumulated as context. We also adapt the training objective to red teaming, where success is judged by the single best sample rather than the average. Red-TTT requires only sampling access to the victim and integrates into existing attack pipelines with no other changes. Against the Best-of-N baseline, Red-TTT raises attack success rate from 55.9% to 72.4% on average at a budget of 120 samples, improving over the baseline in every configuration and cracking many behaviors previous method cannot.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreaking, automated red teaming, LLM safety, test-time training

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_MGehOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Red-TTT, which builds on the simple Best-of-N attack and uses the victim's responses to the attack queries as a signal to train the attacker at test time (a per-behavior LoRA adapter on Qwen2.5-7B-Instruct, updated between rounds with a GRPO-style risk-seeking objective). The paper shows that such test-time training substantially outperforms the frozen Best-of-N baseline (average ASR 55.9% → 72.4% on HarmBench) under the same number of attack queries (N = 120 per behavior).

Critical strengths
  1. The idea is simple and well motivated. For LLM-based attackers that sample adversarial prompts, such as the Best-of-N setting in this paper, it is natural to expect that using feedback from the victim model to optimize the attacker helps elicit more harmful responses.
  2. Clear experimental results. In the Best-of-N control experiments with matched attack queries, the ASR gain is consistent across all 12 configurations (+9.0 to +23.0 points), including on top of the RL-trained Jailbreak-R1 attacker.
Critical weaknesses
  1. The reward signal directly uses the same classifier as the evaluation. The main reward term is the continuous probability from the same HarmBench Llama-2-13B classifier (same model and prompt), whose thresholded output forms the classifier component of the final success criterion. Some reward signal for harmful output is clearly needed, and the paper acknowledges the shared classifier and introduces a relevance guard ρ against reward exploitation. However, ρ gates only the StrongREJECT term, not the dominant classifier term, so directly optimizing the evaluation metric still raises a concern about reward hacking. The classifier term has weight 2.0 out of a maximum reward of 3.3, and in Table 2 (ZeroShot × Llama-3-8B) removing it drops ASR to 46.0, roughly the Best-of-N level (49.0). The gain therefore appears to depend heavily on optimizing the evaluation classifier itself.
  2. The main evaluation relies on a single judge model, which is also used in training. The independent-judge check in Table 5 (Gemini-3-Flash) supports the concern raised in Weakness 1. Although the gain over Best-of-N remains positive for all four victims, the average gain on the ZeroShot substrate falls from +19.1 points under the HarmBench classifier to +9.25 under Gemini with the HarmBench prompt and +7.5 under the stricter rubric, so roughly half of the gain does not transfer to an independent judge. Moreover, on Llama-3-8B, gpt-4.1-mini, and DeepSeek-V4-Pro, Gemini confirms a smaller fraction of Red-TTT's successes than of Best-of-N's (69%, 69%, and 81% vs. 77%, 74%, and 87%), which is consistent with optimization against the training classifier. In addition, this independent check covers only the ZeroShot substrate. It also re-scores only one delivered completion per behavior, whose selection rule is not specified and presumably relies on the HarmBench classifier, instead of letting the independent judge select among all 120 responses.
Questions for authors
  1. If an independent judge is used both to select and to score the responses (e.g., best of 120 under Gemini), how large is the gain over Best-of-N for all 12 configurations?
  2. Given that removing the classifier term in Table 2 eliminates most of the gain, if the reward uses a judge different from the evaluation classifier (e.g., train with another classifier or LLM judge, evaluate with HarmBench), does the gain persist?
  3. For Red-TTT vs. Best-of-N, what are the wall-clock time per behavior, the GPU type and GPU-hours, and the total number of victim calls including StrongREJECT grading (which re-prompts the victim, i.e., roughly 120 attack queries plus 120 grading calls per behavior)? Fig. 3(a) compares against larger Best-of-N budgets only for ZeroShot × Llama-3-8B. How does Best-of-N compare at a matched number of victim calls (e.g., N = 240) in the other configurations?

I'll consider raising score if my questions are addressed.

8 comments
AuthorsOpenReview ↗

Response to W1, W2 and Q1, Q2

We thank the reviewer for the detailed and constructive feedback. We conducted additional experiments to investigate evaluator dependence, independently judged attack effectiveness, and computational efficiency.

W1 / Q2.

We appreciate the reviewer's concern that using the evaluation classifier as part of the training reward may introduce evaluator-specific optimization. We first clarify the reward ablation in Table 2. The original comparison inadvertently mixed results from different evaluation subsets and protocols. We will correct the table to ensure consistent evaluation and explicitly report the matched frozen control. Removing the classifier reward achieves 46% ASR, compared with 30% for frozen Best-of-N, retaining a 16-percentage-point advantage.

Cross-evaluator experiments

To investigate whether Red-TTT's effectiveness depends on optimizing the evaluation classifier, we conducted additional experiments using alternative reward judges. We ran the following experiments on ZeroShot × Llama-3-8B hard set:

Reward judge during adaptationRed-TTT ASR
Original classifier reward33.3%
Without classifier reward24.3%
ShieldGemma-9B23.5%
Llama-Guard-3-8B25.5%

When using ShieldGemma-9B or Llama-Guard-3-8B, the alternative judge replaces the original classifier in both reward computation and candidate selection. The HarmBench classifier is used only for final evaluation. Switching to independent judges, Red-TTT continues to generate successful attacks. These experiments demonstrate that its effectiveness is not solely attributable to optimizing the evaluation classifier.

Additional benchmark

We additionally evaluated Red-TTT on AdvBench, using the same attacker and 120-query budget. On ZeroShot × Llama-3-8B settings,Red-TTT achieves 59.0% ASR, compared with 48.0% for frozen Best-of-N, a 11.0 percentage-point improvement. This further demonstrates the generalization.

W2 / Q1.

We thank the reviewer for highlighting that independently judging only the classifier-selected final response does not fully eliminate selection bias. To address this concern, we conducted a more comprehensive evaluation using Gemini-3-Flash as an independent judge. Rather than evaluating only the final selected response, Gemini-3-Flash judges all 120 stored candidate responses per behavior for both Red-TTT and frozen Best-of-N, under the same evaluation procedure. Gemini is never used during training or candidate selection. This analysis covers four configurations for which all candidate responses were available.

Independently judged attack effectiveness across adaptation rounds

We measured the fraction of candidate responses judged harmful by Gemini during the first and last three adaptation rounds. For frozen Best-of-N, its 120 candidates were divided into 15 groups of eight for comparison.

ConfigurationRed-TTT (first → last)Frozen Best-of-N (first → last)
JB-R1 × gpt-oss-20b0.315% → 0.433%0.236% → 0.354%
JB-R1 × gpt-4.1-mini9.52% → 9.95%8.08% → 9.44%
ReNeLLM × gpt-4.1-mini14.29% → 16.82%13.69% → 14.43%
ZeroShot × DeepSeek-V4-Pro4.31% → 10.65%4.31% → 3.88%

The clearest improvement occurs on ZeroShot × DeepSeek-V4-Pro, where the fraction of independently judged harmful responses increases from 4.31% to 10.65% during adaptation, while the frozen control remains approximately unchanged. This provides direct evidence that Red-TTT can improve attack effectiveness under a judge that participates in neither reward computation nor candidate selection. Together with the cross-evaluator experiments in W1, these results indicate that Red-TTT's gains are not exclusively classifier-specific, while also motivating more careful evaluation of reward-model dependence.

AuthorsOpenReview ↗

Response to Q3

Q3. Computational cost and victim-call accounting

We agree that attack-query matching alone does not establish computational efficiency. We therefore measured end-to-end runtime, victim-call usage, and estimated FLOPs.

Runtime and computational overhead

Both methods were profiled on the same 12 behaviors using one NVIDIA RTX PRO 6000 Blackwell (98 GiB), with the attacker, victim, and 13B classifier co-resident.

ComponentRed-TTTFrozen Best-of-N
Attacker generation28.9 s0.9 s
Victim attack calls40.3 s3.2 s
Reward scoring48.8 s43.4 s
LoRA updates11.8 s0 s
Total runtime132.5 s48.5 s

Red-TTT requires 2.73× wall-clock time, corresponding to approximately 7.4 versus 2.7 GPU-hours for 200 behaviors. LoRA updates account for only 8.9% of Red-TTT's total runtime, with the remaining overhead primarily arising from sequential generation and reward evaluation.

Victim-call accounting

Both methods use 120 victim attack calls. Red-TTT additionally performs 120 StrongREJECT grading calls, yielding 240 total calls under this accounting. Frozen Best-of-N performs additional grading only for candidates passing the classifier gate, requiring approximately 1.4–41.7 additional calls per behavior, depending on the configuration. We will clarify that the original comparison matches the number of attack queries rather than total victim-side evaluation calls.

FLOPs and compute-matched comparison

Using measured token counts and analytical FLOPs estimates, Red-TTT requires 4.11 PFLOPs per behavior, compared with 2.15 PFLOPs for frozen Best-of-N (or 1.88 PFLOPs with shared-prefix caching), corresponding to approximately 1.9–2.2× the computation. We further evaluated frozen Best-of-N under larger sampling budgets on ZeroShot × Llama-3-8B:

Frozen budget NFrozen Best-of-N ASR
12049.0%
24057.0%
48060.0%
1,00060.5%
2,04863.5%
Estimated equal wall-clock (~328)59.0%

Red-TTT achieves 66.0% ASR at N=120. At approximately equal wall-clock time, frozen Best-of-N reaches 59.0%, leaving a 7.0-percentage-point advantage for Red-TTT. We also evaluated the estimated FLOPs-matched budget, corresponding to approximately 230–263 frozen candidates, where frozen Best-of-N achieves 57.0–57.8% ASR. These results show that Red-TTT incurs additional computational overhead but retains an advantage under the estimated compute- and wall-clock-matched comparisons. We will explicitly distinguish attack-query efficiency from overall computational efficiency in the revised paper.

We will incorporate the additional cross-evaluator experiments, clarify the evaluation protocol, and report computational costs more transparently. We thank the reviewer again, and hope these clarifications address the concerns. We are happy to answer any further questions.

Thank you for clarifying the reporting errors and providing the full-benchmark coefficient-sensitivity results. The rerun provides useful evidence of robustness across several nonzero coefficient settings. I also appreciate the additional runtime measurements and clarification of victim-call accounting.

My main concern about evaluator dependence remains insufficiently resolved:

  1. Independent-judge evaluation. For the four configurations where Gemini has evaluated all 120 candidate responses, please report behavior-level ASR@120 for both Red-TTT and frozen Best-of-N. The reported fraction of harmful responses in the first and last adaptation rounds measures a different quantity and does not directly establish how much of the original ASR improvement survives independent evaluation.

  2. Alternative reward judges. The ShieldGemma and Llama-Guard experiments would be more informative with explicit attacker/victim configurations, evaluation populations, and corresponding frozen controls under the same candidate-selection procedure and attack budget. Nonzero ASR after replacing the reward judge does not, by itself, establish an advantage from parameter adaptation.

The new sensitivity results address part of the experimental uncertainty, but they do not yet establish the magnitude of the improvement when reward computation and evaluation are separated. I therefore maintain my current rating of 2 (Weak rejection), while remaining open to reconsideration if these comparisons are provided.

Reviewer_5zHnOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether an automated jailbreak attacker can benefit from parameter adaptation during an attack on a single target behavior. Red-TTT divides a fixed attack-query budget into rounds and updates a behavior-specific LoRA adapter using a dense composite reward, an exponentially weighted group-relative advantage, and elite replay. The adapter is reset for each behavior. Experiments on 200 HarmBench behaviors cover three attacker substrates and four victim models. Relative to frozen controls with the same initial attacker and attack-query budget, the paper reports improvements of 9–23 percentage points across all 12 configurations, including improvements over an offline-trained attacker.

Critical strengths
  1. Each Red-TTT configuration is compared with a frozen control starting from the same attacker and using the same number of attack attempts. The improvements across all 12 configurations provide useful evidence that behavior-specific parameter adaptation can add value beyond additional sampling. The gains on JB-R1 are particularly informative because they persist when the attacker has already undergone offline jailbreak training.

  2. Studying attackers that learn from feedback during an attack is relevant to safety evaluation. The paper includes component ablations, independent-judge re-scoring, input-guard experiments, and transfer tests. These analyses strengthen the empirical case beyond the main success-rate comparison, although they do not resolve all questions about the design rationale.

Critical weaknesses
  1. The motivation does not clearly distinguish parameter adaptation, contextual adaptation, and optimization of a proxy reward. The abstract states that a frozen attacker never adapts its proposal distribution during an attack and consequently depends almost entirely on sampling budget. However, PAIR and TAP condition on changing histories or search states, so their conditional proposal distributions can change without parameter updates. Eq. (4) explicitly permits this. The contribution should therefore be framed as adding parameter adaptation to existing feedback mechanisms, rather than introducing adaptation where none existed.

    The best-case versus average-case argument also needs qualification. For a fixed policy with independent samples and binary success probability p, success within N attempts is 1−(1−p)N1 − (1 − p)^N, which increases with the mean binary success rate. This observation does not characterize the full adaptive search process, but shows that the BoN criterion alone does not establish a need for risk-seeking updates. With continuous proxy rewards, the relevant mismatch may exist, but emphasizing the proxy’s upper tail does not by itself resolve it. These issues weaken the explanation of why the proposed design addresses the stated challenges.

  2. The evidence does not sufficiently distinguish the benefit of enabling online training from the value of Red-TTT’s specific training design. The motivation identifies frozen attacker weights as a limitation, and the main comparisons test Red-TTT against attackers whose weights remain fixed during the attack. These are useful controls for asking whether online adaptation helps. However, they do not establish how Red-TTT compares with other RL-based methods that also update the attacker during the attack. JB-R1 is an offline-RL-trained attacker, so its frozen control does not answer this question. Table 2 partially addresses it through a GRPO-style advantage ablation, with 50% ASR versus 66% for Red-TTT, but only on ZeroShot × Llama-3-8B. This is useful evidence, yet its limited coverage leaves the broader advantage of the proposed training design insufficiently established.

    This distinction is particularly relevant because Eq. (9) has a direct precedent in RS-GRPO[1], which appears uncited. For the same candidate group of size n, rewards, and inverse temperature β > 0, Eq. (8) of Jiang et al. (2025) equals n/β times Eq. (9) here. This concerns the advantage construction, not equivalence of the full methods. TTT-Discover is already cited and uses a related entropic objective, although its finite-sample normalization and temperature selection differ. The paper should clarify the inherited and new elements and provide stronger evidence of the value of its modifications relative to alternative online updates.

    The practical increment also depends on cost. Matching attack queries is valuable, but LoRA updates and reward evaluation add computation; Appendix B acknowledges additional victim inference for StrongREJECT scoring. Without runtime and memory measurements, the experiments establish gains at a fixed attack-query budget, while the practical cost of those gains remains unclear.

[1] Jiang, Y., et al. Risk-Sensitive RL for Alleviating Exploration Dilemmas in Large Language Models. arXiv:2509.24261, 2025.

Questions for authors
  1. Motivation and objective. Can you distinguish frozen parameters from a history-dependent proposal distribution and clarify what “mean reward” refers to in the best-case versus average-case argument? What evidence supports the claim that emphasizing the upper tail of the dense proxy reward is the relevant improvement over mean-centered updates in this setting?

  2. Comparison with other online training methods. Since the paper identifies frozen attacker weights as a limitation, how does Red-TTT compare with a simple online RL baseline that also updates the attacker during the attack? Table 2's GRPO-style advantage ablation partially addresses this question, but only for ZeroShot × Llama-3-8B. Could you extend this comparison to other representative configurations, matching the initial attacker, reward signal, trainable parameters, and attack-query budget? This would help distinguish the benefit of adding online training from the benefit of Red-TTT's specific design.

  3. Relationship to prior work. How does Eq. (9) relate to RS-GRPO [1] and the already cited TTT-Discover [2]? Which modifications are specific to Red-TTT, and how do the existing analyses support their value for adaptation under a small attack-query budget?

  4. Reward coefficients. How were ($\lambda_s=2.0$) and ($\lambda_e=0.3$) selected? Please clarify any tuning data and selection procedure, whether the coefficients were fixed before evaluation and shared across configurations, and any available evidence of robustness to reasonable nonzero alternatives. The component-removal ablations do not establish this sensitivity.

  5. Adaptation choice and cost. What motivated the choice of per-behavior LoRA over full-parameter adaptation? Please clarify whether this was primarily a resource decision or informed by effectiveness or stability comparisons, and report the additional adaptation time, peak memory, and reward-evaluation cost relative to the frozen control, including hardware and timing scope. This requests the design rationale and actual overhead; it does not require a new full-parameter training study.

[1] Jiang, Y., et al. Risk-Sensitive RL for Alleviating Exploration Dilemmas in Large Language Models. arXiv:2509.24261, 2025.

[2] Yuksekgonul, M., et al. Learning to Discover at Test Time. arXiv:2601.16175, 2026.

If the author can answer my question, I will consider increasing the score.

7 comments
AuthorsOpenReview ↗

Response to W1, W2, Q1, Q2, Q3

We thank the reviewer for the thoughtful and constructive feedback. We address the distinction between parameter and contextual adaptation, the relationship to existing risk-seeking estimators, additional online RL comparisons, hyperparameter sensitivity, and the choice of LoRA below.

Q1 / W1.

Parameter versus contextual adaptation. We agree that feedback-based methods such as PAIR and TAP can adapt their proposals through evolving conversational histories, even without parameter updates. Our intended distinction is between contextual adaptation through feedback and parameter adaptation through gradient updates. Red-TTT introduces the latter into automated red-teaming, allowing the attacker to update its proposal policy for each target behavior. These mechanisms are complementary rather than mutually exclusive.

Best-case versus average-case optimization. We appreciate the reviewer's theoretical observation. For a fixed policy with independent binary outcomes, the probability of at least one success within N attempts is 1-(1-p)N, which is monotonic in the per-attempt success probability p. Therefore, the Best-of-N objective alone does not mathematically require risk-seeking optimization. Our motivation instead concerns the continuous, multi-component reward used during test-time adaptation. Since many candidates receive low rewards due to refusals or ineffective attacks, emphasizing higher-reward candidates provides a practical mechanism for directing gradient updates toward more promising attack strategies. We will clarify this motivation and avoid presenting risk-seeking weighting as a theoretical necessity.

Q2 / W2.

We thank the reviewer for suggesting a direct comparison with standard GRPO-style optimization. We extended the advantage-estimator ablation to ZeroShot × gpt-oss-20b, using the same initial attacker, reward function, trainable parameters, and 120-query budget. And we evaluate it using the hard set.

Advantage estimatorASR
Risk-seeking28%
Mean-centered GRPO24%

Computational cost. We additionally profiled Red-TTT and frozen Best-of-N on the same 12 behaviors using one NVIDIA RTX PRO 6000 Blackwell (98 GiB). Red-TTT requires 132.5 s per behavior, compared with 48.5 s for frozen Best-of-N (2.73×). LoRA updates account for only 11.8 s (8.9%) of Red-TTT's runtime. The estimated computational costs are 4.11 versus 2.15 PFLOPs per behavior, respectively. To account for this difference, we also evaluated larger frozen sampling budgets. At approximately equal wall-clock time, frozen Best-of-N can generate around 328 candidates, reaching 59.0% ASR, compared with 66.0% for Red-TTT at N=120. Thus, Red-TTT retains a 7.0-point advantage under the estimated wall-clock-matched comparison. We will include these measurements to distinguish query efficiency from computational efficiency.

Q3 / W2.

We thank the reviewer for identifying the connection between our risk-seeking advantage estimator and RS-GRPO. We agree that Eq. (9) is mathematically equivalent to the RS-GRPO estimator up to a positive scaling factor. Specifically, letting wi=softmax(r)i for a group of size n, RS-GRPO's estimator can be expressed as

AiRS-GRPO=nβ(wi−1n)A_i^{\mathrm{RS\text{-}GRPO}} = \frac{n}{\beta}\left(w_i - \frac{1}{n}\right)

which differs from our Eq. (9) by the factor n/. We will explicitly acknowledge this equivalence, cite RS-GRPO, and remove any claim that the advantage estimator itself is novel.

We will also clarify the relationship to TTT-Discover, including the differences in finite-sample normalization and temperature selection. Importantly, Red-TTT's contribution lies in the overall test-time red-teaming framework, rather than in proposing a new advantage estimator. In particular, Red-TTT combines:

  1. Per-behavior parameter adaptation: A lightweight adapter is optimized for each target behavior and reset afterward, rather than training a single policy across behaviors.
  2. Dense red-teaming feedback: A multi-component reward provides optimization signals during test-time interaction.
  3. Elite replay: Previously discovered high-reward candidates are reused to support iterative adaptation.
  4. Integration with existing attack scaffolds: The framework complements contextual feedback mechanisms, as demonstrated by the improvements in Table 3. Regarding the annealed temperature schedule, we additionally compared the default schedule with fixed-temperature alternatives. The results indicate that the temperature schedule is not essential to the effectiveness of the framework. We will therefore present it as an implementation choice rather than an independent methodological contribution.
AuthorsOpenReview ↗

Response to Q4 and Q5

Q4.

The reward coefficients were fixed in our initial implementation before the reported experiments and applied consistently across all configurations, without victim-specific tuning. We will re run the coefficient experiments following the same configurations in Table 2.

Q5.

We chose LoRA following the practice of prior test-time training work, where parameter-efficient adaptation enables practical online updates. Red-TTT performs online updates while the attacker, victim, and 13B reward classifier are co-resident on a single NVIDIA RTX PRO 6000 Blackwell (98 GiB). Full-parameter adaptation of the 7B attacker would substantially increase memory requirements, whereas LoRA enables practical per-behavior optimization under this hardware constraint. We additionally measured the adaptation overhead on the same 12 behaviors:

QuantityMeasurement
LoRA update time11.8 s per behavior
Total Red-TTT runtime132.5 s per behavior
LoRA share of runtime8.9%
Peak attacker memory16.7 GiB
Estimated cost for 200 behaviors7.4 GPU-hours

These measurements show that parameter updates account for a relatively small fraction of total runtime, with much of the remaining cost arising from sequential generation and reward evaluation. We do not claim that LoRA is inherently more effective than full-parameter adaptation. Rather, it provides a practical way to perform behavior-specific test-time updates within the available memory budget. We will clarify this motivation and report the measured overhead in the revised paper. We will revise the paper to distinguish inherited optimization components from our framework-level contributions, clarify the theoretical motivation, and incorporate the additional experimental evidence. We thank the reviewer again, and hope these clarifications address the concerns. We are happy to answer any further questions.

The response does not adequately explain how the newly reported experiments relate to the submitted results. The unexplained differences and missing supporting material prevent a reliable assessment of the new experimental evidence. The following points require direct, specific answers.

  1. Unexplained 19.5-point difference in the risk-seeking result. Q2 reports 28% ASR for risk-seeking on ZeroShot × gpt-oss-20b with 120 queries, whereas Table 1 reports 47.5% for Red-TTT on the same attacker and victim combination at the same budget. Risk-seeking is the advantage construction used in the original method, as specified in Section 3.2, Eq. (9). Adding a GRPO comparison does not itself explain why the risk-seeking reference result changes. Please explicitly state whether this row uses the full default Red-TTT configuration and explain the difference. Any changes to the behavior set, sample size, success criterion, implementation, or other experimental settings must be specified.

  2. Missing configuration details and unexplained changes in the reward ablations. The new sensitivity tables report 42% ASR at the original coefficient values and 43%/44% when λs=0\lambda_s=0/λe=0\lambda_e=0, respectively. Original Table 2 reports 66% for ZeroShot × Llama-3-8B, decreasing to 46%/54% when the corresponding classifier/engagement terms are removed. The response does not identify the attacker and victim configuration, behavior set, or success criterion for the new tables. Please provide these details and confirm that the coefficients retain their meanings in Eq. (6). If the settings are unchanged, the substantial difference in the reference result and the reversal in the component-removal effects require explanation. If the settings differ, this must be made explicit, together with the implications for the original claim that “the classifier is essential.”

  3. The table reference and claimed integration results do not match the original submission. Q1 and Q3 cite Table 3 as evidence of 9.0–26.5-point gains from integrating Red-TTT into PAIR, TAP, JB-R1, and ReNeLLM. However, Table 3 in the original PDF is a notation table, not an experimental results table. Original Table 1 uses PAIR and TAP as frozen search baselines and reports 9.0–23.0-point gains for the three Red-TTT substrates. Please provide the actual table supporting the claims in the response, identify the relevant manuscript version, and include the experimental configurations and paired results with and without Red-TTT. The current reference does not allow these claims to be checked.

Given the unexplained differences in the reported results, the unspecified experimental settings, and the missing results supporting the cited claims, I am lowering my rating. These issues undermine my confidence in the empirical evidence as currently presented. I remain open to reconsidering this assessment if the authors provide a clear, verifiable reconciliation of the reported results, together with the relevant experimental details and supporting tables.

Reviewer_jGsLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RED-TTT, a test-time training framework that adaptively updates a lightweight LoRA adapter on the attacker LLM via group-relative policy gradient during attacks on individual target behaviors, which overcomes the representational bottleneck of frozen-weight search and achieves notable ASR gains over budget-matched baselines across multiple victim models.

Critical strengths

This paper pioneers the application of test-time training to automated red teaming by formulating per-behavior jailbreak optimization as an online, group-relative policy gradient update on a lightweight LoRA adapter, which effectively circumvents the in-context capacity bottleneck of frozen attackers and demonstrates consistent, non-trivial ASR improvements across diverse substrates on the HarmBench benchmark.

Critical weaknesses
  1. Section 3.2 and Section 5 repeatedly claim a "matched budget" (N=RG=120N=RG=120) against the Best-of-N baseline by counting only the number of attacker generation queries. However, this comparison overlooks the computational asymmetry between pure forward sampling and online test-time training is that RED-TTT requires per-round backpropagation, LoRA gradient updates, and dense evaluations across multiple external evaluators, which incurs substantial overhead from inference-training context switching and wall-clock delay. Could you report the actual end-to-end wall-clock time and total FLOPs for both RED-TTT and all baselines, and provide an evaluation of Best-of-N scaled under an equivalent wall-clock time or computational budget?

  2. In Section 3.2, the authors admit that the elite replay pool Ek\mathcal{E}_k is used as "short-horizon behavioral reuse rather than as an unbiased off-policy policy gradient estimator" without importance sampling correction, while Section 4.4 explicitly states that "the KL penalty is omitted because a nonzero KL coefficient produced no successful transitions". Under such an aggressive setting without KL regularization and where advantages increasingly concentrate on top-1 candidates via annealing βk\beta_k (Eq. 9–10), adapting on small groups is highly prone to sample-level overfitting and policy collapse.

  3. The dense reward formulation in Eq. (6) relies on an ad-hoc combination of four heterogeneous signals with fixed hyperparameters (λs=2.0,λe=0.3\lambda_s=2.0, \lambda_e=0.3). Crucially, the ablation in Table 2 reveals that removing the HarmBench classifier term causes the ASR to collapse from 66.0% down to 46.0%—dropping even below the frozen Best-of-N baseline (49.0%). Given that the primary evaluation metric in Table 1 shares the exact same HarmBench classifier, this raises serious concerns regarding reward hacking and evaluation leakage rather than genuine semantic jailbreaking. Could you conduct a sensitivity analysis on the reward hyperparameters (λs,λe\lambda_s, \lambda_e) and provide cross-evaluator verification by training the policy on an alternative reward model and evaluating on an unseen, held-out safety benchmark?

  4. In Section 4.1 and Table 1, all evaluations are conducted exclusively against four standard base/aligned models (Llama-3-8B-Instruct, gpt-oss-20b, gpt-4.1-mini, and Deepseek-V4-Pro), with no evaluation against explicit, hardened defenses. Since RED-TTT inherently relies on continuous, graded compliance signals from the victim to form non-zero advantages (Eq. 6), strict refusal mechanisms or internal representation disruption (such as Circuit Breaker, dynamic activation steering, or external input-output guardrails) could severely flatten the reward feedback. Could your clarify the effectiveness and robustness of RED-TTT when evaluated against victim models deployed with advanced representation-level defenses or active guardrail filters?

Questions for authors
  1. Could you report the actual end-to-end wall-clock time and total FLOPs for both RED-TTT and all baselines, and provide an evaluation of Best-of-N scaled under an equivalent wall-clock time or computational budget?

  2. Could you conduct a sensitivity analysis on the reward hyperparameters (λs,λe\lambda_s, \lambda_e) and provide cross-evaluator verification by training the policy on an alternative reward model and evaluating on an unseen, held-out safety benchmark?

  3. Could your clarify the effectiveness and robustness of RED-TTT when evaluated against victim models deployed with advanced representation-level defenses or active guardrail filters?

4 comments
AuthorsOpenReview ↗

Response to W1 and Q1

We thank the reviewer for the detailed and positive assessment, and for pinpointing the questions so precisely. Several points are addressed with new experiments below.

W1 and Q1.

We agree that matching the number of attack queries alone does not establish computational efficiency. We therefore measured end-to-end runtime and estimated FLOPs, and compared Red-TTT against frozen Best-of-N at larger sampling budgets.

Measured runtime

We profiled both methods on the same subset of 12 behaviors using one NVIDIA RTX PRO 6000 Blackwell (98 GiB), with the attacker, victim, and 13B model classifier co-resident. The following values are median per-behavior measurements.

ComponentRed-TTTFrozen Best-of-N
Attacker generation28.9 s0.9 s
Victim attack calls40.3 s3.2 s
Reward scoring, including StrongREJECT48.8 s43.4 s
LoRA updates11.8 s0 s
Total runtime132.5 s48.5 s
Runtime ratio2.73x1.00x

Extrapolated to 200 behaviors on this GPU, the corresponding costs are approximately 7.4 versus 2.7 GPU-hours. LoRA updates account for 11.8 s, or approximately 8.9% of Red-TTT's total runtime. The remaining overhead arises largely from sequential generation and reward evaluation. Specifically, frozen Best-of-N generates 120 candidates in a single vLLM batch, whereas Red-TTT uses 15 sequential rounds of eight candidates through a training-compatible HuggingFace backend. The current implementation also performs updates through a per-sample Python loop. These are genuine implementation costs, and we will explicitly report them rather than characterize the methods as equally compute-efficient.

Victim-call accounting

Both methods use 120 victim attack calls. Red-TTT additionally performs 120 StrongREJECT grading calls, giving 240 calls per behavior under this accounting. The frozen control performs additional grading only for candidates passing the classifier gate, corresponding to approximately 1.4–41.7 additional calls per behavior, depending on the configuration. Thus, its total victim-call count is not always exactly 120. The original comparison is therefore matched on attack queries, not total victim calls. We will distinguish these quantities explicitly.

FLOPs and compute-matched comparison

Using measured token counts and analytical FLOPs accounting (approximately 2ND for forward passes and 4ND for LoRA training passes):

QuantityRed-TTTFrozen Best-of-N
FLOPs per behavior4.11 PFLOPs2.15 PFLOPs
With frozen shared-prefix caching4.11 PFLOPs1.88 PFLOPs
Relative compute1.9–2.2x1.0x

These are analytical estimates of required computation rather than hardware-achieved FLOPs. The attention-score term, approximately 2.4% of the matrix-multiplication term in this setting, is omitted.

Under the estimated compute-equivalent budget, the frozen control can generate approximately 230–263 candidates, yielding 57.0–57.8% ASR, compared with 66.0% for Red-TTT at N=120 on ZeroShot × Llama-3-8B. Similarly, using measured throughput to estimate an equal-wall-clock budget gives approximately 328 frozen candidates, yielding 59.0% ASR, or a 7.0-point gap relative to Red-TTT.

Scaling the frozen control

We generated 2,048 independent frozen-attacker candidates per behavior across the 200 ZeroShot × Llama-3-8B behaviors (409,600 generations), and calculated the expected ASR.

Frozen budget NASRDifference from Red-TTT@120 (66.0%)
12049.0%+17.0
24057.0%+9.0
48060.0%+6.0
1,00060.5%+5.5
2,04863.5%+2.5
Estimated equal wall-clock (~328)59.0%+7.0

These results show that additional frozen sampling improves performance, but Red-TTT retains an advantage under the estimated compute- and wall-clock-matched budgets.

AuthorsOpenReview ↗

Response to W2, W3, W4 and Q2, Q3

W2.

We thank the reviewer for raising the concern that repeated on-policy updates without importance correction or explicit KL regularization might lead to policy collapse. We conducted additional diagnostics to examine diversity and concentration throughout test-time adaptation. On ZeroShot × Llama-3-8B across 200 behaviors and 15 adaptation rounds, within-group bigram self-similarity increases only from 0.143 to 0.240, while the group's best classifier probability increases from 0.148 to 0.477. Moreover, the fraction of groups containing identical candidates decreases from 0.215 to 0.185. On the evaluated 50-behavior subset, this fraction decreases from 0.44 to 0.18, whereas frozen sampling remains between 0.34 and 0.46 across rounds. These observations indicate that the policy continues to generate diverse candidates while improving attack quality, rather than collapsing to repeated outputs. We also evaluated smaller KL coefficients:

KL coefficientASR
0 (default)66.0
0.0162.0
0.0364.5

The additional KL settings do not consistently improve performance over the default. We will revise Section 4.4 to report these results and clarify that our evidence supports stable behavior over the evaluated adaptation horizon, rather than guaranteeing stability under arbitrary numbers of updates.

W3 / Q2.

Clarifying the reward ablation

We agree that the reward ablation requires clearer evaluation accounting. We will revise the table and caption to specify and report the corresponding frozen control under the same setting. Importantly, removing the classifier-based reward does not eliminate the benefit of test-time adaptation. On the ablation, the frozen control achieves 30.0% , while Red-TTT without the classifier reward achieves 46.0%. This indicates that the observed improvement is not solely attributable to optimizing the classifier reward.

New and Complete sensitivity results

We have now rerun the coefficient-sensitivity experiments on the full 200-behavior HarmBench benchmark, following the standard experimental configuration and evaluation protocol used in the original submission. Experimental configuration Attacker: ZeroShot, using Qwen2.5-7B as the attacker backbone. Victim: Llama-3-8B. Reward coefficients: As defined in Eq. (6), s controls the classifier-based fulfillment reward (default 2.0), while e controls the engagement reward (default 0.3). We vary one coefficient at a time while holding the other at its default value.

Classifier-reward coefficient sensitivity:

s01.02.0 (default)3.05.0
ASR (%)50.563.064.566.564.0

Engagement-reward coefficient sensitivity:

e00.150.3 (default)0.60.8
ASR (%)56.561.064.566.066.5

Cross-evaluator reward experiments

We ran the following experiments on ZeroShot × Llama-3-8B hard set:

Reward judge during adaptationRed-TTT ASR
Original classifier reward33.3%
Without classifier reward24.3%
ShieldGemma-9B23.5%
Llama-Guard-3-8B25.5%

When using ShieldGemma-9B or Llama-Guard-3-8B, the alternative judge replaces the original classifier in both reward computation and candidate selection. The HarmBench classifier is used only for final evaluation. Switching to independent judges, Red-TTT continues to generate successful attacks. These experiments demonstrate that its effectiveness is not solely attributable to optimizing the evaluation classifier.

Additional benchmark

We additionally evaluated Red-TTT on AdvBench [1] , using the same attacker and 120-query budget. On ZeroShot × Llama-3-8B settings,Red-TTT achieves 59.0% ASR, compared with 48.0% for frozen Best-of-N, a 11.0 percentage-point improvement. This further demonstrates the generalization.

W4 / Q3. Robustness against stronger defenses

We extended our experiments to include multiple input-stage guards, an output-stage guard, and a circuit-breaker defense. For each configuration, the defense is active during Red-TTT's adaptation loop, so the attacker interacts with the defended victim throughout training.

DefenseFrozen Best-of-N ASRRed-TTT ASRImprovement
Llama-Guard-3 (input)50.5%63.0%+12.5
Qwen3-Guard-4B (input)45.5%59.0%+13.5
ShieldGemma-9B (input)49.5%62.0%+12.5
Llama-Guard-3 (output)56.5%66.5%+10.0
Llama-3-8B-Instruct-RR (circuit breakers)30.5%36.5%+6.0

We thank the reviewer again, and hope these clarifications address the concerns. We are happy to answer any further questions.

[1] Zou et al, Universal and Transferable Adversarial Attacks on Aligned Language Models, 2023

Dear Reviewer jGsL,

Thank you again for your detailed review. As the discussion period ends soon, we would like to kindly check whether our responses have addressed your concerns. In brief, we added: (1) wall-clock- and FLOPs-matched comparisons against scaled Best-of-N; (2) diversity diagnostics and KL ablations regarding policy collapse; (3) full-benchmark reward-coefficient sensitivity, alternative reward judges, and an independent Gemini-based evaluation and (4) evaluations against input/output guardrails and circuit breakers.

If any concerns remain, we would be glad to address them within the remaining time. We would also be grateful if you would consider re-evaluating the score in light of these results.

Best regards,

Authors

Dear Reviewer jGsL,

Thank you for your valuable and constructive feedback on our submission! Your insightful suggestions have greatly helped us enhance the quality of our work and strengthen our claims. Following your recommendations, we conducted many additional experiments and uncovered further insights.

As the discussion window is closing soon, we are eager to know whether our new clarifications have successfully addressed your concerns and questions. We would greatly appreciate it if you could let us know, and we are happy to answer any further questions you may have during the remaining time !

Thank you again for your efforts and insightful comments, which have helped us improve our work!

Best wishes,

Authors

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

This paper proposes Red-TTT that adapts an LLM attacker using LoRA-based test-time training. The method combines dense rewards, risk-seeking policy updates, and elite replay to improve candidate generation over iterations. Experiments on HarmBench show improved attack success rates (ASR) over frozen baselines under matched attack-query budgets.

Critical strengths
  • Important problem. The paper addresses how to improve automated red teaming under limited query budgets, an important problem for identifying safety vulnerabilities and evaluating the robustness of LLMs.
  • Well-motivated approach. Updating the attacker using feedback collected during an attack is a solid approach to improving subsequent proposals. Also, risk-seeking updates and elite replay reasonably emphasize promising candidates.
Critical weaknesses
  • W1. The budget-matched comparison does not account for the computational cost of test-time adaptation. The main comparison matches the number of attack queries but does not account for the additional computation required to update the attacker. While the frozen baselines only perform inference, Red-TTT additionally requires forward and backward passes to optimize its adapter between attack rounds. Although LoRA reduces the number of trainable parameters, it does not eliminate the cost of backpropagation, and the paper does not quantify this overhead. A stronger comparison should measure the adapter-update overhead and evaluate performance under matched total compute or wall-clock time. Additional reward-evaluation costs, including the victim-model calls used for StrongREJECT scoring, should also be included in this accounting.
  • W2. The contribution and robustness of the combined recipe remain insufficiently disentangled. The proposed framework combines online parameter adaptation, a multi-component dense reward, risk-seeking advantage weighting, and elite replay. Hence, the comparison against feedback-based search does not fully separate parameter adaptation from differences in feedback and candidate reuse. Some of these mechanisms could also benefit PAIR or TAP; for example, PAIR with same dense reward. Also, Table 2 provides useful ablations, but only for ZeroShot × Llama-3-8B. The decline from 66% to 56% when increasing the replay size from two to four, together with the substantial changes under alternative budget splits, raises questions about robustness. Broader sensitivity analyses and repeated runs would help establish the robustness of the method.
  • W3. Missing important experimental details. Several details in Table 1 and Figure 2(a) are missing. PAIR and TAP consistently underperform independent Best-of-N in Table 1, while AutoDAN-Turbo achieves 27% ASR in Figure 2(a), substantially below the approximately 62% reported in prior HarmBench evaluations against Llama-3-8B [1] . Differences in several setups could explain this gap, but the current draft does not provide enough information to assess these possibilities.
  • W4. The hard-set analysis is somewhat unclear. Line 351-354 define the hard set as behaviors on which the frozen Best-of-N control fails within 120 queries. However, Figure 3(b) shows the frozen control achieving nonzero cumulative success on this subset within the same budget. This ambiguity weakens the interpretation of the figure as evidence that adaptation recovers behaviors that frozen sampling cannot solve within the matched budget.

Minor

  • No citation for TTT-Discover in line 58.

[1] Liu et al., AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs, ICLR 2025

Questions for authors

Please address the weaknesses above. Also, one more specific question is:

  • Why N=120N=120 is considered for main experiments? How about the increased budget like N=1000N=1000? Is TTT still effective with such a larger NN?
10 comments
AuthorsOpenReview ↗

Response to W1

We thank the reviewer for the detailed and positive assessment, and for pinpointing the questions so precisely. Several points are addressed with new experiments below.

W1.

We agree that matching the number of attack queries alone does not establish computational efficiency. We therefore measured end-to-end runtime and estimated FLOPs, and compared Red-TTT against frozen Best-of-N at larger sampling budgets.

Measured runtime

We profiled both methods on the same subset of 12 behaviors using one NVIDIA RTX PRO 6000 Blackwell (98 GiB), with the attacker, victim, and 13B model classifier co-resident. The following values are median per-behavior measurements.

ComponentRed-TTTFrozen Best-of-N
Attacker generation28.9 s0.9 s
Victim attack calls40.3 s3.2 s
Reward scoring, including StrongREJECT48.8 s43.4 s
LoRA updates11.8 s0 s
Total runtime132.5 s48.5 s
Runtime ratio2.73x1.00x
Extrapolated to 200 behaviors on this GPU, the corresponding costs are approximately 7.4 versus 2.7 GPU-hours. LoRA updates account for 11.8 s, or approximately 8.9% of Red-TTT's total runtime. The remaining overhead arises largely from sequential generation and reward evaluation. Specifically, frozen Best-of-N generates 120 candidates in a single vLLM batch, whereas Red-TTT uses 15 sequential rounds of eight candidates through a training-compatible HuggingFace backend. The current implementation also performs updates through a per-sample Python loop. These are genuine implementation costs, and we will explicitly report them rather than characterize the methods as equally compute-efficient.

Victim-call accounting

Both methods use 120 victim attack calls. Red-TTT additionally performs 120 StrongREJECT grading calls, giving 240 calls per behavior under this accounting. The frozen control performs additional grading only for candidates passing the classifier gate, corresponding to approximately 1.4–41.7 additional calls per behavior, depending on the configuration. Thus, its total victim-call count is not always exactly 120. The original comparison is therefore matched on attack queries, not total victim calls. We will distinguish these quantities explicitly. The N=240 frozen-control result below is an additional sampling-budget comparison, not an exact match of all victim-side evaluation calls.

FLOPs and compute-matched comparison

Using measured token counts and an analytical FLOPs accounting (approximately 2ND for forward passes and 4ND for LoRA training passes):

QuantityRed-TTTFrozen Best-of-N
FLOPs per behavior4.11 PFLOPs2.15 PFLOPs
With frozen shared-prefix caching4.11 PFLOPs1.88 PFLOPs
Relative compute1.9–2.2x1.0x
These are analytical estimates of required computation rather than hardware-achieved FLOPs. The attention-score term, approximately 2.4% of the matrix-multiplication term in this setting, is omitted. Under the estimated compute-equivalent budget, the frozen control can generate approximately 230–263 candidates, yielding 57.0–57.8% ASR, compared with 66.0% for Red-TTT at N=120 on ZeroShot × Llama-3-8B. Similarly, using measured throughput to estimate an equal-wall-clock budget gives approximately 328 frozen candidates, yielding 59.0% ASR, or a 7.0-point gap relative to Red-TTT.

Scaling the frozen control

We generated 2,048 independent frozen-attacker candidates per behavior across the 200 ZeroShot × Llama-3-8B behaviors (409,600 generations). We then calculated the expected ASR.

Frozen budget NASRDifference from Red-TTT
12049.0%+17
24057.0%+9
48060.0%+6
1,00060.5%+5.5
2,04863.5%+2.5
Estimated equal wall-clock (~328)59.0%+7.0
AuthorsOpenReview ↗

Response to W2

W2.

(a) Separating parameter adaptation from reward-based selection

To address the reviewer's suggestion more directly, we additionally evaluated frozen PAIR equipped with our composite reward as its selection signal, without gradient updates. The comparison uses the same attacker checkpoint, Llama-3-8B victim, 200 standard behaviors, and 120-query budget, with committed-prompt ASR evaluated.

MethodASR
PAIR15.0%
Frozen Best-of-N49.0%
Frozen PAIR + composite reward56.5%
Red-TTT66.0%
We will add this stronger frozen baseline to the main results and revise the contribution claims accordingly.

(b) Ablations on additional victims and scaffolds

We extended the advantage-estimator comparison to ZeroShot × gpt-oss-20b, using the behaviors where the frozen control failed. The attacker, reward, trainable parameters, and attack-query budget were matched.

EstimatorASR
Risk-seeking28%
Mean-centered GRPO24%

(c) Replay size and budget splits.

As shown in Table 2, elite replay improves performance over the no-replay variant, while different budget allocations affect the balance between exploration and iterative adaptation. Notably, the 40×3 and 4×30 configurations achieve 46% and 50% ASR, respectively, demonstrating that Red-TTT remains effective across different allocations of the same query budget. These results suggest that the default configuration provides a favorable balance between candidate diversity and adaptation frequency.

AuthorsOpenReview ↗

Response to W3 and W4

W3.

We investigated the evaluation protocols and actual query utilization and found several details that require clarification.

(a) AutoDAN-Turbo result

The two results are obtained under different evaluation settings. In the previous authors' released Llama-3 configuration, AutoDAN-Turbo uses Llama-3-8B-Instruct as the attacker, summarizer, and target, with a separate google/gemma-1.1-7b-it scorer. Success is determined using that scorer's 1–10 assessment with a threshold of 8.5, without our additional refusal test. The lifelong procedure includes up to 150 attempts per request during warm-up, followed by strategy accumulation and four additional lifelong iterations. By contrast, our Figure 2(a) uses a unified Qwen2.5-7B attacker substrate, a 120-attempt budget, one committed prompt per behavior, and success determined by the HarmBench classifier together with a non-refusal requirement.

Evaluation aspectOriginal AutoDAN-Turbo settingOur Figure 2(a)
AttackerLlama-3-8B-InstructQwen2.5-7B-Instruct
ScorerGemma-based scorerHarmBench classifier
Success ruleScorer threshold 8.5HarmBench success AND non-refusal
BudgetLifelong warm-up and subsequent iterations120 attempts
These differences prevent a direct comparison between the two reported ASRs. We have not isolated the individual contribution of each difference, so we do not claim that any single factor explains the entire gap. We will clarify that Figure 2(a) is a unified-protocol comparison, not a reproduction of each method's originally published result.

(b) PAIR and TAP: nominal versus consumed queries

We found that PAIR and TAP do not necessarily consume their full nominal 120-query budget. Both methods can terminate when their internal judge reaches its stopping threshold. On ZeroShot × Llama-3-8B, their realized victim-query utilization is:

QuantityPAIRTAP
Nominal query limit120120
Median consumed victim queries2012
Behaviors exhausting budget50/200None
The original PAIR and TAP comparisons use the same maximum budget of 120 victim queries, although both methods may terminate earlier. To examine the effect of early stopping, we disabled PAIR's stopping threshold on gpt-4.1-mini, increasing its median query usage from 20 to 116, while its ASR remained at 11%. For TAP, query usage remained limited by its internal pruning mechanism. We will clarify this distinction in the revised paper.

(c) Attacker refusals in PAIR/TAP scaffolds

We additionally found that the ZeroShot attacker sometimes declines to generate an attack. If the scaffold terminates early, this refusal can become the final delivered prompt. Such refusals occur in approximately 26–35% of behaviors in the evaluated ZeroShot arms and cannot produce successful attacks. Restricting the diagnostic evaluation to behaviors for which an actual attack was delivered yields:

ConfigurationOriginal PAIR ASRASR conditional on attack deliveredBest-of-N ASR on all behaviors
ZeroShot × Llama-3-8B15.0%21.7%49.0%
ZeroShot × gpt-oss-20b10.0%13.5%27.5%
The conditional ASRs use a restricted denominator and therefore should not be treated as strictly matched comparisons with the all-behavior Best-of-N results. We will retain the full evaluation denominator in the main results and report these conditional values as diagnostics.

W4.

Figure 3(b) mistakenly used the hard set constructed under an earlier, incorrect version in the repo. We have corrected the analysis using the intended hard set defined in the paper. Under the same 120-query budget, frozen Best-of-N achieves 0% ASR, while Red-TTT achieves 43.7% ASR. We will replace Figure 3(b) and update the corresponding discussion accordingly.

Comments on the paper

1 comment
~Kautzer_Ginton1OpenReview ↗

Questions on reported numbers and the source of the gains

Dear authors,

Thanks for your work and the detailed responses. While browsing through the forum, I noticed a few inconsistent numbers, and I also have a question about the main claim.

1. Some values don’t seem possible with the stated number of behaviors. The authors stated that the corrected Figure 3(b) and the reward-judge experiments both use the 102-behavior hard set defined in the paper (response to TrSU, W4; follow-up response to MGeh), so every ASR on it should be K/102 for some integer K. However:

  • Red-TTT scores 43.7% in the corrected Figure 3(b), but 44/102 = 43.1% and 45/102 = 44.1%.
  • “Without classifier reward” gets 24.3% in the reward-judge table (listed as 34.3% in an earlier version of the table), but 24/102 = 23.5% and 25/102 = 24.5%, while the other rows in the same table (33.3%, 23.5%, 25.5%) all match K/102.

Could the authors clarify how they were computed?

2. Two different hard-set numbers. In the reward-judge table, the reported ASR for default Red-TTT is 33.3% (responses to jGsL and MGeh), but 43.7% in the corrected Figure 3(b). Could the authors clarify which value is correct, or whether the two use different definitions of success?

Also, the numbers in Figure 2(b) look like they were directly computed from Table 1. All 12 cells exactly equal (ASRRed-TTT−ASRBoN)/(100(\mathrm{ASR}_{\text{Red-TTT}} - \mathrm{ASR}_{\text{BoN}}) / (100\\% - \mathrm{ASR}_{\text{BoN}}) from Table 1. This would only happen if Red-TTT solves every behavior BoN solves. Could the authors report, in each configuration, how many behaviors BoN solves but Red-TTT does not?

3. Effect of TTT. The paper’s central claim is that test-time training improves jailbreak success. However, a large part of the gain from 49.0% to 66.0% (ZeroShot × Llama-3-8B) seems to come from reward-guided selection. As reported in the follow-up response to TrSU, frozen PAIR with the composite reward reaches 58.5% without any weight updates. The compute-matched experiment only compares against frozen BoN, a rather weak baseline, and frozen PAIR with the reward was never compared with matched compute. Overall, the evidence that test-time training drives the gain is not fully convincing to me. Could the authors report this comparison?

Thank you for your patience, and I’d appreciate it if the authors could provide any further clarification in their revisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.