acceptodds
Under review as a conference paper at ICLR 2027

RAISED: Self-Distillation for Robustness to Prompt Injection in LLM Agents

27 views

Abstract

Tool-using language-model agents are vulnerable to indirect prompt injection because they must act on untrusted external content. Existing training-time defenses can reduce attack success rates, but often at the cost of general capabilities. We show that training-based defenses induce substantial drift in the model's output distribution, altering its behavior even in benign settings and providing a potential mechanism for utility degradation. We further identify a failure mode of these defenses: On benign tool-use tasks, the model refrains from a step needed to finish an authorized task, particularly when that step is indicated by a tool output. To address these limitations, we introduce RAISED (Robust Attack Invariance through SElf-Distillation), a training framework that combines self-generation and self-distillation. The model first generates its own tool-use scenarios, with an emphasis on cases where task completion requires acting on legitimate guidance from tool outputs. Then, through self-distillation, the student is trained to match the teacher's clean-context behavior on both clean and injected variants of the same trajectory. RAISED substantially reduces the attack success rate of prompt injections in tool responses while, unlike prior training-based defenses, preserving utility on both agentic and general-purpose benchmarks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents safety, indirect prompt injection, tool use, self-distillation, synthetic data generation

Reviews

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

The paper studies defenses against indirect prompt injection. In particular, authors argue that current trained based defenses suffer from two limitations:

  • They suffer from distribution shift failures when the data at inference time differs from the data used in constructing the defense.
  • They ignore any instructions not given by the user, making it impossible for an agent or model to complete a task that requires following instructions/guidance from a tool output.

To address this, they propose RAISED, a framework consisting of two stages: "self-generation" and "self-distillation". In self generation, a frozen copy of the model to be defended generates data for cases where successful task completion requires following guidance from tool outputs. They generate both benign scenarios and scenarios where malicious instructions are injected, and the continuation ignores it. In self distillation, a trainable version of the same model (student) is trained on these continuations with a KL divergence objective.

They apply their method to 3 models and compare the robustness and utility (under attack) against existing defenses across multiple benchmarks (including general knowledge and reasoning ones, to stress test utility preservation). They find their method to reduce ASR to a similar or lower level while better preserving utility.

Critical strengths
  • The paper focuses on an important problem, defenses against indirect prompt injections, which remains interesting and unsolved.
  • The paper is well written and easy to follow.
  • The paper discusses two intuitive shortcomings of existing training based defenses: distribution shift and the inability to follow benign instructions in tool outputs.
  • The inclusion of general knowledge and reasoning benchmarks to evaluate utility preservation is a nice addition on top of utility measurement in the common prompt injection benchmarks.
  • The paper carefully describes its framework, and evaluates this rigorously across models and benchmarks against other defenses.
Critical weaknesses

First, while the limitations of existing training based defenses are intuitive, I do not find the corresponding results compelling.

If I understand things correctly, the results in Section 4.1 are as such by design? SFT or DPO do not directly aim to maintain KL or top 1, so a KL objective is very much expected to better maintain this. While the authors acknowledge that this does not necessarily mean that this causes downstream utility loss, I think it would be valuable to compute exactly that for these three objectives, considering the exact same data? That would make the argument cleaner in my opinion. Authors include similar results in Appendix E and I personally find this to be more compelling. It might be worth incorporating these (or other) clean ablations more directly in the main paper, or at least refer to them earlier.

Similarly, Section 5.2 on the inability of models to follow benign guidance in tool output, authors provide illustrative evidence, but it would be much more compelling to quantify this properly. Could they not create/sample data testing for exactly this and showcase the difference more thoroughly?

Second, compared to Secalign, both the reduction in ASR and improvement in utility (on prompt injection benchmarks or the general utility benchmarks) are quite marginal, and often lie within one standard deviation. While the framework studies a meaningful, different approach to training time defenses, the current results do not compellingly show that their data construction + KL objective leads to a fundamentally better defense than existing ones.

Lastly, when proposing a new defense against prompt injection, it's particularly valuable to also evaluate the robustness against stronger, adaptive attacks such as the ones from Wen et al. (2025) and Nasr et al. (2026), see below.

Nasr, M., Carlini, N., Sitawarin, C., Schulhoff, S. V., Hayes, J., Ilie, M., ... & Tramèr, F. (2026). The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against {LLM} Jailbreaks and Prompt Injections. In 35th USENIX Security Symposium (USENIX Security 26) (pp. 1467-1486).

Wen, Y., Zharmagambetov, A., Evtimov, I., Kokhlikyan, N., Goldstein, T., Chaudhuri, K., & Guo, C. (2025). Rl is a hammer and llms are nails: A simple reinforcement learning recipe for strong prompt injection. arXiv preprint arXiv:2510.04885.

Questions for authors

Mostly see weaknesses.

Minor suggestions:

  • the citation style could be edited to make the paper more readable, citep versus citet in latex.
  • In the introduction, when explaining self generation and distillation, it would be helpful to emphasize that the injections are malicious and the generated data that follows the injections ignores it. This was not directly clear to me.
6 comments

We thank the reviewer for their feedback, and answer each point below.

First, while the limitations of existing training based defenses are intuitive, I do not find the corresponding results compelling.

If I understand things correctly, the results in Section 4.1 are as such by design? SFT or DPO do not directly aim to maintain KL or top 1, so a KL objective is very much expected to better maintain this. While the authors acknowledge that this does not necessarily mean that this causes downstream utility loss, I think it would be valuable to compute exactly that for these three objectives, considering the exact same data? That would make the argument cleaner in my opinion. Authors include similar results in Appendix E and I personally find this to be more compelling. It might be worth incorporating these (or other) clean ablations more directly in the main paper, or at least refer to them earlier.

A KL-based objective is indeed expected to better preserve the base model's predictive distribution. However, Section 4.1 evaluates this preservation in a substantially different setting from training: all three methods are trained on identical injection-augmented tool-use trajectories, while KL and top-1 agreement are measured on held-out, benign non-agentic questions. Our results therefore reveal a nontrivial finding: SFT and DPO induce substantial distributional drift even on benign tasks outside the tool-use setting used for training, while self-distillation keeps the model close to its original distribution. We also observe the same pattern on tool-use data, with an even larger gap between self-distillation and SFT/DPO.

The motivation for limiting this drift is to preserve the base model's existing capabilities. In the ideal case of zero drift across all benign contexts, the defended model would retain the same behavior and downstream performance. The original submission already included a controlled comparison of downstream utility across the three objectives in Appendix E, using the same Glaive training runs as the longer-training drift analysis in Appendix B.1, and we have now expanded this ablation with additional benign tool-use and general-capability results, brought the key findings into the main text (Section 7.4, Table 4), and made the connection to the distributional-drift analysis explicit (Appendix E). Specifically, self-distillation induces substantially less distributional drift than DPO and SFT while also preserving more downstream utility. For example, it reaches 55.5 benign utility on AgentDyn compared with 48.6 for DPO and 17.5 for SFT; 51.2 on BFCL compared with 28.8 and 11.8; and 71.7 on GPQA compared with 68.4 and 63.0. While this does not establish causality, it shows that lower distributional drift is associated with better capability preservation when training data and inputs are held fixed.

Similarly, Section 5.2 on the inability of models to follow benign guidance in tool output, authors provide illustrative evidence, but it would be much more compelling to quantify this properly. Could they not create/sample data testing for exactly this and showcase the difference more thoroughly?

To quantify this failure mode beyond the illustrative examples, we conducted an additional evaluation on 400 independently generated and validated benign tool-use tasks, where completing the user's request requires following legitimate guidance revealed through a tool output (Appendix B.2.2, Table 8).

We measure how often models receive this guidance but fail to follow it. Relative to the base model, SFT increases this rate by 53.3 and 35.1 percentage points on Gemma and Nemotron, respectively, and SecAlign++ by 24.8 and 25.7 points. In contrast, RAISED limits these increases to only 2.0 and 3.4 points. These results quantify the behavioral drift illustrated in Section 4.2 and show that RAISED largely preserves the base model's ability to follow legitimate guidance from tool outputs.

Second, compared to Secalign, both the reduction in ASR and improvement in utility (on prompt injection benchmarks or the general utility benchmarks) are quite marginal, and often lie within one standard deviation. While the framework studies a meaningful, different approach to training time defenses, the current results do not compellingly show that their data construction + KL objective leads to a fundamentally better defense than existing ones.

The improvements over SecAlign++ are not uniform across all benchmarks. However, several benchmarks reveal substantial utility degradation with SecAlign++ that RAISED largely avoids, as well as significant improvements in robustness.

On Gemma, the clearest distinction is how much vanilla utility each defense retains while achieving similarly low AgentDojo ASR (about 1% for both). Relative to vanilla, SecAlign++ loses 26.1% of AgentDyn benign utility (61.7% → 45.6 ± 0.8%), whereas RAISED loses 2.8% (61.7% → 60.0 ± 2.4%). On AgentDojo, the corresponding losses are 8.5% (84.5% → 77.3 ± 2.2%) and 1.2% (84.5% → 83.5 ± 0.8%). The pattern extends beyond agentic tasks: relative to vanilla, SecAlign++ loses 4.4% on GPQA and 4.2% on MBPP, compared with a 1.2% gain and a 1.0% loss for RAISED, respectively.

On Nemotron, RAISED reduces ASR further while staying close to vanilla utility. Relative to vanilla, SecAlign++ loses 11.2% of AgentDyn benign utility (25.0% → 22.2 ± 0.8%), whereas RAISED preserves it (25.0% → 25.0 ± 0.0%). On AgentDojo, vanilla ASR is 11.7%; SecAlign++ lowers it to 4.0 ± 0.7%, and RAISED to 0.5 ± 0.1%. On AgentDyn, the corresponding figures are 17.6%, 7.8 ± 0.3%, and 1.9 ± 0.6%. Extending SecAlign++ to 625 steps lowers ASR to 0.7% and 3.0%, but moves general capabilities farther from vanilla: MBPP falls by 80.2% relative to vanilla (67.8% → 13.4 ± 2.7%), Minerva Math by 13.6% (76.5% → 66.1 ± 0.9%), and MMLU by 13.5% (49.5% → 42.8 ± 0.3%). RAISED, which was trained for 625 steps, remains close to vanilla on these benchmarks, scoring 66.2 ± 0.6%, 76.7 ± 0.1%, and 49.6 ± 0.1%, respectively.

Lastly, when proposing a new defense against prompt injection, it's particularly valuable to also evaluate the robustness against stronger, adaptive attacks such as the ones from Wen et al. (2025) and Nasr et al. (2026), see below.

Thank you for pointing out these relevant references. We agree that evaluating a new defense against adaptive attacks is important for assessing its robustness. As discussed in Appendix F, we initially investigated the optimization-based PISmith framework (Yin et al., 2026). However, we observed severe reward sparsity across both undefended and defended models. While PISmith eventually succeeded against the undefended models, it did not achieve successful attacks against the defended models. Therefore, we extended our experiments on Gemma-4-12B-it using the stronger search-based adaptive attack proposed by Nasr et al. (2026) on AgentDyn, with a budget of 30 evaluations per case. In this setting, the LLM attacker iteratively refines injections using the agent's execution traces, tool outputs and LLM critic feedback, leading to higher attack success rates even against defended models. Importantly, this stronger attack assumes a less realistic threat model, as the adversary can repeatedly attempt the exact same task with feedback on each attempt. Nevertheless, it provides a valuable stress test of defense robustness beyond the static attacks considered in our main evaluation.

The results reported in Table 19 show that RAISED is more robust than SecAlign++ against this adaptive attack, achieving an ASR of 56.0% compared to 77.4% for SecAlign++. Moreover, the attack budget curve (figure 8) shows that RAISED always outperforms SecAlign++ at each iteration k. For example, 36.9% vs. 40.5% at k=5, 44.0% vs. 54.8% at k=10, and 51.2% vs. 70.2% at k=20. Thus, breaking RAISED is more difficult and expensive than breaking SecAlign++ using this adaptive attack.

Table 19: ASR (%) across evaluation budgets kk on AgentDyn (84 cases)

Defensek=1k=1k=5k=5k=10k=10k=20k=20k=30k=30
Vanilla45.2%95.2%97.6%98.8%98.8%
SecAlign++1.2%40.5%54.8%70.2%77.4%
RAISED0.0%36.9%44.0%51.2%56.0%
Δ\Delta (RAISED vs. SecAlign++)-1.2-3.6-10.8-19.0-21.4

the citation style could be edited to make the paper more readable, citep versus citet in latex.

In the introduction, when explaining self generation and distillation, it would be helpful to emphasize that the injections are malicious and the generated data that follows the injections ignores it. This was not directly clear to me.

Thank you for these suggestions. We revised the citation formatting in the updated version of the paper. We also clarified the introduction: the model first generates a clean task-solving trajectory, and malicious instructions are then inserted into its tool outputs. The training targets remain those obtained from the clean trajectory, teaching the student to continue the legitimate task despite the injected instructions.

Reviewer_fM18OpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RAISED, a self-distillation framework for defending LLM agents against indirect prompt injection attacks. The authors identify two limitations of existing training-based defenses: changes in the model's output distribution and reduced ability to complete legitimate tool-use tasks. To address these issues, RAISED first generates tool-use scenarios and attack examples using the base model. It then trains a student model to match the teacher's clean-context predictions on both clean and injected inputs. Experiments show that RAISED improves robustness while largely preserving the model's original capabilities.

Critical strengths
  1. The paper addresses the trade-off between prompt injection robustness and model utility, which is an important challenge for LLM agents.

  2. Using self-distillation to preserve the model's original behavior under prompt injection is a reasonable and interesting approach.

  3. The authors evaluate multiple models and benchmarks. RAISED achieves low attack success rates while largely preserving benign task performance. The ablation studies also provide useful evidence for the effectiveness of the proposed components.

Critical weaknesses
  1. The main idea of matching model predictions across clean and perturbed inputs is closely related to existing consistency training and self-distillation methods. Although RAISED applies this idea to multi-turn tool-use scenarios, the paper does not sufficiently explain what fundamentally distinguishes its learning mechanism from existing approaches.

  2. The paper shows that self-distillation reduces output-distribution drift, but does not fully establish why this leads to better preservation of downstream capabilities. Similarly, the explanation of why existing defenses fail to follow legitimate tool guidance relies mainly on behavioral examples. A deeper analysis would strengthen the motivation and conclusions.

  3. RAISED uses the base model's clean-context predictions as training targets. However, these predictions may contain incorrect or unsafe behaviors. The proposed method mainly preserves the original model's behavior rather than improving its correctness. It is unclear how well the method works when the teacher generates imperfect trajectories.

Questions for authors
  1. What is the key difference between RAISED and existing consistency training methods? Could the authors clarify the main methodological novelty beyond applying self-distillation to tool-use scenarios?

  2. How does the quality of the self-generated training data affect the final defense performance? What happens when the base model generates incorrect or incomplete task trajectories?

2 comments

We thank the reviewer for their feedback, and answer each point below.

The main idea of matching model predictions across clean and perturbed inputs is closely related to existing consistency training and self-distillation methods. Although RAISED applies this idea to multi-turn tool-use scenarios, the paper does not sufficiently explain what fundamentally distinguishes its learning mechanism from existing approaches.

What is the key difference between RAISED and existing consistency training methods? Could the authors clarify the main methodological novelty beyond applying self-distillation to tool-use scenarios?

We do not claim to introduce a novel consistency training objective. Rather, to the best of our knowledge, RAISED introduces two novel methodological contributions to prompt-injection defence in multi-turn tool-use scenarios. First, we adapt self-distillation-based consistency training to agentic trajectories: the student learns to match the teacher's clean-context predictions despite receiving a context containing injected instructions, encouraging it to distinguish malicious instructions from legitimate guidance within tool responses. Second, we introduce a targeted self-generation and verification pipeline designed to support this distinction. The data is self-generated, executed, and verified, and deliberately includes scenarios where tool outputs provide legitimate guidance that must be followed to complete the task, alongside injections that must be ignored. The ablations in Table 4 empirically demonstrate that both contributions improve the results. We did clarify these contributions more explicitly in the revised manuscript.

The paper shows that self-distillation reduces output-distribution drift, but does not fully establish why this leads to better preservation of downstream capabilities.

Reducing output-distribution drift in benign settings keeps the defended model closer to the base model's original behavior, thereby helping preserve its existing capabilities. In the ideal case of zero drift across all benign contexts, the defended model would behave identically to the base model and thus achieve the same downstream performance. Empirically, the controlled ablation in Section 7.4 (Table 4) shows that lower distributional drift is associated with better downstream utility when training data and inputs are held fixed.

Similarly, the explanation of why existing defenses fail to follow legitimate tool guidance relies mainly on behavioral examples. A deeper analysis would strengthen the motivation and conclusions.

Existing training objectives push the model away from following instructions found in tool outputs and give almost no signal for acting on legitimate guidance from the same channel, so not acting on tool-output guidance is a shortcut that satisfies the objective. To quantify this behavior, we conducted an additional evaluation on 400 independently generated and validated benign tool-use tasks requiring models to follow legitimate guidance from tool outputs. Relative to the base models, SFT increases failure to follow guidance by 53.3pp and 35.1pp on Gemma and Nemotron, respectively, and SecAlign++ by 24.8pp and 25.7pp. In contrast, RAISED limits these increases to only 2.0pp and 3.4pp (Appendix B.2.2, Table 8). These results provide quantitative evidence that existing training-based defenses substantially reinforce this failure mode, while RAISED largely preserves the base model's ability to act on legitimate tool-output guidance.

RAISED uses the base model's clean-context predictions as training targets. However, these predictions may contain incorrect or unsafe behaviors. The proposed method mainly preserves the original model's behavior rather than improving its correctness. It is unclear how well the method works when the teacher generates imperfect trajectories. How does the quality of the self-generated training data affect the final defense performance? What happens when the base model generates incorrect or incomplete task trajectories?

Our goal is to improve robustness to prompt injection, rather than improve the base model's tool-use performance. Thus, our objective is to keep the model invariant to prompt injections. Our KL objective matches the teacher's soft predictive distribution instead of treating individual generated tokens as ground-truth targets, preserving the teacher's probability distribution. So even if a sampled continuation is imperfect, the student is not trained to imitate that specific sequence of tokens, but instead learns to preserve the teacher's behavior through its predictive distribution. Moreover, we only retain trajectories that pass execution checks, satisfy their model-generated success criteria, and pass structure-specific validations (Section 5.1 and Appendix C.5.1), providing validated reference behaviors that the student learns to preserve under injection.

Dear Reviewer fM18,

As the discussion period is approaching its end, we wanted to kindly follow up and ask whether you have had a chance to review our responses. We would appreciate hearing whether your concerns have been addressed and would be happy to provide any further clarifications or answer additional questions.

Thank you very much!

Comments on the paper

3 comments

We thank the reviewer for their feedback, and answer each point below.

The method inherits clean-model errors. Matching the original model cannot correct its wrong tool choices, poor plans, or inappropriate behavior.

RAISED does not aim to correct wrong tool choices, poor plans, or inappropriate behavior already present in the base model. Our goal is to improve robustness to prompt injection, rather than correcting errors or inappropriate behavior that occur even in benign settings.

Synthetic-task filtering can favor easier cases.

We agree that filtering can favor tasks that the base model finds easier. However, we retain trajectories that satisfy their model-generated task-success checks and pass structure-specific validation. This is consistent with our objective: improving robustness to prompt injection by preserving verified clean-context behaviors, rather than teaching the model to solve tasks it cannot already complete.

The prediction-drift study uses a selected benign set. It retains questions the base model answers correctly and excludes likely refusal cases.

The prediction-drift evaluation set was deliberately constructed to focus on benign questions that the base model answers correctly. This allows us to measure whether defensive training alters behavior on tasks the model can already solve, while excluding likely refusal cases avoids confounding the analysis with changes in refusal behavior. As explained in Appendix B.1, the goal is a controlled analysis of distributional drift, rather than measuring drift across all possible inputs.

Some capability losses remain visible.

Some capability losses do remain visible. However, RAISED substantially reduces capability degradation compared with SecAlign++. Aggregating the reported benign performance results across Gemma, Nemotron, and Qwen, RAISED achieves an average baseline-normalized score of 99.23%, compared with 97.67% for SecAlign++. Each score is normalized to its corresponding base model and capped at 100%, so that improvements on some benchmarks cannot offset losses on others. Importantly, RAISED avoids the large capability drops observed with SecAlign++. For instance, on Gemma, SecAlign++ reduces benign utility by 16.1 points on AgentDyn and 7.2 points on AgentDojo, compared with only 1.7 and 1.0 points for RAISED. Similarly, SecAlign++ loses 3.2 points on GPQA and 3.5 points on MBPP, whereas RAISED gains 0.9 points and loses only 0.8 points, respectively. More broadly, RAISED's performance changes are mixed rather than a constant degradation. Without capping, allowing gains to offset losses, its average normalized score reaches 100.04% of baseline, compared with 98.78% for SecAlign++, indicating that its aggregate performance remains close to the original models.

The training attacker has restricted information. It sees a tool output and attack objective but not the user request or full trajectory. Real attackers may know those details and craft more convincing task-specific injections. How would the defense respond to an attacker that repeatedly adapts an injection to the final trained model?

We intentionally restrict the training attacker's information to reflect a realistic setting in which an attacker controls external tool content without access to the user's request or full execution history. This choice is consistent with task-independent attack construction in prior work, including SecAlign++, InjecAgent, and AgentDyn, as well as generic attacks in AgentDojo.

As discussed in Appendix F, we initially investigated the optimization-based PISmith framework (Yin et al., 2026). However, we observed severe reward sparsity across both undefended and defended models. While PISmith eventually succeeded against the undefended models, it did not achieve successful attacks against the defended models. Therefore, we extended our experiments on Gemma-4-12B-it using the stronger search-based adaptive attack proposed by Nasr et al. (2026) on AgentDyn, with a budget of 30 evaluations per case. In this setting, the LLM attacker iteratively refines injections using the agent's execution traces, tool outputs and LLM critic feedback, leading to higher attack success rates even against defended models. Importantly, this stronger attack assumes a less realistic threat model, as the adversary can repeatedly attempt the exact same task with feedback on each attempt. Nevertheless, it provides a valuable stress test of defense robustness beyond the static attacks considered in our main evaluation.

The results reported in Table 19 show that RAISED is more robust than SecAlign++ against this adaptive attack, achieving an ASR of 56.0% compared to 77.4% for SecAlign++. Moreover, the attack budget curve (figure 8) shows that RAISED always outperforms SecAlign++ at each iteration k. For example, 36.9% vs. 40.5% at k=5, 44.0% vs. 54.8% at k=10, and 51.2% vs. 70.2% at k=20. Thus, breaking RAISED is more difficult and expensive than breaking SecAlign++ using this adaptive attack.

Table 19: ASR (%) across evaluation budgets kk on AgentDyn (84 cases)

Defensek=1k=1k=5k=5k=10k=10k=20k=20k=30k=30
Vanilla45.2%95.2%97.6%98.8%98.8%
SecAlign++1.2%40.5%54.8%70.2%77.4%
RAISED0.0%36.9%44.0%51.2%56.0%
Δ\Delta (RAISED vs. SecAlign++)-1.2-3.6-10.8-19.0-21.4

How can the method preserve legitimate tool guidance when malicious text is tightly mixed with information that the user actually needs?

Our data generation framework allows the attacker to tightly mix malicious instructions with legitimate content. The attacker has access to the clean tool output and can embed malicious instructions within existing fields, alongside information or guidance already present in that output (Appendix C.5.1). Self-distillation trains the student to match the teacher's clean-context behavior, including following legitimate guidance, despite the injected instructions (Section 5.2). Together, these two components encourage the model to distinguish legitimate guidance from malicious instructions even when they are tightly mixed within the same tool output, rather than disregarding instructions from tool outputs altogether. However, our evaluation datasets do not isolate the effect of such tightly mixed content, and we leave its quantitative assessment to future work.

Would independently written scenarios and attacks reveal weaknesses hidden by using the same model for generation, solving, and teaching?

Thank you for raising this point. Our evaluation already tests this separation: InjecAgent, AgentDojo, and AgentDyn provide external tasks and attacks that are not generated by RAISED’s self-generation pipeline. The reported robustness gains therefore extend beyond scenarios and attacks produced by the model itself. We also test capability preservation on independent tool-use benchmarks, including ToolTalk, BFCL, WorkBench, Multi-Mission Tool Bench, and AsyncTool, as well as general knowledge and reasoning benchmarks. These results provide evidence that RAISED generalizes beyond its self-generated training data while preserving useful capabilities.

Dear Reviewer UcPr,

With the discussion period ending in two days, we wanted to check if you have had a chance to look over our responses. Please let us know if we have adequately addressed your concerns or if there are any remaining points you would like us to clarify.

Thank you very much!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.