acceptodds
Under review as a conference paper at ICLR 2027

Intent-Sensitive Language Models for Robust Safety

15 views

Abstract

Despite extensive safety-tuning, large language models (LLMs) remain vulnerable to jailbreak attacks via adversarially crafted instructions, reflecting a persistent trade-off between safety and task performance. In this work, we propose Intent-FT, a simple and lightweight fine-tuning approach that explicitly trains LLMs to infer the underlying intent of an instruction before responding. By fine-tuning on a targeted set of adversarial instructions, Intent-FT enables LLMs to generalize intent deduction to unseen attacks, thereby substantially improving their robustness. We comprehensively evaluate both parametric and non-parametric attacks across open-source and proprietary models, considering harmfulness from attacks, utility, over-refusal, and impact against white-box threats. Empirically, Intent-FT consistently mitigates all evaluated attack categories, with no single attack exceeding a 50% success rate—whereas existing defenses remain only partially effective. Importantly, our method preserves the model's general capabilities and reduces excessive refusals on benign instructions containing superficially harmful keywords. Furthermore, models trained with Intent-FT accurately identify hidden harmful intent in adversarial attacks, and these learned intentions can be effectively transferred to enhance vanilla model defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Safety, Jailbreak, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zyJqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a lightweight supervised fine-tuning approach that trains LLMs to explicitly infer the underlying intent of an instruction before generating a response. The authors construct intention-aware training data using adversarial instructions paired with their corresponding harmful instructions, and fine-tune the model to generate an intent analysis followed by a safe or benign response. Experiments across Llama, Qwen, and GPT-4.1-mini show improved robustness against several jailbreak attacks while largely preserving utility and reducing over-refusal.

Critical strengths
  1. The motivation and method are clear, and the visualization is helpful.

  2. This paper is well-written, making it easy to follow.

Critical weaknesses
  1. Limited methodological novelty and insufficient comparison with closely related work. The central idea of identifying the underlying intent before responding has already been explored by several jailbreak defences, including Intention Analysis and SENTINEL [1]. Moreover, recent training-based safety-reasoning approaches [2-5] similarly train models to perform safety/intent reasoning before response generation. The main distinction of INTENT-FT, therefore, appears to be a relatively direct SFT-based instantiation of intention deduction. Stronger differentiation and direct comparisons with these closely related methods are needed to substantiate the methodological novelty.

  2. The supervision may make it unclear whether the model learns general intent understanding. The harmful training labels are generated by a teacher that is explicitly given both the original harmful instruction and its adversarial version. Thus, the model is trained on relatively direct mappings from adversarial prompts to known harmful intents, rather than independently learning to infer intent. The paper would benefit from stronger evaluation on unseen attack constructions and more challenging intent obfuscation.

  3. The empirical evaluation could be strengthened in several concrete aspects. First, the open-source evaluation is limited to 7–8B models; experiments on larger and more recent models (e.g., Qwen 3.5 27B, Llama-3.3-70B and MOE gpt-oss-20b) would help establish scalability. Second, evaluation is conducted on only 100 harmful instructions from JailbreakBench; additional benchmarks such as HarmBench and WildJailbreak would provide substantially broader coverage of harmful behaviors and naturally occurring jailbreaks. Third, several strong and closely related training-based defenses are missing, particularly Reasoning-to-Defend and Safety Chain-of-Thought, which are important baselines given that they similarly introduce safety reasoning into model training.

  4. The claimed cross-model generalisation is only partially supported. The paper shows that generated intentions can transfer to other vanilla models, but the transfer is notably weaker for Qwen. This raises questions about whether the learned intention representation is genuinely model-independent or is partly tied to the model that generated the reasoning.

  5. The inference procedure introduces an additional reasoning step, and its cost is not fully characterised. INTENT-FT requires the model to generate an explicit intention analysis before producing the final response. Although the paper argues that this improves robustness, the additional tokens and latency, as well as the effect of intention-generation errors on downstream responses, deserve more systematic analysis.

[1] Intention Matching Stops Jailbreaks, https://openreview.net/forum?id=vKHCM9O25U
[2] Reasoning-to-defend: Safety-aware reasoning can defend large language models from jailbreaking
[3] Safechain: Safety of language models with long chain-of-thought reasoning capabilities
[4] SafeBehavior: Simulating human-like multistage reasoning to mitigate jailbreak attacks in large language models
[5] CoT defender: Preemptive chain-of-thought occupation for jailbreak attack mitigation

Questions for authors

Refer to weaknesses.

Reviewer_TK6rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Intent-FT, an SFT defense that trains LLMs to produce an explicit intent deduction (<intent>…</intent>) before responding. Training uses ~100 WildJailbreak adversarial prompts annotated with teacher-generated intent traces and refusals, mixed with benign prompts. On Llama-3.1-8B, Qwen-2.5-7B, and GPT-4.1-mini, the authors report the following:

  • lower ASR than prompt- and fine-tuning-based baselines across PAIR, DeepInception, Adaptive Attack, prefilling, and harmful fine-tuning
  • preserved utility and reduced over-refusal on XSTest
  • in the white-box setting, robustness to refusal-direction ablation, with stronger steering degrading WMDP accuracy (a “jailbreak tax”)
Critical strengths
  1. Consistency across attacks and models. Baselines each fail on some attack family. Intent-FT is the only method competitive across all five, using little data and with lower over-refusal than vanilla on Llama and GPT-4.1.
  2. The jailbreak-tax framing. Treating capability cost as a defense objective for open-weight models is a useful perspective, as is the observation that the refusal direction weakens once the early tokens no longer encode the refusal decision.
Critical weaknesses
  1. Non-adaptive evaluation.
  • The AA objective was changed specifically for Intent-FT from log-probability to refusal-phrase matching, which weakens the attack (App. C.1).
  • PAIR runs at only 3 iterations × 15 streams, and Intent-FT’s points in Fig. 9a sit at the iteration cap. Its PAIR ASR is therefore budget-limited, which undercuts Contribution 1’s claim of robustness to attacks with “substantial query budgets.”
  • No attack targets the mechanism itself: prefilling a fake benign intent, omitting the <intent> tag, or running a defense-aware PAIR.
  • In Harmful-FT, only the defender’s safety examples carry <intent>, so the tag may act as a backdoor trigger, as in BD-A.
  • The white-box refusal direction is extracted before </intent>, although App. E.3 says the decision is made later.
  1. The gain is not attributed to intent deduction.
  • Safety-FT trains on vanilla prompts, while Intent-FT trains on adversarial ones. The needed control is the same adversarial data without intent traces.
  • The Intent-FT system prompt (Tab. 5) instructs the model to adhere to safety and ethical guidelines. This is effectively a Self-Reminder baked into training, and it contradicts §4.3, which credits the low over-refusal to “not explicitly prompting the model to follow safety standards.”
  • The training tactics (WildJailbreak) closely resemble PAIR and DI.
  • There is no comparison to closely related reasoning-based safety methods (Reasoning-to-Defend, which is cited, and deliberative-alignment-style training) or to circuit breakers and LAT.
Questions for authors
  1. What is the AA result when you keep its original log-probability objective, targeting compliance after </intent>?
  2. How robust is Intent-FT to defense-aware attacks: fake-intent prefilling, omitting the <intent> tag, and Harmful-FT data that includes <intent> blocks?
  3. How does SFT on the same adversarial prompts and refusals compare when the intent traces are removed, and separately, when the safety system prompt is removed?
  4. Does refusal-direction ablation still fail if the direction is extracted after </intent>?
Reviewer_hHDuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Intent-FT, a lightweight fine-tuning approach that trains large language models (LLMs) to infer the underlying intent of an instruction before generating a response. The central motivation is that conventional safety alignment rely too heavily on surface-level refusal patterns, making models vulnerable to jailbreaks that disguise harmful requests through role-playing, contextual framing, or optimized adversarial prompts. The proposed method constructs an intention dataset using 100 samples from Wild Jailbreak, with a teacher model generating intention deductions and safe continuations. The training data is augmented with benign instructions to encourage intent analysis without excessive refusal. Extensive evaluations suggest that the proposed method improves robustness against the evaluated attacks while largely preserving general capabilities and reducing unnecessary refusals on benign instructions.

The paper addresses an important problem, and the reported results are promising. However, the central claim that explicitly generating an intention deduction produces robust, generalizable safety improvements is not yet convincingly established. The evaluation relies on a small training set and a limited set of target models and attack methods, while key comparisons do not fully isolate the contribution of intention deduction from other changes introduced by fine-tuning. Several experimental choices also make the robustness claims difficult to interpret, including the exclusion of attacks based on their baseline success rates, reliance on an automatic judge, and the use of a single decoding configuration. The evidence supports Intent-FT as a potentially useful fine-tuning heuristic, but not yet as a broadly robust defense.

Critical strengths
  1. Improving jailbreak robustness without unnecessarily blocking benign requests is an important objective for deployed language models.

  2. Intent-FT can be implemented as supervised fine-tuning and does not require a separate inference-time classifier or a complex defense pipeline.

  3. The paper considers judged harmfulness, attack success, general capability, and over-refusal. This is a more informative evaluation strategy than measuring refusal phrases alone.

Critical weaknesses
  1. The proposed method uses merely 100 adversarial training samples but claims broad generalization across attacks. The representativeness of samples and stability across sample subsets and random seeds are not verified. Fine-tuning on this tiny dataset may let the model memorize training patterns instead of acquiring general malicious intent inference capability.

  2. The evaluation lacks rigorous tests against strong adaptive adversaries. Some attack families are excluded for low success on undefended models, and adaptive attacks use a reduced budget. Thus the results cannot validate robustness against stronger, defense-aware attacks, and white-box results call for tempered robustness claims.

  3. The paper does not sufficiently distinguish the benefit of intention deduction from the effects of adversarial safety examples, special tokens, training duration, and benign-data composition. This is a major attribution problem: the observed improvements may be real without validating the proposed explanation for them.

  4. The core safety metric depends on GPT-4.1 as the automatic judge, using a strict threshold of harm score = 10. This misses partial harmful compliance, and judge bias may skew comparisons. Also, the paper does not establish sufficient agreement with human evaluations or sensitivity to alternative thresholds to show that the main conclusions are robust to the evaluation procedure.

  5. General capability scores and refusal-string matching do not fully measure helpfulness or over-refusal in realistic settings. The paper also varies the training mixture for the proprietary model, making direct comparisons less controlled.

Questions for authors
  1. How sensitive is Intent-FT to the choice of the 100 training examples? Can the authors report results across multiple independently sampled training sets and random seeds, with uncertainty estimates?

  2. Why were several established attack families excluded solely because their baseline attack success rates were below 50%? How does Intent-FT perform against a stronger adaptive attacker that explicitly optimizes prompts to bypass or manipulate the intention-deduction stage, using comparable and larger attack budgets?

  3. Can the authors provide controlled ablations comparing (a) ordinary safety fine-tuning on the same adversarial data, (b) fine-tuning with intent tags but without intention deductions, (c) fine-tuning with intention deductions but without the special tags, and (d) the full method? These comparisons should control for data, training tokens, and compute as far as possible.

  4. How well do GPT-4.1 judge scores agree with human judgments of harmful compliance? Do the conclusions remain consistent under alternative ASR thresholds, different judges, and explicit assessment of partial compliance?

  5. Can the authors evaluate benign requests using a stronger, validated benchmark and report human-verified over-refusal rates? How do they distinguish a genuine refusal from an answer that contains a conventional refusal phrase but still provides useful information?

  6. How do the authors distinguish an increase in the cost of successful misuse from a general loss of model capability under steering interventions? Can they provide evidence that the resulting models are less useful for harmful tasks without materially degrading relevant benign capabilities?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.