acceptodds
Under review as a conference paper at ICLR 2027

HiRoute: Hierarchical Routed Prompt Tuning for Safety Alignment of Large Language Models

Abstract

Large language models (LLMs) remain vulnerable to harmful requests and jailbreak attacks. Parameter-efficient safety alignment methods based on prompt tuning typically rely on a single global prompt or externally selected prompt modules. Such static designs struggle to maintain a cross-category safety boundary while generating constructive responses tailored to specific risks and avoiding over-refusal of benign inputs. To address these limitations, we propose HiRoute, an input-adaptive hierarchical prompt-tuning framework that separates category-agnostic safety control from category-specific response guidance. HiRoute first trains a lightweight hierarchical router on representations extracted from a frozen LLM to jointly detect harmful intent and predict multi-label risk scores. It then freezes both the backbone model and the router and uses preference optimization with alternating gradient updates to learn a shared coarse-grained prompt and a set of fine-grained prompt experts as continuous embeddings. At inference time, benign inputs bypass the safety branch, whereas risky inputs are processed using the shared prompt together with a router-weighted mixture of risk-specific prompt experts. Experiments across three instruction-tuned models show that HiRoute achieves high safety rates across multiple safety benchmarks while preserving safe-response helpfulness, reducing over-refusal, and maintaining competitive performance on general-purpose tasks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Safety Alignment, Prompt Tuning, Jailbreak Defense

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_gbU5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

HiRoute uses a classifier to activate a shared safety prompt and mix risk-specific soft prompts over a frozen LLM. Experiments on three backbones report higher average safety and safe-response helpfulness, with lower over-refusal than the compared defenses.

Critical strengths
  • Reports substantial safety improvements over the base models.

  • Evaluates safe-response helpfulness, general utility, and over-refusal alongside safety.

Critical weaknesses
  • The motivation for separate experts is weak. A shared prompt still conditions on the request and can produce tailored safe responses. The paper does not establish an advantage over one prompt trained on the same combined data, with the same gate and training budget. Figure 2’s separate diagnostic does not provide this comparison.

  • The over-refusal comparison is confounded by gating. No gate-matched results are reported for competing defenses, leaving the source of HiRoute’s advantage unclear.

  • All evaluated backbones are from 2023–2024. The benefit on recent and more relevant models remains untested.

  • The GCG evaluation uses transferred attacks. Adaptive attacks against the complete defense remain untested, as the authors acknowledge.

Questions for authors
  • Does HiRoute outperform a single prompt under matched data, gating, and training budget?

  • Does the over-refusal advantage persist when competing defenses use the same gate?

  • Do the gains hold on a recent instruction-tuned model?

  • How does the complete defense perform under adaptive attacks, including attempts to bypass the router?

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

This paper proposes HiRoute, an input-adaptive and parameter-efficient framework for LLM safety alignment. A lightweight router trained on hidden representations from a frozen backbone first predicts whether the user input is safe and then assigns multi-label scores to four risk categories. For inputs requiring safety intervention, a shared, category-agnostic soft prompt is concatenated with a weighted mixture of risk-specific soft-prompt experts; inputs judged confidently safe bypass these additional prompts. The backbone and router remain frozen during prompt optimization, which uses DPO with alternating gradient updates for the shared and fine-grained prompts. Experiments on Mistral-7B-Instruct-v0.3, Vicuna-7B-v1.5, and Zephyr-7B-Beta report improvements in safety rate and safe-response helpfulness over several prompt-based baselines. The paper also evaluates general-task utility, over-refusal, component and training ablations, extension to Vicuna-13B, and robustness to transferred GCG attacks.

Critical strengths
  1. Clear and well-motivated hierarchical safety control. A shared coarse-grained prompt provides category-independent safety guidance, while a multi-label router combines fine-grained experts for input-specific risks. Table 3(b) provides supporting evidence: on Mistral, the weighted fine-only variant achieves 69.8% average safety / 7.4 helpfulness, Top-1 + coarse achieves 82.5% / 6.9, and HiRoute achieves 93.2% / 7.4.
  2. Broad evaluation of safety, helpfulness, and utility, with meaningful ablations. The experiments cover StrongReject, AdvBench, JailbreakBench, safe-response helpfulness, general-task performance, and XSTest over-refusal. The study also examines routing and aggregation, prompt length, training strategy, update ratio, a 13B backbone, and transferred GCG attacks. Relative to the tested safety-aligned baselines, HiRoute achieves lower over-refusal rates while improving safety.
Critical weaknesses
  1. Insufficient discussion and comparison with closely related routing-based safety alignment methods. The paper does not cite or discuss closely related works such as MoGU (Du et al., NeurIPS 2024) and MidPO (Qi et al., Findings of EMNLP 2025), which have already explored input-dependent expert routing to balance safety and usability/helpfulness. HiRoute follows a related direction through a frozen LLM backbone and hierarchical soft-prompt routing with shared safety control and risk-specific experts. However, the experimental comparisons mainly focus on prompt-based defenses (ACD, RPO, and DRO), without including representative routing-based safety alignment methods. This makes it difficult to assess the additional benefit of the proposed design over existing approaches. Given that MoGU and HiRoute both evaluate on AdvBench, a controlled comparison under consistent model and evaluation settings would be particularly valuable. Furthermore, since handling overlapping risk categories is a central motivation, an evaluation distinguishing single-risk from multi-risk requests would help substantiate the benefits of the proposed fine-grained routing mechanism.
  2. Insufficient validation of the main automatic evaluation results. Response safety and safe-response helpfulness are evaluated using GPT-5.4, but Appendix E reports a manual check of only 20 responses without quantitative agreement, annotator information, or a sampling breakdown. Helpfulness is also calculated only on responses classified as safe, potentially yielding different evaluated subsets across methods. HiRoute's average safety-rate gains over the strongest aligned baselines are 1.7, 1.3, and 2.9 percentage points on the three backbones, with no reported confidence intervals or multi-seed variability. The reliability of these relatively small margins remains insufficiently established. 3.Furthermore, the main experiments are limited to Mistral-7B-Instruct-v0.3, Vicuna-7B-v1.5, and Zephyr-7B-Beta, primarily 2023–2024-era checkpoints. Since Zephyr is derived from the Mistral model family, the evaluation also provides limited backbone-family diversity. Although the Vicuna-13B experiment examines scalability, it does not establish generalization to more recent, independent model families. This leaves the applicability of HiRoute to contemporary instruction-tuned LLMs insufficiently validated.
Questions for authors
  1. Prior work and comparative evidence. Please discuss MoGU and MidPO and specify the contribution beyond existing routing-based safety alignment. Can a representative routing-based baseline be evaluated under a consistent backbone and judge? Alternatively, can the authors provide direct evidence for the benefit of risk-category routing, particularly on single-risk versus overlapping-risk requests?
  2. Evaluation reliability. How were the 20 manually checked responses sampled, and what annotation procedure was used? Please report quantitative human–judge agreement for safety classification and discrepancies in helpfulness scores. Can the authors provide additional human validation or uncertainty analysis to support the margins over the strongest baselines, and clarify how safety-conditioned helpfulness affects comparisons?
  3. Risk-label construction. PKU-SafeRLHF associates risk labels with question–answer pairs, whereas the fine-grained router predicts labels from the user input alone. How are the original labels converted into input-level multi-label targets, including cases with multiple responses to the same prompt?
  4. Backbone representativeness and generalization. What motivates the selection of these relatively early instruction-tuned models? Could the authors evaluate HiRoute on a more recent, independent backbone (e.g., Qwen3-8B) to assess whether the reported safety–helpfulness benefits generalize to contemporary models? Alternatively, the scope of the generalization claims should be clarified.
Reviewer_QBUEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the limitations of static prompt-tuning approaches for LLM safety, which can struggle to enforce consistent protections across risk categories while avoiding over-refusal of benign inputs. It proposes HiRoute, an input-adaptive hierarchical framework that detects harmful intent and risk levels, then combines a shared safety prompt with router-weighted, risk-specific prompt experts. Experiments on three instruction-tuned models indicate that HiRoute achieves strong safety while preserving helpfulness, reducing over-refusal, and maintaining competitive performance on general-purpose tasks.

Critical strengths

(1) The paper presents an interesting approach to improving LLM safety while reducing over-refusal.

(2) The authors conduct extensive experiments and report promising results, which support the effectiveness of their proposed method.

Critical weaknesses

(1) My main concern is the paper’s novelty. The work appears incremental, combining existing ideas such as risk routing and prompt learning without providing substantial new insight into the safety alignment of large language models.

(2) The paper does not clearly identify its contributions. In particular, it is difficult to determine how the proposed method differs from existing work, especially given the substantial body of research on LLM safety alignment.

(3) The paper does not report computational costs, making it difficult to assess the practical efficiency of the proposed method.

(4) The paper would benefit from a thorough revision to improve its clarity and readability. Beyond correcting typographical errors, the authors should revise passages that are difficult to understand and explain the method and its contributions more clearly.

Questions for authors

(1) What is the paper’s main contribution, and how does the proposed work relate to and differ from existing approaches?

(2) Do the authors compare the computational costs of the proposed method with those of existing approaches?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.