Reducing the Safety Tax in LLM Safety Alignment with On-Policy Self-Distillation

• URL: http://arxiv.org/abs/2605.15239v1
• Date: Thu, 14 May 2026 03:40:07 GMT
• Title: Reducing the Safety Tax in LLM Safety Alignment with On-Policy Self-Distillation
• Abstract summary: Safety alignment often improves robustness to harmful queries at the cost of reasoning ability.<n>We study on-policy self-distillation for safety alignment, which we call OPSA.<n> OPSA achieves a stronger safety--reasoning tradeoff than off-policy self-distillation and external-teacher distillation.
• Score: 22.972195366680694
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Safety alignment often improves robustness to harmful queries at the cost of reasoning ability, a tradeoff known as the safety tax. A common cause is distributional mismatch: supervised fine-tuning trains the target model on safety demonstrations produced by humans, external models, or fixed self-generated traces, rather than on trajectories sampled from its own policy. We identify off-policy training mismatch as a second source of this tax and study on-policy self-distillation for safety alignment, which we call OPSA. The model generates its own rollouts and receives dense per-token KL supervision from a frozen teacher copy of itself conditioned on a privileged safety context. Because this teacher must be safer than the sampled student trajectory, we introduce \emph{teacher flip rate}: a criterion that measures how often a privileged context converts unsafe responses into safe ones. We use this signal to search for contexts that activate latent safety reasoning rather than merely elicit safe-looking demonstrations. Across two reasoning-model families and five model scales, OPSA achieves a stronger safety--reasoning tradeoff than off-policy self-distillation and external-teacher distillation under matched data and full-parameter fine-tuning, with the largest gains on smaller models (+8.85 points on R1-Distill-1.5B and +5.49 points on Qwen3-0.6B). The gains persist across training-set sizes and adaptive jailbreak evaluations. Token-level analyses further show that OPSA concentrates updates near early compliance-decision tokens, providing a mechanism for improving safety while preserving general reasoning.

Related papers

• Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models [31.940554924036807]
  Large Reasoning Models (LRMs) remain highly vulnerable to sophisticated jailbreaks and direct harmful queries.<n>We employSupervised Fine-Tuning (SFT) to explicitly induce safe tags to trigger safety analysis and guidance.<n>We also apply Direct Preference Optimization (DPO) to further enhance the correctness and stability of the safety analysis and guidance.
  arXiv  Detail & Related papers  (2026-06-15T14:51:34Z)
• Constitutional On-Policy Safe Distillation [64.02536207601112]
  On-policy self-distillation (OPSD) has emerged as an efficient post-training paradigm by using a teacher conditioned on privileged information to provide dense token-level supervision.<n>We propose Constitutional On-Policy Safe Distillation (COPSD), which first calibrates the teacher through a Cross-SFT cold-start and then performs constitution-conditioned on-policy distillation.
  arXiv  Detail & Related papers  (2026-06-02T03:17:56Z)
• Learning to Stay Safe: Adaptive Regularization Against Safety Degradation during Fine-Tuning [2.9184958249079975]
  Existing defenses offer limited protection or force a trade-off between safety and utility.<n>We introduce a training framework that adapts regularization in response to safety risk.<n>We empirically verify that harmful intent signals are predictable from pre-generation activations.
  arXiv  Detail & Related papers  (2026-02-19T16:59:54Z)
• THINKSAFE: Self-Generated Safety Alignment for Reasoning Models [60.10077024249373]
  We propose ThinkSafe, a framework that restores safety alignment without external teachers.<n>Our key insight is that while compliance suppresses safety mechanisms, models often retain latent knowledge to identify harm.<n> Experiments on DeepSeek-R1-Distill and Qwen3 show ThinkSafe significantly improves safety while preserving reasoning proficiency.
  arXiv  Detail & Related papers  (2026-01-30T16:31:02Z)
• Unified Defense for Large Language Models against Jailbreak and Fine-Tuning Attacks in Education [32.70143887942455]
  Large Language Models (LLMs) are increasingly integrated into educational applications.<n>LLMs are vulnerable to jailbreak and fine-tuning attacks, which can compromise safety alignment and lead to harmful outputs.<n>We propose a three-stage shield framework (TSSF) for educational LLMs that simultaneously mitigates both jailbreak and fine-tuning attacks.
  arXiv  Detail & Related papers  (2025-11-18T12:27:51Z)
• When Models Outthink Their Safety: Mitigating Self-Jailbreak in Large Reasoning Models with Chain-of-Guardrails [74.63933201261595]
  Large Reasoning Models (LRMs) demonstrate remarkable capabilities on complex reasoning tasks.<n>LRMs remain vulnerable to severe safety risks, including harmful content generation and jailbreak attacks.<n>We propose the Chain-of-Guardrail (CoG), a training framework that recomposes or backtracks unsafe reasoning steps.
  arXiv  Detail & Related papers  (2025-10-24T09:32:25Z)
• Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention [53.25106308403173]
  We show that existing methods overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users.<n>We propose Intervened Preference Optimization (IPO), an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals.
  arXiv  Detail & Related papers  (2025-09-29T07:41:09Z)
• Shape it Up! Restoring LLM Safety during Finetuning [65.75757313781104]
  Finetuning large language models (LLMs) enables user-specific customization but introduces critical safety risks.<n>We propose dynamic safety shaping (DSS), a framework that uses fine-grained safety signals to reinforce learning from safe segments of a response while suppressing unsafe content.<n>We present STAR-DSS, guided by STAR scores, that robustly mitigates finetuning risks and delivers substantial safety improvements across diverse threats, datasets, and model families.
  arXiv  Detail & Related papers  (2025-05-22T18:05:16Z)
• SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning [76.56522719330911]
  Large Reasoning Models (LRMs) introduce a new generation paradigm of explicitly reasoning before answering.<n>LRMs pose great safety risks against harmful queries and adversarial attacks.<n>We propose SafeKey to better activate the safety aha moment in the key sentence.
  arXiv  Detail & Related papers  (2025-05-22T03:46:03Z)
• STAIR: Improving Safety Alignment with Introspective Reasoning [44.780098674618614]
  We propose STAIR, a framework that integrates SafeTy Alignment with Itrospective Reasoning.<n>We show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies.<n>With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks.
  arXiv  Detail & Related papers  (2025-02-04T15:02:55Z)
• Vulnerability Mitigation for Safety-Aligned Language Models via Debiasing [12.986006070964772]
  Safety alignment is an essential research topic for real-world AI applications.<n>Our study first identified the difficulty of eliminating such vulnerabilities without sacrificing the model's helpfulness.<n>Our method could enhance the model's helpfulness while maintaining safety, thus improving the trade-off-front.
  arXiv  Detail & Related papers  (2025-02-04T09:31:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.