Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away

• URL: http://arxiv.org/abs/2602.11096v1
• Date: Wed, 11 Feb 2026 18:09:17 GMT
• Title: Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away
• Authors: Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh, Furong Huang, Dinesh Manocha, Amrit Singh Bedi,
• Abstract summary: We propose a lightweight inference-time defense that treats safety recovery as a satising constraint rather than an objective.<n>In our evaluations across six open-source MLRMs and four jailbreak benchmarks, SafeThink reduces attack success rates by 30-60%.
• Score: 97.11976870616273
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60% (e.g., LlamaV-o1: 63.33% to 5.74% on JailbreakV-28K, R1-Onevision: 69.07% to 5.65% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20% to 65.00%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first 1-3 reasoning steps typically suffices to redirect the full generation toward safe completions.

Related papers

• 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)
• How Does the Thinking Step Influence Model Safety? An Entropy-based Safety Reminder for LRMs [10.526176863220988]
  We find that the emergence of safe-reminding phrases within thinking steps plays a pivotal role in ensuring LRM safety.<n>Motivated by this finding, we propose SafeRemind, a decoding-time defense method that dynamically injects safe-reminding phrases into thinking steps.<n>By leveraging entropy triggers to intervene at decision-locking points, SafeRemind redirects potentially harmful trajectories toward safer outcomes without requiring any parameter updates.
  arXiv  Detail & Related papers  (2026-01-07T07:26:31Z)
• Think-Reflect-Revise: A Policy-Guided Reflective Framework for Safety Alignment in Large Vision Language Models [58.17589701432514]
  Think-Reflect-Revise (TRR) is a training framework designed to enhance the safety alignment of Large Vision Language Models (LVLMs)<n>We first build a Reflective Safety Reasoning (ReSafe) dataset with 5,000 examples that follow a think-reflect-revise process.<n>We then fine-tune the target model using the ReSafe dataset to initialize reflective behavior, and finally reinforce policy-guided reflection through reinforcement learning.
  arXiv  Detail & Related papers  (2025-12-08T03:46:03Z)
• Bag of Tricks for Subverting Reasoning-based Safety Guardrails [62.139297207938036]
  We present a bag of jailbreak methods that subvert the reasoning-based guardrails.<n>Our attacks span white-, gray-, and black-box settings and range from effortless template manipulations to fully automated optimization.
  arXiv  Detail & Related papers  (2025-10-13T16:16:44Z)
• Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Check [32.82170313959032]
  We introduce a novel safety alignment approach called Answer-Then-Check.<n>Our method enables models to directly answer the question in their thought and then critically evaluate its safety.<n>We find that training on a small subset of just 500 examples can achieve comparable performance to using the full dataset.
  arXiv  Detail & Related papers  (2025-09-15T06:47:35Z)
• ARMOR: Aligning Secure and Safe Large Language Models via Meticulous Reasoning [64.32925552574115]
  ARMOR is a large language model that analyzes jailbreak strategies and extracts the core intent.<n> ARMOR achieves state-of-the-art safety performance, with an average harmful rate of 0.002 and an attack success rate of 0.06 against advanced optimization-based jailbreaks.
  arXiv  Detail & Related papers  (2025-07-14T09:05:54Z)
• SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment [12.206811117510448]
  We introduce SAFEPATH, a lightweight alignment method that fine-tunes LRMs to emit a short, 8-token Safety Primer at the start of their reasoning.<n> Empirical results indicate that SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance.
  arXiv  Detail & Related papers  (2025-05-20T17:54:54Z)
• Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment [97.38766396447369]
  Despite training-time safety alignment, Multimodal Large Language Models (MLLMs) remain vulnerable to jailbreak attacks.<n>We propose Immune, an inference-time defense framework that leverages a safe reward model through controlled decoding to defend against jailbreak attacks.
  arXiv  Detail & Related papers  (2024-11-27T19:00:10Z)

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.