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.
CogSteer: Cognition-Inspired Selective Layer Intervention for Efficiently Steering Large Language Models
Department of Informatics, Universität Hamburg（汉堡大学信息学院）；The University of Sydney（悉尼大学）；Alibaba International Digital Commerce（阿里巴巴国际数字商业）；Institute of Psychology, Chinese Academy of Sciences（中国科学院心理研究所）
STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations
Jinesis AI Lab, University of Toronto & Vector Institute（Jinesis AI实验室，多伦多大学及向量研究所）；Max Planck Institute for Intelligent Systems, Tübingen, Germany（智能系统马克斯·普朗克研究所，图宾根，德国）；Thoughtworks（Thoughtworks公司）；Martian；ELLIS Institute, Tübingen, Germany（图宾根ELLIS研究所，德国）；EuroSafeAI
Artificial entrepreneurial cognition: Locating and causally steering an opportunity recognition dial inside large language models (LLMs)
Interdisciplinary Centre for Security, Reliability and Trust (SnT), University of Luxembourg（卢森堡大学跨学科安全、可靠性与信任中心（SnT））；Amsterdam Business School, University of Amsterdam（阿姆斯特丹大学阿姆斯特丹商学院）；Stanford Graduate School of Business, Stanford University（斯坦福大学斯坦福商学院）
Steering Language Generation: Harnessing Contrastive Expert Guidance and Negative Prompting for Coherent and Diverse Synthetic Data Generation
Mathematical Sciences Institute, Australian National University（澳大利亚国立大学数学科学研究所）；Research School of Astronomy & Astrophysics, Australian National University（澳大利亚国立大学天文与天体物理研究学院）；School of Computing, Australian National University（澳大利亚国立大学计算学院）；Department of Astronomy, The Ohio State University（俄亥俄州立大学天文学系）
