How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study.

Qi Zhu 0011 Huawei Noah's Ark Lab, Guangdong, China Qi Zhu Qi Zhu 0001 Nanjing University of Aeronautics and Astronautics, Department of Computer Science and Technology, China Harbin Institute of Technology, Shenzhen Graduate School, Bio-Computing Research Center, China https://orcid.org/0000-0001-5497-9478 https://orcid.org/0000-0001-7740-292X https://ieeexplore.ieee.org/author/37085727803 Qi Zhu 0002 Northwestern University, Department of Electrical Engineering and Computer Science, USA University of California Riverside, Department of Electrical and Computer Engineering, CA, USA Intel Corporation, Strategic CAD Laboratories, USA University of California Berkeley, Department of Electrical Engineering and Computer Science, CA, USA Tsinghua University, Computer Science and Technology Department, Beijing, China https://orcid.org/0000-0002-7700-4099 https://www.wikidata.org/entity/Q60012633 Qi Zhu 0003 Nanjing University of Posts and Telecommunications, School of Telecommunication and Information Engineering, Key Wireless Laboratory of Jiangsu Province, China https://orcid.org/0000-0002-8190-8912 Qi Zhu 0004 National University of Defense Technology, College of Mechatronics and Automation, Changsha, China Qi Zhu 0005 Central South University, School of Information Science and Engineering, Changsha, China https://orcid.org/0000-0002-4307-3694 Qi Zhu 0006 University of Science and Technology of China, Department of Electronic Engineering and Information Science, Hefei, China Nanjing University of Aeronautics and Astronautics, China https://orcid.org/0000-0002-9786-2621 Qi Zhu 0007 Beijing National Research Center for Information Science and Technology, Institute for Artificial Intelligence, China Tsinghua University, China Qi Zhu 0008 University of Illinois at Urbana-Champaign, Department of Computer Science, IL, USA https://openreview.net/profile?id=~Qi_Zhu7 https://orcid.org/0000-0003-0129-8542 https://www.wikidata.org/entity/Q130841415 Qi Zhu 0009 National Oceanic and Atmospheric Administration, NOAA, Center for Satellite Applications and Research, National Environmental Satellite, Data, and Information Service, NESDIS, MD, USA Qi Zhu 0010 University of Science and Technology of China, Hefei, China https://openreview.net/profile?id=~Qi_Zhu9 https://orcid.org/0000-0002-1545-1854 Zhexin Zhang Xian Qi Loye Victor Shea-Jay Huang Junxiao Yang Qi Zhu 0011 Shiyao Cui Fei Mi Lifeng Shang Yingkang Wang Hongning Wang Minlie Huang How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study. 20438-20457 2026 ACL (1) https://doi.org/10.18653/v1/2026.acl-long.936 https://aclanthology.org/2026.acl-long.936/ conf/acl/2026-1 db/conf/acl/acl2026-1.html#ZhangLHYZCMSWWH26 Bowen Ding Yuhan Chen Jiayang Lyu Jiyao Yuan Qi Zhu 0011 Shuangshuang Tian Dantong Zhu Futing Wang Heyuan Deng Fei Mi Lifeng Shang Tao Lin Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning. 33081-33106 2026 ACL (1) https://doi.org/10.18653/v1/2026.acl-long.1528 https://aclanthology.org/2026.acl-long.1528/ conf/acl/2026-1 db/conf/acl/acl2026-1.html#DingCLYZTZWDMSL26

Hanbin Wang Jingwei Song Jinpeng Li Qi Zhu 0011 Fei Mi Ganqu Cui Yasheng Wang Lifeng Shang Teaching Large Reasoning Models Effective Reflection. 2026 January abs/2601.12720 CoRR https://doi.org/10.48550/arXiv.2601.12720 db/journals/corr/corr2601.html#abs-2601-12720 streams/journals/corr

Wenshuo Zhao Qi Zhu 0011 Xingshan Zeng Fei Mi Lifeng Shang Yi R. Fung 0001 Entropy Centroids as Intrinsic Rewards for Test-Time Scaling. 2026 April abs/2604.26173 CoRR https://doi.org/10.48550/arXiv.2604.26173 db/journals/corr/corr2604.html#abs-2604-26173 streams/journals/corr

Xujun Li Kehan Zheng Mingyuan Zhao Yize Geng Jinfeng Zhou Qi Zhu 0011 Fei Mi Lifeng Shang Minlie Huang Hongning Wang You Live More Than Once: Towards Hierarchical Skill Meta-Evolving. 2026 May abs/2605.28390 CoRR https://doi.org/10.48550/arXiv.2605.28390 db/journals/corr/corr2605.html#abs-2605-28390 streams/journals/corr

Xian Qi Loye Qinglin Su Zhexin Zhang Shiyao Cui Qi Zhu 0011 Fei Mi Hongning Wang Minlie Huang RUBAS: Rubric-Based Reinforcement Learning for Agent Safety. 2026 June abs/2606.04051 CoRR https://doi.org/10.48550/arXiv.2606.04051 db/journals/corr/corr2606.html#abs-2606-04051 streams/journals/corr

Zige Wang Qi Zhu 0011 Fei Mi Yasheng Wang Haotian Wang Lifeng Shang Dynamic data selection with normalized gradient-based influence approximation for targeted fine-tuning of LLMs. 114144 2025 327 Knowl. Based Syst. https://doi.org/10.1016/j.knosys.2025.114144 db/journals/kbs/kbs327.html#WangZMWWS25 streams/journals/kbs

Erxin Yu Jing Li 0049 Ming Liao Qi Zhu 0011 Boyang Xue Minghui Xu Baojun Wang Lanqing Hong Fei Mi Lifeng Shang Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning. 8504-8519 2025 ACL (1) https://doi.org/10.18653/v1/2025.acl-long.417 https://aclanthology.org/2025.acl-long.417/ conf/acl/2025-1 db/conf/acl/acl2025-1.html#YuLLZXXWHMS25 Zige Wang Qi Zhu 0011 Fei Mi Minghui Xu Ruochun Jin Wenjing Yang 0002 ClusterUCB: Efficient Gradient-Based Data Selection for Targeted Fine-Tuning of LLMs. EMNLP (Findings) 2025 18867-18880 conf/emnlp/2025f https://doi.org/10.18653/v1/2025.findings-emnlp.1026 https://aclanthology.org/2025.findings-emnlp.1026/ db/conf/emnlp/emnlp2025f.html#WangZMXJY25 streams/conf/emnlp Yen-Ju Lu Ting-Yao Hu Hema Swetha Koppula Hadi Pouransari Jen-Hao Rick Chang Yin Xia Xiang Kong Qi Zhu 0011 Simon Wang Oncel Tuzel Raviteja Vemulapalli Mutual Reinforcement of LLM Dialogue Synthesis and Summarization Capabilities for Few-Shot Dialogue Summarization. 7237-7256 2025 NAACL (Findings) https://doi.org/10.18653/v1/2025.findings-naacl.404 https://aclanthology.org/2025.findings-naacl.404/ conf/naacl/2025f db/conf/naacl/naacl2025f.html#LuHKPCXKZWTV25 streams/conf/naacl

Yen-Ju Lu Ting-Yao Hu Hema Swetha Koppula Hadi Pouransari Jen-Hao Rick Chang Yin Xia Xiang Kong Qi Zhu 0011 Simon Wang Oncel Tuzel Raviteja Vemulapalli Mutual Reinforcement of LLM Dialogue Synthesis and Summarization Capabilities for Few-Shot Dialogue Summarization. 2025 February abs/2502.17328 CoRR https://doi.org/10.48550/arXiv.2502.17328 db/journals/corr/corr2502.html#abs-2502-17328 streams/journals/corr

Boyang Xue Qi Zhu 0011 Hongru Wang 0003 Rui Wang 0092 Sheng Wang Hongling Xu Fei Mi Yasheng Wang Lifeng Shang Qun Liu 0001 Kam-Fai Wong DAST: Difficulty-Aware Self-Training on Large Language Models. 2025 March abs/2503.09029 CoRR https://doi.org/10.48550/arXiv.2503.09029 db/journals/corr/corr2503.html#abs-2503-09029 streams/journals/corr

Zhexin Zhang Xian Qi Loye Victor Shea-Jay Huang Junxiao Yang Qi Zhu 0011 Shiyao Cui Fei Mi Lifeng Shang Yingkang Wang Hongning Wang Minlie Huang How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study. 2025 May abs/2505.15404 CoRR https://doi.org/10.48550/arXiv.2505.15404 db/journals/corr/corr2505.html#abs-2505-15404 streams/journals/corr

Yuzhen Huang 0002 Weihao Zeng 0003 Xingshan Zeng Qi Zhu 0011 Junxian He Pitfalls of Rule- and Model-based Verifiers - A Case Study on Mathematical Reasoning. 2025 May abs/2505.22203 CoRR https://doi.org/10.48550/arXiv.2505.22203 db/journals/corr/corr2505.html#abs-2505-22203 streams/journals/corr

Erxin Yu Jing Li 0049 Ming Liao Qi Zhu 0011 Boyang Xue Minghui Xu Baojun Wang Lanqing Hong Fei Mi Lifeng Shang Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning. 2025 May abs/2505.22591 CoRR https://doi.org/10.48550/arXiv.2505.22591 db/journals/corr/corr2505.html#abs-2505-22591 streams/journals/corr

Hongling Xu Qi Zhu 0011 Heyuan Deng Jinpeng Li Lu Hou Yasheng Wang Lifeng Shang Ruifeng Xu 0001 Fei Mi KDRL: Post-Training Reasoning LLMs via Unified Knowledge Distillation and Reinforcement Learning. 2025 June abs/2506.02208 CoRR https://doi.org/10.48550/arXiv.2506.02208 db/journals/corr/corr2506.html#abs-2506-02208 streams/journals/corr

Zige Wang Qi Zhu 0011 Fei Mi Minghui Xu Ruochun Jin Wenjing Yang 0002 ClusterUCB: Efficient Gradient-Based Data Selection for Targeted Fine-Tuning of LLMs. 2025 June abs/2506.10288 CoRR https://doi.org/10.48550/arXiv.2506.10288 db/journals/corr/corr2506.html#abs-2506-10288 streams/journals/corr

Boyang Xue Qi Zhu 0011 Rui Wang 0092 Sheng Wang Hongru Wang 0003 Fei Mi Yasheng Wang Lifeng Shang Qun Liu 0001 Kam-Fai Wong ReliableMath: Benchmark of Reliable Mathematical Reasoning on Large Language Models. 2025 July abs/2507.03133 CoRR https://doi.org/10.48550/arXiv.2507.03133 db/journals/corr/corr2507.html#abs-2507-03133 streams/journals/corr

Bowen Ding Yuhan Chen Jiayang Lv Jiyao Yuan Qi Zhu 0011 Shuangshuang Tian Dantong Zhu Futing Wang Heyuan Deng Fei Mi Lifeng Shang Tao Lin Rethinking Expert Trajectory Utilization in LLM Post-training. 2025 December abs/2512.11470 CoRR https://doi.org/10.48550/arXiv.2512.11470 db/journals/corr/corr2512.html#abs-2512-11470 streams/journals/corr

Shijue Huang Wanjun Zhong Jianqiao Lu Qi Zhu 0011 Jiahui Gao 0002 Weiwen Liu Yutai Hou Xingshan Zeng Yasheng Wang Lifeng Shang Xin Jiang 0002 Ruifeng Xu 0001 Qun Liu 0001 Planning, Creation, Usage: Benchmarking LLMs for Comprehensive Tool Utilization in Real-World Complex Scenarios. 4363-4400 2024 ACL (Findings) https://doi.org/10.18653/v1/2024.findings-acl.259 https://aclanthology.org/2024.findings-acl.259 conf/acl/2024f db/conf/acl/acl2024f.html#HuangZLZGLHZWSJ24

Jianqiao Lu Wanjun Zhong Yufei Wang 0005 Zhijiang Guo Qi Zhu 0011 Wenyong Huang Yanlin Wang 0001 Fei Mi Baojun Wang Yasheng Wang Lifeng Shang Xin Jiang 0002 Qun Liu 0001 YODA: Teacher-Student Progressive Learning for Language Models. 2024 abs/2401.15670 CoRR https://doi.org/10.48550/arXiv.2401.15670 db/journals/corr/corr2401.html#abs-2401-15670

Shijue Huang Wanjun Zhong Jianqiao Lu Qi Zhu 0011 Jiahui Gao 0002 Weiwen Liu Yutai Hou Xingshan Zeng Yasheng Wang Lifeng Shang Xin Jiang 0002 Ruifeng Xu 0001 Qun Liu 0001 Planning, Creation, Usage: Benchmarking LLMs for Comprehensive Tool Utilization in Real-World Complex Scenarios. 2024 abs/2401.17167 CoRR https://doi.org/10.48550/arXiv.2401.17167 db/journals/corr/corr2401.html#abs-2401-17167

Zige Wang Wanjun Zhong Yufei Wang 0005 Qi Zhu 0011 Fei Mi Baojun Wang Lifeng Shang Xin Jiang 0002 Qun Liu 0001 Data Management For Large Language Models: A Survey. 2023 abs/2312.01700 CoRR https://doi.org/10.48550/arXiv.2312.01700 db/journals/corr/corr2312.html#abs-2312-01700

Jen-Hao Rick Chang Yuhan Chen Ganqu Cui Shiyao Cui Heyuan Deng Bowen Ding Yi R. Fung 0001 Jiahui Gao 0002 Yize Geng Zhijiang Guo Junxian He Lanqing Hong Lu Hou Yutai Hou Ting-Yao Hu Minlie Huang Shijue Huang Victor Shea-Jay Huang Wenyong Huang Yuzhen Huang 0002 Xin Jiang 0002 Ruochun Jin Xiang Kong Hema Swetha Koppula Jing Li 0049 Jinpeng Li Xujun Li Ming Liao Tao Lin Qun Liu 0001 Weiwen Liu Xian Qi Loye Jianqiao Lu Yen-Ju Lu Jiayang Lv Jiayang Lyu Fei Mi Hadi Pouransari Lifeng Shang Jingwei Song Qinglin Su Shuangshuang Tian Oncel Tuzel Raviteja Vemulapalli Baojun Wang Futing Wang Hanbin Wang Haotian Wang Hongning Wang Hongru Wang 0003 Rui Wang 0092 Sheng Wang Simon Wang Yanlin Wang 0001 Yasheng Wang Yingkang Wang Yufei Wang 0005 Zige Wang Kam-Fai Wong Yin Xia Hongling Xu Minghui Xu Ruifeng Xu 0001 Boyang Xue Junxiao Yang Wenjing Yang 0002 Erxin Yu Jiyao Yuan Weihao Zeng 0003 Xingshan Zeng Zhexin Zhang Mingyuan Zhao Wenshuo Zhao Kehan Zheng Wanjun Zhong Jinfeng Zhou Dantong Zhu