STAIR: Improving Safety Alignment with Introspective Reasoning

Yichi Zhang, Siyuan Zhang, Yao Huang, Zeyu Xia, Zhengwei Fang, Xiao Yang, Ranjie Duan, Dong Yan, Yinpeng Dong, Jun Zhu
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:76754-76777, 2025.

Abstract

Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). Specifically, we design a theoretically grounded reward for outcome evaluation to seek balance between helpfulness and safety. We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. We have open-sourced our code, datasets and models at https://github.com/thu-ml/STAIR.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-zhang25cx, title = {{STAIR}: Improving Safety Alignment with Introspective Reasoning}, author = {Zhang, Yichi and Zhang, Siyuan and Huang, Yao and Xia, Zeyu and Fang, Zhengwei and Yang, Xiao and Duan, Ranjie and Yan, Dong and Dong, Yinpeng and Zhu, Jun}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {76754--76777}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/zhang25cx/zhang25cx.pdf}, url = {https://proceedings.mlr.press/v267/zhang25cx.html}, abstract = {Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). Specifically, we design a theoretically grounded reward for outcome evaluation to seek balance between helpfulness and safety. We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. We have open-sourced our code, datasets and models at https://github.com/thu-ml/STAIR.} }
Endnote
%0 Conference Paper %T STAIR: Improving Safety Alignment with Introspective Reasoning %A Yichi Zhang %A Siyuan Zhang %A Yao Huang %A Zeyu Xia %A Zhengwei Fang %A Xiao Yang %A Ranjie Duan %A Dong Yan %A Yinpeng Dong %A Jun Zhu %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-zhang25cx %I PMLR %P 76754--76777 %U https://proceedings.mlr.press/v267/zhang25cx.html %V 267 %X Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). Specifically, we design a theoretically grounded reward for outcome evaluation to seek balance between helpfulness and safety. We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. We have open-sourced our code, datasets and models at https://github.com/thu-ml/STAIR.
APA
Zhang, Y., Zhang, S., Huang, Y., Xia, Z., Fang, Z., Yang, X., Duan, R., Yan, D., Dong, Y. & Zhu, J.. (2025). STAIR: Improving Safety Alignment with Introspective Reasoning. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:76754-76777 Available from https://proceedings.mlr.press/v267/zhang25cx.html.

Related Material