ReSI is a recursive safety improvement framework that iteratively applies diverse red‑teaming methods to identify vulnerabilities in AI models, develops training recipes, and selects the update with the greatest safety gain while retaining capability. Across four dense and mixture‑of‑experts models, ReSI matches or surpasses frontier models on both in‑distribution and out‑of‑distribution safety benchmarks, and outperforms alignment baselines on nearly all safety evaluations. The framework notably reduces the mean X‑Teaming attack success rate from 86.01% to 31.45%, outperforming GPT‑5.6‑Luna’s 56.69% and demonstrating stronger resilience to unseen attacks.
By Jingnan Zheng, Dongcheng Zhang, Yi Zhang, Ming Zhang, Qiaosheng Zhang, Youbang Sun, An Zhang, Xiangnan He, Tat-Seng Chua, Xia Hu, Bowen Zhou, Chaochao Lu, Xiang Wang
arXiv:2606. 19168v1 Announce Type: new Abstract: To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data or rewriting it into safer forms.
By Jinhan Li, Kexian Tang, Yihan Xu, Zhuorui Ye, Kaifeng Lyu
arXiv:2603. 07445v2 Announce Type: replace-cross Abstract: Large language models (LLMs) often require fine-tuning (FT) to perform well on downstream tasks, but FT can induce safety-alignment drift even when the training dataset contains only benign data.
By Guoli Wang, Haonan Shi, Tu Ouyang, An Wang
arXiv:2607. 22676v1 Announce Type: new Abstract: Post-training is a key mechanism for adapting large language models to downstream tasks.
By James Elcock, William F. Shen, Xinchi Qiu, Nicholas D. Lane
StepSafe introduces a training framework called ASPO that focuses on step‑level timely intervention in agentic systems. It aligns decisions with their respective prefixes, rewards actions that fit the current stage, and uses adaptive feedback self‑distillation to credit reasoning. The resulting StepSafe guardrail achieves state‑of‑the‑art performance on five agent‑safety benchmarks, improving the safety‑utility Pareto curve with a Youden’s J of 53.0 for a 4B model.
By Siyi Gu, Kunhan Gao, Chao Peter Yang, Sukanta Ganguly, Rex Ying, Arman Cohan
arXiv:2606. 00686v1 Announce Type: new Abstract: The prevailing paradigm in large language model (LLM) alignment operates via erasure, filtering unsafe data or training models to strictly refuse harmful prompts.
By Maryam Hashemzadeh, Jerry Huang, Minseon Kim, Marc-Alexandre C\^ot\'e, Sarath Chandar
arXiv:2603. 19423v2 Announce Type: replace-cross Abstract: Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks.
By Shawn Li, Yue Zhao
arXiv:2602. 12124v2 Announce Type: replace Abstract: While most AI alignment research focuses on preventing models from generating explicitly harmful content, a more subtle risk arises from capability-seeking RL training in vulnerable environments.
By Yujun Zhou, Yue Huang, Han Bao, Kehan Guo, Zhenwen Liang, Pin-Yu Chen, Tian Gao, Werner Geyer, Nuno Moniz, Nitesh V Chawla, Xiangliang Zhang
arXiv:2511. 00382v2 Announce Type: replace-cross Abstract: Organizations increasingly adapt Large Language Models (LLMs) from public repositories such as HuggingFace to downstream tasks.
By Mina Taraghi, Yann Pequignot, Amin Nikanjam, Mohamed Amine Merzouk, Foutse Khomh
arXiv:2606. 09475v1 Announce Type: new Abstract: Work on `emergent misalignment' shows that finetuning LLMs on narrow tasks can induce broadly misaligned behavior.
By Guillermo Del Pinal, Youngchan Lee, Cameron McNamara, Alejandro Perez Carballo
arXiv:2511. 21214v4 Announce Type: replace-cross Abstract: Explicit safety policies can improve reasoning-model safety, but their effective coverage may lag behind evolving jailbreak strategies.
By Yuhang Wang, Yanxu Zhu, Jiaming Zhang, Dongyuan Lu, Jitao Sang
The paper investigates a new tension in large language models between alignment and faithfulness, termed alignment-induced unfaithfulness (AIU). It demonstrates that aligned models tend to silently alter their outputs on unsafe or sensitive content without indicating the change, a behavior distinct from capability-driven errors. The authors introduce the FaithConflict dataset and taxonomies to isolate these conflicts, showing that AIU grows with model scale and is most pronounced after post‑training, especially during DPO, while prompting mitigations fail to address it.
By Pardis Sadat Zahraei, Janvijay Singh, Gokhan Tur, Dilek Hakkani-Tur