The paper introduces MINT‑Safe, a new open‑source dataset of 11,270 multi‑image dialogues and 500 refusal VQA pairs designed to expose safety risks in multi‑modal large language models during open‑ended conversations. It also proposes TAD‑Align, a turn‑aware dual‑objective reward framework that dynamically up‑weights dialogue turns with inconsistent safety behavior, improving safety metrics on Qwen2.5‑VL‑7B‑Instruct and LLaVA‑Next‑7B. The results show over 10% reduction in attack success rate and notable gains in harmlessness and helpfulness while maintaining overall model performance.
By Han Zhu, Jiale Chen, Chengkun Cai, Shengjie Sun, Haoran Li, Yujin Zhou, Chi-Min Chan, Pengcheng Wen, Lei Li, Yike Guo, Sirui Han
Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but losing sight of the victim's replies, wasting tur...
arXiv:2610.35743v1 Announce Type: new
Abstract: Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but...
By Mingqian Feng, Xiaodong Liu, Weiwei Yang, Chenliang Xu, Jialin Song, Qiyao Ma, Chengquan Guo, Carolyn Buractaon, Christopher White, Jianfeng Gao
arXiv:2602. 13379v2 Announce Type: replace-cross Abstract: LLM-based agents are becoming increasingly capable, yet their safety lags behind.
By Xu Li, Simon Yu, Minzhou Pan, Yiyou Sun, Bo Li, Dawn Song, Xue Lin, Weiyan Shi
The paper introduces NegWAM, a safety post‑training framework for unified world‑action models (WAMs) used in end‑to‑end autonomous driving. NegWAM generates negative trajectories that are close to expert paths but contain safety risks, using them as repulsive anchors to steer predictions away from dangerous behavior while encoding scene‑trajectory compatibility via rider tokens. The method achieves a 4.1% relative improvement on the NAVSIM v2 navhard benchmark and 93.3 EPDMS on navtest, demonstrating strong safety performance without sacrificing routine driving, and shows cross‑benchmark generalization in zero‑shot closed‑loop tests on HUGSIM.
By Zihao Zhang, Xinyu Zhang, Bing Zeng, Xingyu Chen, Zhenning Li, Bin Yu, Ziyuan Pu
arXiv:2607. 18063v1 Announce Type: cross Abstract: LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation.
By Devina Jain, David Hartmann, Chuan Li
FPD-Drive is an end‑to‑end autonomous driving planner that uses future‑privileged multi‑target distillation to transfer diverse, high‑quality trajectory targets from a teacher model with access to future observations to a student policy. It also introduces safety‑aware score correction to improve candidate selection by adjusting rankings between safe and unsafe trajectories. The approach achieves state‑of‑the‑art performance on NAVSIM v1 and v2 benchmarks and demonstrates strong zero‑shot closed‑loop results on HUGSIM.
By Li-Heng Chen, Qifeng Chen, Chengwei Wu, Yingjie Zhang, Sheng Yang, Hongbo Fu, Shaoqing Ren
arXiv:2606. 19047v1 Announce Type: new Abstract: Multi-turn tool-use RL is bottlenecked by the rapid depletion of informative samples in static datasets.
By Ruishan Fang, Siyuan Lu, Chenyi Zhuang, Tao Lin
arXiv:2607. 16244v1 Announce Type: cross Abstract: Training multi-turn evidence-reading agents with outcome-only reinforcement learning is unstable because intermediate turns receive little direct credit.
By Hao Dou
arXiv:2606. 02530v1 Announce Type: new Abstract: Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, termed the alignment tax.
By Hao Li, Jingkun An, Zijun Song, Pengyu Zhu, Rui Li, Hao Wang, Wendi Feng, Yesheng Liu, Lijun Li, Jin-Ge Yao, Lei Sha
arXiv:2603.04364v2 Announce Type: replace-cross
Abstract: Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet th...
By Haoyu Liu, Dingcheng Li, Lukas Rutishauser, Zeyu Zheng
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