The study investigates whether large language models (LLMs) can reliably detect when their own responses have been manipulated by adversarial prefill attacks. Across ten instruction‑tuned LLMs ranging from 3B to 70B parameters and four safety benchmarks, none consistently recognized compromised outputs, with models claiming intent on prefilled responses at an average of 25.3%. The research identifies that introspective signals mainly arise from safety reasoning and refusal, and that training to improve introspection can paradoxically increase attack success, underscoring the fragility of LLM self‑reporting in safety contexts.
By Quang Minh Nguyen, Uzair Ahmed, Taegyoon Kim
arXiv:2608. 03632v1 Announce Type: new Abstract: On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals.
By Yinuo Jiang, Yongjie Ye, Zhou Tao, Xiang Zhuang, Qiang Zhang, Huajun Chen, Tiankai Li
Video-OPSD introduces a post‑training framework for Video Large Language Models that leverages privileged visual evidence to enhance on‑policy self‑distillation. The method constructs a self‑teacher conditioned only on annotated evidence frames, while the student processes the full video, allowing the teacher to provide more focused supervision. Additionally, an evidence‑guided token optimization scheme weights distillation based on each token’s reliance on privileged evidence, improving perceptually grounded reasoning. Experiments demonstrate consistent gains over standard OPSD and comparable performance to GRPO with less training time.
By Ziyue Wang, Shiqi Huang, Weiwen Xu, Bihan Wen, Xudong Jiang
EOPSA (Efficient On-Policy Self-Distilled Safety Alignment) addresses inefficiencies in On-Policy Self-Distillation (OPSD) for safety alignment by focusing training on safety-critical tokens. It introduces Adaptive Rollout Scheduling, which limits generation length based on a Teacher Rescue Rate metric, and Selective Distillation, which filters out safety-neutral tokens to concentrate gradient updates on safety-pivotal transitions. Experiments on models up to 32B parameters show that EOPSA reduces rollout computation by about 50% and backpropagates through only roughly 2% of tokens, outperforming full-token distillation baselines in safety compliance and reasoning retention.
The paper introduces "cunning questions"—non‑safety prompts that contain misleading premises or subtle inconsistencies—to train large language models (LLMs) to scrutinize underlying intent and assumptions. Experiments show that incorporating these questions improves robustness against out‑of‑distribution jailbreak attacks and enhances subsequent safety fine‑tuning, achieving a new state‑of‑the‑art reduction in mean ASR from 17.40% to 15.05% across nine backbone–benchmark combinations. The authors argue that this training fosters vigilance, enabling models to prioritize safety judgments before engaging in harmful planning.
By Youjia Wang, Lin Xu, Yang Sun, Yuxiao Lu, Chengfang Fang, Jie Shi
The paper investigates how large language models balance helpfulness and safety by refusing harmful queries while responding to benign ones. It decomposes safety-tuning responses into a boilerplate refusal statement and a rationale, finding that the statement causes false refusals by relying on superficial cues. Training on rationales alone reduces false refusals without compromising safety performance, suggesting that fine‑grained safety supervision is essential for better alignment.
By Minji Kim, Hyounghun Kim
arXiv:2606. 03089v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) has emerged as an efficient post-training paradigm by using a teacher conditioned on privileged information to provide dense token-level supervision.
By Ming Wen, Yuxuan Liu, Kun Yang, Yunhao Feng, Zhuoer Xu, Yuhao Sun, Shiwen Cui, Xiang Zheng, Xingjun Ma, Yu-Gang Jiang
arXiv:2608.23313v1 Announce Type: new
Abstract: Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view c...
By Xuetong Li, Gaofeng Liu
arXiv:2607. 02047v1 Announce Type: cross Abstract: Safe completion requires models to provide useful assistance without enabling harm, but this behavior is difficult to evaluate with isolated prompts.
By Rheeya Uppaal, Seungwoo Lyu, Selina Sung, Junjie Hu
arXiv:2609.36562v1 Announce Type: cross
Abstract: While Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, their reliability is threatened by multimodal im...
By Ruochen Zhang, Yao Huang, Yitong Sun, Jiahe Xie, Jin Yan, Jifan Ma, Yuanfang Guo, Xingxing Wei
arXiv:2606. 05718v1 Announce Type: cross Abstract: On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher.
By Kanghui Tian, Siyuan Liu, Ziang Yan, Sheng Xia, Shuai Dong, Yi Wang
On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. Recent selective OPD methods improve this process by prioritizing signals that are confident, informative, or learnable.